Deep learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction

Cenitta D; VIijaya Arjunan Ranganathan; Tanuja Shailesh; Andrew J; Arul N; Praveen Pai T

doi:10.12688/f1000research.165575.2

Home Browse Deep learning based hybrid residual attention and echo state network...

ALL Metrics

Views

Downloads

Get PDF

Get XML

Export

▬

✚

Research Article

Revised

Deep learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction

[version 2; peer review: 1 approved, 1 approved with reservations, 1 not approved]

Cenitta D¹, VIijaya Arjunan Ranganathan ¹, Tanuja Shailesh¹, Andrew J¹, Arul N², Praveen Pai T¹

Cenitta D¹, VIijaya Arjunan Ranganathan ¹, [...] Tanuja Shailesh¹, Andrew J¹, Arul N², Praveen Pai T¹

PUBLISHED 16 Sep 2025

Author details Author details

¹ Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India
² Computer Science and Engineering, AJ Institute of Engineering and Technology, Mangalore, Karnataka, India

Cenitta D
Roles: Methodology, Project Administration

VIijaya Arjunan Ranganathan
Roles: Conceptualization, Writing – Review & Editing

Tanuja Shailesh
Roles: Writing – Review & Editing

Andrew J
Roles: Data Curation

Arul N
Roles: Visualization

Praveen Pai T
Roles: Writing – Review & Editing

OPEN PEER REVIEW

REVIEWER STATUS

This article is included in the Artificial Intelligence and Machine Learning gateway.

This article is included in the Manipal Academy of Higher Education gateway.

Abstract

Background

Early and accurate prediction of ischemic heart disease (IHD) is essential for reducing mortality and enabling timely intervention. Misdiagnosis can lead to severe health outcomes, emphasizing the need for robust and intelligent predictive models. Deep learning approaches have shown strong potential in identifying hidden patterns in medical data and aiding clinical decision-making.

Methods

This study proposes a novel Hybrid Residual Attention with Echo State Network (HRAESN) model that integrates Attention Residual Learning (ARL) with Echo State Networks (ESN) to enhance feature extraction and temporal data learning. The hybrid model is designed to refine feature attention through residual learning while leveraging ESN for efficient time-series prediction. Two publicly available benchmark datasets were used for evaluation: the Kaggle Cardiovascular Disease dataset comprising 70,000 instances and the UCI Heart Disease dataset containing 303 instances. Missing values in both datasets were handled using a multiple imputation technique tailored for ischemic heart disease. Model performance was assessed using standard classification metrics, including accuracy, sensitivity, specificity, precision, recall, and F-measure.

Results

The proposed HRAESN model demonstrated superior classification performance compared to traditional machine learning and deep learning approaches. It achieved an accuracy of 98.4% on the Kaggle dataset and 97.7% on the UCI dataset. Additionally, the model showed high sensitivity and specificity, indicating strong diagnostic capability and reliability in identifying both diseased and non-diseased cases.

Conclusions

The HRAESN model effectively combines the strengths of residual attention mechanisms and echo state networks, resulting in improved accuracy and stability for ischemic heart disease prediction. Its strong performance on benchmark datasets confirms its potential as a valuable clinical decision support tool for early detection of IHD. Future work may focus on optimizing model complexity and integrating real-time medical IoT data to enhance practical deployment in healthcare systems.

Keywords

UCI, Kaggle, Heart Disease, Imputation, Deep Learning, Echo State Network, Residual Attention.

Corresponding authors: VIijaya Arjunan Ranganathan, Tanuja Shailesh

Competing interests: No competing interests were disclosed.

Grant information: The author(s) declared that no grants were involved in supporting this work.

Copyright: © 2025 D C et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

How to cite: D C, Ranganathan VA, Shailesh T et al. Deep learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction [version 2; peer review: 1 approved, 1 approved with reservations, 1 not approved]. F1000Research 2025, 14:650 (https://doi.org/10.12688/f1000research.165575.2) First published: 03 Jul 2025, 14:650 (https://doi.org/10.12688/f1000research.165575.1) Latest published: 16 Sep 2025, 14:650 (https://doi.org/10.12688/f1000research.165575.2)

Revised Amendments from Version 1

In this revised version, we have substantially strengthened the methodological transparency, statistical rigor, and reproducibility of our study. The Introduction was rewritten for improved flow and updated to reflect current diagnostic practices in ischemic heart disease (IHD), replacing outdated modalities with contemporary techniques (cardiac CT, RbPET, coronary angiography). The Related Works section was streamlined and expanded to cover prior attention–Echo State Network (ESN) combinations, thereby clarifying the novelty of our Hybrid Residual Attention with Echo State Network (HRAESN) model.
In the Methods, we now provide detailed definitions of heart disease/IHD in the UCI and Kaggle datasets, describe missingness and imputation using the IHD Multiple Imputation Technique, and explain how ESNs were adapted for structured tabular data. Evaluation metrics, including Cohen’s kappa and Jaccard index, are introduced earlier for consistency.
To strengthen statistical robustness, we re-ran all experiments using 5-fold and 10-fold stratified cross-validation, reporting mean ± standard deviation across folds. We added statistical significance testing (McNemar’s test and Wilcoxon signed-rank) and expanded performance evaluation with ROC curves, AUC values, precision–recall curves, and calibration plots. Confidence intervals (95%) were computed via bootstrap resampling.
Figures and tables were revised to improve clarity: Figure captions specify dataset scope, confusion matrices now include raw counts, and baseline population characteristics are summarized in a new table. Comparative analysis was clarified to explain baseline method selection.
The Discussion and Limitations were expanded to address external validation, imputation bias, dataset imbalance, and interpretability of the Attention Residual Learning module. Claims regarding clinical readiness were moderated to emphasize proof-of-concept status.
Finally, to enhance reproducibility, we expanded algorithmic details of the imputation method and provide code availability upon request. Collectively, these revisions address reviewer feedback and significantly improve the rigor, transparency, and interpretability of our work.

See the authors' detailed response to the review by MUHAMMAD HAMMAD MEMON
See the authors' detailed response to the review by Amalie Dahl Haue
See the authors' detailed response to the review by Dhadkan Shrestha

1. Introduction

Ischemic heart disease (IHD) arises when coronary arteries are narrowed or blocked, leading to reduced blood flow and oxygen supply to the heart muscle. Persistent restriction of coronary circulation results in myocardial ischemia, which can progress to coronary artery disease and, in severe cases, myocardial infarction. Silent ischemia, in particular, occurs without overt symptoms but still poses a high risk of sudden cardiac events, especially in individuals with diabetes or a prior history of heart attack. In current clinical practice, the diagnosis and assessment of IHD relies on advanced imaging and invasive modalities, including cardiac computed tomography (CT), Rubidium positron emission tomography (RbPET), and coronary angiography, which provide accurate evaluation of coronary artery stenosis and perfusion deficits. These methods, while effective, remain costly, invasive, and not always feasible for large-scale population screening, motivating the exploration of non-invasive, AI-based predictive approaches for early detection.¹

The World Health Organization (WHO) reports that cardiovascular diseases (CVDs) continue as the main cause of global mortality since 17.9 million people died from CVDs in 2019 which amounted to 32% of worldwide fatalities. Heart attacks and strokes lead to 85% of fatal outcomes among the tested patients.² The worldwide fatalities from noncommunicable diseases reached 17 million during 2019 before people turned 70 years old and cardiovascular conditions caused 38% of those premature deaths. Medical detection of CVDs remains vital because behavioral prevention through risk control methods such as smoking and food control and weight management cannot substitute for early medical discovery to achieve both effective treatment and lower mortality rates. Heart disease poses a major financial challenge and increasing health burden because of high surgical expenses and rising population incidence mainly affecting developing countries. Knowledge about how patient characteristics link to heart disease risk serves as the basis for preventing the condition and detecting it early for treatment purposes.

Deep learning has become an integral part of computer vision, object recognition, natural language processing, speech recognition, medical diagnostics, bioinformatics, and drug discovery. Similar to traditional artificial neural networks (ANNs), deep learning models consist of input, hidden, and output layers, with patient risk factors serving as input features. The research demonstrates that artificial neural networks deliver outstanding results when used for identifying and foretelling coronary heart disease.³ Medical AI applications experience rapid growth because of three main factors including Internet of Things (IoT) and powerful computing hardware (e.g., GPUs and TPUs) together with big medical datasets. Essential information needed by deep learning models comes from Medical IoT devices together with electronic health records as well as genomic data and central medical databases. The critical challenges include preserving data privacy as well as successfully deploying the models and optimizing service quality despite their importance.³

Time-series prediction has seen increased popularity among researchers who use recurrent neural networks (RNNs) as deep learning-based approaches. RNNs work with sequential data sets through the process of feeding output data from previous components to next steps making them ideal for ECG signal processing and patient health surveillance. RNNs differ from regular neural networks by retaining previous input data thus they produce enhanced forecasts for temporal information patterns. Traditional RNNs experience gradient vanishing problems because of which they become problematic for handling long sequences. The development of both Hochreiter and Schmidhuber led to long short-term memory (LSTM) networks which incorporated memory gates to control information transmission and suppress gradient deterioration.⁴

Time-series extrapolation along with fast learning occurs efficiently through Echo State Networks (ESN) which function as a preferred substitute to normal RNNs.⁵ An Echo State Network functions through its reservoir of recurrent neurons connected haphazardly that helps the network learn complex patterns yet uses few processing resources. The forecast capabilities of time-series prediction and representation learning capabilities improve through the use of Deep ESNs (DESNs) that include multiple serially connected reservoirs.⁶

A transformation of conventional convolutional neural networks (CNNs) called Residual Attention Network brings attention mechanism integration for feature enhancement.⁷ The advanced feed-forward framework permits end-to-end training which enables it to learn hierarchical features independently. Gremlin Deep Residual Attention Networks provide an efficient mechanism for deep learning systems to reach hundreds of layers through their implementation of Attention Residual Learning (ARL).⁸ Different algorithms can achieve maximum strength performance through hybrid deep learning models which integrate multiple techniques. Medical diagnostic accuracy along with efficiency can experience significant improvement by combining residual attention learning methods with Echo State Networks. The appropriate addressing of missing values through the Ischemic Heart Disease Multiple Imputation Technique creates improved data reliability and completeness.⁹

1.1 Objective of this study

The main goal of this research work is to create a Hybrid Residual Attention with Echo State Network (HRAESN) model used to predict ischemic heart disease (IHD) at an early stage while maintaining high accuracy. The proposed method integrates Residual Attention Learning (RAL) with Echo State Networks (ESNs) to boost both feature extraction and time-series classification and general model performance. This study solves data preprocessing problems with Ischemic Heart Disease Multiple Imputation Technique while using hybrid deep learning effectively for robust classification. The research uses two recognized heart disease data sets including 70,000 records from the Kaggle Cardiovascular Disease dataset and 303 records from the UCI Heart Disease dataset to evaluate the proposed method. The objective is to prove that this approach outperforms current state-of-the-art heart disease prediction methods. ART-based analysis findings will enhance clinical diagnosis along with IHD detection and patient care through AI-powered diagnostic systems.

The following research questions are the focus of the study’s search and synthesis of the literature.

1. How do deep learning models, particularly Echo State Networks (ESNs) and Residual Attention Learning (RAL), improve the accuracy and stability of ischemic heart disease prediction compared to traditional machine learning approaches?
2. What are the key challenges associated with handling missing data in medical datasets, and how can the Ischemic Heart Disease Multiple Imputation Technique enhance data completeness and reliability?
3. How does the proposed Hybrid Residual Attention with Echo State Network (HRAESN) model perform on benchmark datasets (Kaggle Cardiovascular Disease and UCI Heart Disease) compared to existing state-of-the-art heart disease prediction models?

1.2 Problem statement

One of the main causes of death is ischemic heart disease (IHD), which needs to be predicted early and accurately in order to be effectively treated. While current machine learning models have trouble managing missing data, time-series dependencies, and computational inefficiencies, traditional diagnostic techniques are costly, time-consuming, and rely on expert interpretation. Vanishing gradients and high complexity are two drawbacks of deep learning techniques like recurrent neural networks (RNNs) and long short-term memory (LSTM) networks. To address these challenges, this study proposes a Hybrid Residual Attention with Echo State Network (HRAESN) model, integrating Residual Attention Learning (RAL) for feature extraction and Echo State Networks (ESNs) for efficient time-series processing, ensuring improved predictive accuracy and robustness.

2. Related works

Numerous studies have explored machine learning (ML) and deep learning (DL) techniques for cardiovascular disease prediction. Traditional ML methods such as Decision Trees, Random Forests, Naïve Bayes, and Support Vector Machines have shown moderate success but often struggle with missing data, feature complexity, and generalization.¹⁰^–¹³

Deep learning models, particularly Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) architectures, have been widely applied to ECG-based diagnosis and patient monitoring, achieving improved accuracy in handling sequential data.¹⁴ Hybrid RNN–LSTM models, for example, demonstrated higher classification performance than standalone approaches.¹⁴^,¹⁵

Echo State Networks (ESNs) and related reservoir computing methods have also been applied in cardiovascular applications due to their efficiency in time-series prediction. Li et al.¹⁶ showed effective heartbeat classification using a residual squeeze-and-excitation framework, while Gao et al.⁵ and Sun et al.¹⁷ combined ESNs with wavelet transformation and Deep Belief Networks, respectively, to improve temporal modeling. Optimized ESN variants, including bidirectional¹⁸ and adaptive evolutionary models,¹⁹^–²¹ have further enhanced performance, and hardware-efficient ESN implementations have demonstrated low-power solutions for clinical settings.²²

Attention mechanisms and residual learning have similarly strengthened feature representation in medical tasks. Residual Attention Graph Convolutional Networks²³ and deep residual attention models⁸^,²⁴ and Residual Attention Graph Convolutional Networks²⁵ demonstrated improvements in complex classification tasks. Feature refinement approaches such as Recursion-Enhanced Random Forest²⁶ and SVM-based ensembles with feature elimination²⁷ have also been explored for cardiovascular disease detection.

Hybrid frameworks that combine different learning paradigms have become increasingly popular. Examples include CNN–reservoir computing hybrids,²⁸ CNN–reservoir computing hybrids,²⁹ clustering-enhanced prediction models,¹³^,¹⁵ and DBN–RNN integrations optimized by metaheuristics.³⁰ Ensemble-based strategies, including two-tier classifiers and hybrid Random Forest/Gradient Boosting methods, have further improved classification outcomes.³¹^,³²

Overall, existing studies highlight three main trends: (i) ESNs provide efficient temporal modelling for cardiovascular data, (ii) attention-based residual learning enhances feature extraction, and (iii) hybrid frameworks that integrate these methods yield superior predictive accuracy. Building on these findings, our proposed HRAESN model integrates Residual Attention Learning with ESNs to address limitations in prior models and achieve higher accuracy, stability, and robustness in ischemic heart disease prediction.

3. Materials and methods

3.1 Dataset

This study utilizes data from two publicly available repositories: Kaggle and the UCI (University of California, Irvine) Machine Learning Repository. These datasets provide comprehensive patient records used for cardiovascular disease prediction and ischemic heart disease classification.

3.1.1 Kaggle cardiovascular disease dataset

There are 70,000 patient records with 11 distinct features in the Kaggle Cardiovascular Disease dataset.³³ When medical practitioners performed clinical examinations, these characteristics were noted. Three types of input features make up the dataset:

1. Objective Characteristics (Real patient data): Gender, Age, Height, and Weight
2. Features of the Examination (Medical Test Results): Blood Pressure Systolic and Diastolic, Blood Pressure Levels of Cholesterol and Glucose
3. Subjective Features (patient data as self-reported): Alcohol use, smoking, and physical activity

3.1.2 UCI heart disease dataset

The UCI Heart Disease dataset contains 76 features, of which 14 are highly relevant for heart disease diagnosis.³⁴ The predictive class attribute is typically listed last, indicating the presence or absence of heart disease. Table 1 and Table 2 provide detailed descriptions of the dataset attributes.

Table 1. Kaggle cardiovascular disease dataset description.

Attribute	Description
Age	Objective Feature\|age\|int (days)
Height	Objective Feature\|height\|int (cm)\|
Weight	Objective Feature\|weight\|float (kg)\|
Gender	Objective Feature\|gender\|categorical code\|
Systolic blood pressure	Examination Feature\|ap_hi\|int\|
Diastolic blood pressure	Examination Feature\|ap_lo\|int\|
Cholesterol	Examination Feature\|cholesterol\| 1: normal, 2: above normal, 3: well above normal
Glucose	Examination Feature\|gluc\| 1: normal, 2: above normal, 3: well above normal
Smoking	Subjective Feature\|smoke\|binary\|
Alcohol intake	Subjective Feature\|alco\|binary\|
Physical activity	Subjective Feature\|active\|binary\|
Presence or absence of cardiovascular disease	Target Variable\|cardio\|binary\|

Table 2. UCI heart disease dataset description.

Attribute	Description	Domain of value
Age	Age in year	29 to 77
Sex	Sex	Male (1)
Sex	Sex	Female (0)
Cp	Chest pain type	Typical angina (1)
		Atypical angina (2)
		Non-anginal (3)
		Asymptomatic (4)
Trestbps	Resting blood sugar	94 to 200 mm Hg
Chol	Serum cholesterol	126 to 564 mg/dl
Fbs	Fasting blood sugar	>120 mg/dl
		True (1)
		False (0)
Restecg	Resting ECG result	Normal (0)
		ST-T wave
		Abnormality (1)
		LV hypertrophy (2)
Thalach	Maximum heart rate achieved	71 to 202
Exang	Exercise induced angina	Yes (1)
Exang	Exercise induced angina	No (0)
Oldpeak	ST depression induced by exercise relative to rest	0 to 6.2
Slope	Slope of peak exercise ST segment	Upsloping (1)
		Flat (2)
		Downsloping (3)
Ca	Number of major vessels coloured by fluoroscopy	0 – 3
Thal	Defect type	Normal (3)
		Fixed defect (6)
		Reversible defect (7)
Num	Heart disease	0-4

3.1.3 Datasets and ethical considerations

This study utilizes two publicly available datasets: the Heart Disease dataset from the UCI Machine Learning Repository and the Cardiovascular Disease dataset from Kaggle. These datasets contain anonymized patient records and are publicly released for academic and research purposes.

3.1.4 Ethical approval statement

As this research involves only the use of publicly accessible, anonymized datasets, no formal ethical approval was required. The study complies with the ethical principles outlined in the Declaration of Helsinki. No intervention or interaction with human subjects occurred.

3.1.5 Informed consent statement

Because this study used pre-existing anonymized data from public repositories, informed consent from participants was not required. All necessary ethical permissions and participant consents were obtained by the original data providers as per their respective institutional and data-sharing policies.

3.1.6 Definition of heart disease in the datasets

In the UCI Heart Disease dataset, the target variable “num” (values 0–4) indicates the severity of disease as determined by coronary angiography. For this study, we followed prior works³¹^,³⁴^–³⁷ and binarised the variable: 0 = absence of disease, 1–4 = presence of heart disease. In the Kaggle Cardiovascular Disease dataset, the binary target variable “cardio” was defined during the original data collection based on combined clinical assessment and diagnostic test results (blood pressure, cholesterol, ECG). Here, 0 = healthy and 1 = diagnosed cardiovascular disease.

3.1.7 Dataset inclusion and missing values

All available records were included: 70,000 instances in the Kaggle dataset and 303 in the UCI dataset. The Kaggle dataset contained ~0.3% missing values across features, while the UCI dataset had six missing entries. These were imputed using the Ischemic Heart Disease Multiple Imputation Technique,⁹ ensuring that no records were discarded and data completeness was preserved.

To ensure that our training and testing sets were representative, we verified that baseline characteristics (age, sex, cholesterol, and blood pressure) were similarly distributed between the two subsets. Table 3 presents the distributions of these key features for both training and testing populations in the UCI and Kaggle datasets.

Table 3. Baseline characteristics of training and testing populations for the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

Values are mean ± SD for continuous variables and % for categorical variables.

Variable	UCI Train (n=242)	UCI Test (n=61)	Kaggle Train (n=56,000)	Kaggle Test (n=14,000)
Age (years)	54.3 ± 9.2	54.7 ± 8.9	54.3 ± 6.7*	54.3 ± 6.8*
Male (%)	68.20%	68.90%	34.90%	35.10%
Cholesterol (mg/dl)	244.9 ± 47.8	251.7 ± 65.5	–	–
Systolic BP (mmHg)	131.7 ± 18.0	131.3 ± 15.7	129.0 ± 161.6	127.9 ± 119.2
Cholesterol categories^†	–	–	1 = 74.8%, 2 = 13.7%, 3 = 11.5%	1 = 74.9%, 2 = 13.5%, 3 = 11.6%

3.2 Hybrid data classification algorithm

The classification of ischemic heart disease (IHD) in this study is based on a hybrid deep learning model that integrates machine learning (ML), soft computing techniques, and optimization methods to enhance accuracy and robustness. Different classification models are created by integrating various ML methods and ensemble learning methods that involve bagging and boosting. Multiple classifiers work together in ensemble methods to generate better generalization as well as decrease overfitting.

HRAESN model combines the following key elements:

1. Echo State Networks (ESNs) for efficient time-series processing
2. Attention Residual Learning (ARL) for enhanced feature extraction

By combining ESN and ARL, the model achieves higher accuracy, better generalization, and improved stability compared to conventional ML classifiers.

3.3 Echo State Network (ESN)

Echo State Networks (ESNs), a subset of recurrent neural networks (RNNs) created for effective sequential data processing, are a part of the reservoir computing paradigm. In contrast to conventional RNNs, an ESN’s hidden layer (reservoir) is fixed and randomly initialized, whereas only the output layer is trained.

Key features of ESNs include:

• The reservoir exhibits two weight sets which are fixed by random values without training: W_in for input-to-lateral connections and W_r for lateral connections.
• During ESN operation researchers only train output weights but maintain simple computational design for efficient pattern learning capability.
• The hidden layer connectivity of ESNs remains sparse which decreases computational complexity.
• Nonlinear Embedding: The reservoir state provides a nonlinear transformation of input data, which can then be mapped to the desired output using a trainable readout layer.
Since ESNs retain past information in a fixed reservoir, they are highly effective for time-series forecasting and real-time signal processing, making them a suitable choice for ischemic heart disease prediction.

3.4 Attention Residual Learning (ARL)

Attention Residual Learning (ARL) is a deep learning technique that enhances feature extraction by selectively focusing on relevant information while reducing noise in deep neural networks. It is particularly beneficial in medical image analysis and time-series classification.

Key challenges in deep residual networks include:

• Performance Degradation: Stacking multiple narrow attention modules can lead to a decline in performance.
• Feature Suppression: Soft mask layers may inadvertently reduce the importance of relevant features.

To address these issues, ARL modifies feature representation using an attention mask. The transformation is mathematically represented as:

H_{i} (t + 1) = (1 + M_{i} (t)) * X_{F}^{i} (t)

Where:

i: Index position in the input matrix

M_i (t): Gradient of the input feature mask during the t-th iteration

H_i (t+1): Updated attention module output at the (t+1)-th iteration

This formulation ensures that:

1. Relevant features are amplified, while irrelevant features are suppressed.
2. Deep residual networks maintain stable performance even with hundreds of layers.
3. Computational efficiency is preserved without significantly increasing model complexity.

The integration of ESNs with ARL enables the proposed HRAESN model to merge its time-series learning functionality with attention-based feature refinement that results in precise and stable outcomes for ischemic heart disease predictions.

3.5 Methodology

The prediction model utilizes heart disease records from UCI Heart Disease Data Set and the Cardiovascular Disease dataset from Kaggle. Pre-processing starts with performing the Ischemic Heart Disease Multiple Imputation Technique to identify and imputation missing values before proceeding further.¹ The HRAESN model combines Echo State Networks (ESNs) for short-term memory processing with Attention Residual Learning (ARL) for enhancing features to classify heart disease.

Figure 1. Workflow of the proposed experiment using the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

Workflow of the proposed experiment using the UCI Heart Disease (303 records, 14 features, 6 missing values) and Kaggle Cardiovascular Disease dataset (70,000 records, 11 features, ~0.3% missing values). Missing values were imputed using the Ischemic Heart Disease Multiple Imputation Technique. Labels were defined as binary: UCI “num” attribute (0 = healthy, 1–4 = disease present, recoded to 0/1) and Kaggle “cardio” attribute (0 = healthy, 1 = disease present). The preprocessed datasets were fed into the HRAESN model, combining Echo State Networks for reservoir-based representation of clinical features with Attention Residual Learning for enhanced feature selection. Model training and evaluation used an 80:20 split, with multiple performance metrics reported.

Figure 2. Overall system model of the proposed Hybrid Residual Attention with Echo State Network (HRAESN).

Patient data (70,000 Kaggle records, 303 UCI records) were preprocessed and imputed before being passed into an Echo State Network reservoir, which captures nonlinear feature interactions. The reservoir outputs were refined using Attention Residual Learning, which selectively enhances relevant clinical patterns while suppressing noise. A final sigmoid activation layer produces binary predictions (0 = healthy, 1 = ischemic heart disease). This architecture leverages ESN efficiency and attention-driven feature refinement for improved classification accuracy.

Experiment workflow

1. Load and preprocess datasets: The Heart Disease Data Set and Cardiovascular Disease dataset are loaded, and missing values are imputed using the Ischemic Heart Disease Multiple Imputation Technique.^9,38
2. Feature extraction and classification: The HRAESN model applies ESNs for sequence modeling and ARL for refining feature representation.
3. Model evaluation: A confusion matrix assesses the model’s performance, ensuring accurate classification of heart disease cases.

3.5.1 Hybrid Residual Attention with Echo State Network (HRAESN) algorithm

The input feature matrix (X_F) is obtained from the Ischemic Heart Disease Multiple Imputation Technique and labeled according to class 0 (normal) or class 1 (heart disease).

Echo State Network (ESN) Hidden Layer Dynamics

(1)

X_{F} (t + 1) = f_{a} (W^{i} u (t) + W^{r} X_{F} (t))

Where:

• $X_{F} (t + 1)$ and $X_{F} (t)$ are the feature matrices at iterations t and t + 1.
• $W^{i}$ is the input reservoir weight matrix derived from the input data.
• $W^{r}$ is the reservoir weight matrix representing internal states.
• $u (t + 1)$ represents the internal states computed at iteration t.
• $f_{a} (.)$ is the activation function applied at the reservoir.

Attention Residual Learning (ARL) transformation

(2)

H_{i} (t + 1) = (1 + M_{i} (t)) * X_{F}^{i} (t)

Where:

• $i$ represents the input matrix’s index positions.
• $M_{i} (t)$ is the gradient of the input feature mask at iteration t.
• $H_{i} (t + 1)$ is the attention module output at iteration t + 1.

The reservoirs in HRAESN are linked in series, meaning each reservoir state depends on the previous reservoir’s output and its own past state:

(3)

X_{F}^{1} (t + 1) = f_{a} (W^{i} u (t) + W^{1} X_{F}^{1} (t))

(4)

X_{F}^{2} (t + 1) = f_{a} (W^{i} X_{F}^{1} (t) + W^{2} X_{F}^{2} (t))

(5)

X_{F}^{M} (t + 1) = f_{a} (W^{i} X_{F}^{(M - 1)} (t) + W^{M} X_{F}^{M} (t))

Where:

• $W^{i} = H_{i} (t + 1)$ represents the attention module output.

Activation Functions and Output Computation

1. Final Activation Function

(6)

A_{n} = Y_{L} \cdot sigmoid (X_{F}^{M} (t + 1))

Where:

• $sigmoid (.)$ is the activation function applied to the final output layer.

Dynamic Echo State Network Output

(7)

P_{R} (t + 1) = g_{a} (W^{o} X_{F}^{M} (t + 1))

Where:

• $W^{o}$ represents the output reservoir weight matrix.
• $g_{a} (.)$ is the final activation function used at step 4.

Algorithm. Hybrid Residual Attention with Echo State Network (HRAESN).

Input: features data $X_{F}$ , label data $Y_{L}$

Output: Predicted result P_r

1: begin

2: for each Compute the Hidden layer of dynamic ESN

3: $X_{F} (t + 1) = f_{a} (W^{i} u (t) + W^{r} X_{F} (t))$

4: end for

5: for each compute the attention residual learning

6: $H_{i} (t + 1) = (1 + M_{i} (t)) * X_{F}^{i} (t)$

7: end for

8: for x=1 to M do:

9: $X_{F}^{1} (t + 1) = f_{a} (W^{i} u (t) + W^{1} X_{F}^{1} (t))$

10: $X_{F}^{2} (t + 1) = f_{a} (W^{i} X_{F}^{1} (t) + W^{2} X_{F}^{2} (t))$

11: …

12: $X_{F}^{M} (t + 1) = f_{a} (W^{i} X_{F}^{M - 1} (t) + W^{M} X_{F}^{M} (t))$

13: end

14: end

Evaluation metrics

The predictive performance of the proposed HRAESN model and baseline classifiers was assessed using multiple evaluation metrics. Standard measures included:

• Accuracy: the proportion of correctly classified instances among all instances.
• Sensitivity (Recall): the proportion of true positive cases (IHD present) correctly identified.
• Specificity: the proportion of true negative cases (IHD absent) correctly identified.
• Precision: the proportion of predicted positives that are true positives.
• F1-score: the harmonic mean of precision and recall, balancing sensitivity and specificity.

In addition, we introduced two supplementary metrics to capture model agreement and similarity beyond traditional measures:

• Cohen’s Kappa Coefficient: quantifies agreement between predicted and actual classifications beyond chance, with values closer to 1 indicating stronger agreement.
• Jaccard Coefficient: measures the similarity between predicted and actual sets of positive cases, defined as the intersection divided by the union of the sets.

For statistical robustness, 95% confidence intervals (CIs) were estimated for all major performance metrics using a bootstrap resampling strategy (1000 resamples). These CIs provide an indication of the reliability and significance of the reported values.

3.5.2 Hyperparameter tuning

The Hyperparameter Tuning process optimizes the performance of the Hybrid Residual Attention with Echo State Network (HRAESN) model by carefully selecting key parameters for both Echo State Networks (ESN) and Attention Residual Learning (ARL). The reservoir size (500 neurons) and spectral radius (0.8) ensure stable memory retention for time-series processing, while 10% sparse connectivity enhances computational efficiency. The input scaling (0.5) and leaky rate (0.2) regulate data flow within the reservoir, preventing overfitting. The attention module depth (3 layers) and mask range ([0,1]) refine feature selection, improving model interpretability. The model is trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 100 epochs for optimal convergence. The model prevents overfitting through dropout rate 0.3 while 80:20 train-test split maintains evaluation stability. The optimized parameters lead to precise and efficient and stable ischemic heart disease predictions as described in Table 4.

Table 4. Summary of hyperparameter settings used in the proposed model training.

Parameter	Value	Description
Number of Reservoir Neurons (N_res)	500	Number of neurons in the ESN reservoir. Determines the capacity of the reservoir to store and process sequential information.
Spectral Radius (ρ)	0.8	Controls the stability of the ESN. A value < 1 ensures echo state property for long-term memory.
Reservoir Connectivity (%)	10%	Percentage of nonzero connections in the reservoir matrix W^r, ensuring sparse connectivity.
Input Scaling (W_in)	0.5	Determines how input data is mapped into the reservoir.
Leaky Rate (α)	0.2	Defines how much of the previous state is retained in the ESN for time-series processing.
Readout Regularization (λ)	10⁻⁴	Ridge regression parameter to prevent overfitting in the output layer of ESN.
Attention Module Depth	3	Number of stacked attention modules in ARL to enhance feature learning.
Attention Mask Range (M_i (t))	[0,1]	Defines the range of soft masks applied in attention residual learning.
Activation Function (f_a(.))	Tanh	Non-linear activation function used in the ESN reservoir.
Output Activation Function (g_a(.))	Sigmoid	Activation function used in the final output layer to predict class labels.
Batch Size	32	Number of training samples processed before updating model weights.
Optimizer	Adam	Optimization algorithm used to update model parameters.
Learning Rate (η)	0.001	Controls the step size of weight updates during training.
Dropout Rate	0.3	Fraction of neurons randomly dropped during training to prevent overfitting.
Number of Epochs	100	Total number of times the model iterates over the entire dataset during training.
Train-Test Split Ratio	80:20:00	Data split for training (80%) and testing (20%).

4. Results and analysis

To predict the existence of ischemic heart disease (IHD), a number of classification methods were employed, including Naïve Bayes (NB), Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), and AdaBoost. Data from the Cardiovascular Disease dataset (Kaggle) and the Heart Disease Data Set (UCI) were used in the experiments.

4.1 Experiment setup and data preprocessing

The datasets contain various medical indicators that serve as input features for classification. The target variable is binary:

• Class 1: Presence of ischemic heart disease
• Class 0: Absence of disease

The proposed hybrid HRAESN model is trained using 80% of the dataset, and the remaining 20% is used for testing. Principal Component Analysis (PCA) was applied to highlight variance and distinct patterns in the dataset. Figure 3 shows the PCA plot, where:

• Principal Component 1 (X-axis) and Principal Component 2 (Y-axis) capture most of the variance.
• Blue (0) represents healthy individuals, while Red (1) represents patients with heart disease.

Figure 3. PCA plot showing data distribution in the heart disease dataset based on the first two principal components.

Additionally, six records in the UCI dataset had missing values, which were imputed using the Ischemic Heart Disease Multiple Imputation technique, producing a complete dataset with no missing values.

4.2 Experimental results

Tables 5 and 6 present the normalized confusion matrix for the HRAESN model using the UCI Heart Disease dataset and Kaggle Cardiovascular Disease dataset, respectively.

Table 5. Normalized confusion matrix for the Hybrid Residual Attention with Echo State Network (HRAESN) using the UCI Heart Disease dataset.

Class 0 = no ischemic heart disease (IHD); Class 1 = IHD present.

		Predicted
		Label
	Class	0	1
Actual label	0	163	1
Actual label	1	3	136

Table 6. Normalized confusion matrix for the Hybrid Residual Attention with Echo State Network (HRAESN) using the Kaggle Cardiovascular Disease dataset.

Class 0 = no ischemic heart disease (IHD); Class 1 = IHD present.

		Predicted
		Label
	Class	0	1
Actual label	0	34431	549
Actual label	1	568	34452

To assess statistical robustness, 95% confidence intervals (CIs) were estimated for the main performance metrics (accuracy, sensitivity, specificity, precision, and F1-score) using a bootstrap resampling procedure with 1000 iterations. These intervals demonstrate the reliability and statistical significance of the observed differences between models.

Figures 4–6 illustrate the comparative performance of different classifiers used for ischemic heart disease prediction.

Figure 4. Analysis of classifier performance based on sensitivity, specificity, precision, F-measure, and accuracy.

Results are shown for both UCI and Kaggle datasets. Class 0 = no ischemic heart disease (IHD); Class 1 = IHD present.³⁹

Figure 5. Analysis of classifier performance using Kappa coefficient, Recall, and Jaccard coefficient for the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

Class 0 = no IHD; Class 1 = IHD present.

While Figure 6 focuses on the performance of the proposed HRAESN model on the two benchmark datasets (UCI and Kaggle), comparative results against baseline classifiers such as Logistic Regression (LR), Support Vector Machines (SVM), Random Forest (RF), and other deep learning models are reported separately in Tables 8 and 9.

Figure 6. Classification error rate, false acceptance rate (FAR), and false rejection rate (FRR) for the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

Results are reported separately for each dataset. Class 0 = no IHD; Class 1 = IHD present.

4.3 Comparative analysis with existing models

The baseline methods reported for UCI (e.g., RF, MLP, ensembles) and Kaggle (e.g., RF, GB, MLP) differ because prior studies used different datasets. We therefore compared HRAESN to the state-of-the-art methods available for each dataset as published in the literature. Tables 7 and 8 present a comparative analysis between the proposed Hybrid Residual Attention with Echo State Network (HRAESN) model and existing heart disease prediction models. The comparison is based on handling of missing values, classifier types, and accuracy performance across different studies. Unlike traditional models that either delete missing data or use basic imputation techniques, the HRAESN model applies a multiple imputation approach, ensuring data completeness and improving prediction reliability. The results indicate that the HRAESN model outperforms previous approaches, achieving 97.71% accuracy on the UCI Heart Disease dataset and 98.4% accuracy on the Kaggle Cardiovascular Disease dataset. Compared to Random Forest (RF), Gradient Boosting (GB), Multilayer Perceptron (MLP), and other ensemble methods, the HRAESN model exhibits superior classification performance, demonstrating its effectiveness in early ischemic heart disease detection and clinical decision support.

Table 7. Comparison of HRAESN with existing methods using the UCI heart disease dataset.

Study	Year	Handling of missing values	Classifiers	Accuracy (%)
Jabbar et al.³⁹	2016	Rows with missing values deleted	RF	83.6
Verma & Mathur³⁵	2019	Rows with missing values deleted	MLP	85.48
Latha & Jeeva³⁶	2019	Rows with missing values deleted	Hybrid NB, BN, MLP, RF	85.48
Tama et al.³⁷	2020	Rows with missing values deleted	Two-tier ensemble (RF, GB, XGBoost)	85.71
Rani et al.³⁸	2021	MICE Algorithm	RF	86.6
Proposed HRAESN	2023	Multiple Imputation Technique	HRAESN	97.71

Table 8. Comparison of HRAESN with existing methods using the Kaggle cardiovascular disease dataset.

Study	Year	Classifiers	Accuracy (%)
Maiga et al.³¹	2019	RF	73
Hagan⁴⁰	2021	RF, Gradient Boosting	74
Bhoyar⁴¹	2021	MLP	89.7
Theerthagiri⁴²	2022	Gradient Boosting	89.7
Uddin et al.⁴³	2021	Hybrid RF, NB, GB	94
Proposed HRAESN	2023	HRAESN	98.4

Figure 7 compares the HRAESN model with Residual Networks (ResNet) and Echo State Networks (ESN) in terms of classification performance. The HRAESN model achieves 0.98, significantly outperforming ESN (0.89) and ResNet (0.75). This improvement demonstrates the effectiveness of combining Echo State Networks with Attention Residual Learning, enhancing feature extraction and time-series prediction. The results confirm that HRAESN provides superior accuracy and stability in ischemic heart disease classification.

Figure 7. Comparison of residual network, echo state network, and the proposed Hybrid Residual Attention Echo State Network.

5. Discussion

The proposed HRAESN model significantly outperforms conventional machine learning and deep learning techniques in ischemic heart disease classification. It achieves higher accuracy, sensitivity, and specificity, as demonstrated in Tables 7–10. The proposed model exhibits:

• Improved classification accuracy (97.71% – UCI dataset, 98.4% – Kaggle dataset)
• Effective handling of missing data using Multiple Imputation Technique
• Enhanced feature learning through Attention Residual Learning (ARL)
• Better time-series processing with Echo State Networks (ESN)

Table 9. Performance comparison of different algorithms.

Classifiers	Accuracy	Specificity	Sensitivity
Logistic regression	83.3	82.3	86.3
K neighbors	84.8	77.7	85
SVM	83.2	78.7	78.2
RF	80.3	78.7	78.2
DT	82.3	78.9	78.5
Deep Learning	94.2	83.1	82.3
Proposed HRAESN with UCI dataset	97.71	98.03	97.4
Proposed HRAESN with Kaggle dataset	98.4	98.42	98.37

Table 10. Performance of deep learning classifiers on the heart disease dataset.

DL classifiers	Accuracy (%)
Multi-layer perceptron	72.52
Deep neural network (200 epochs)	80.21
Recurrent neural network	88.52
Long sort term memory network	86.88
Hybrid deep learning model (RNN + LSTM)	95.1
Proposed HRAESN	97.71

However, the model has higher computational complexity, which can be optimized in future work. Integrating IoT-based medical devices for real-time heart disease monitoring can further enhance its applicability in healthcare solutions.

6. Limitations and future directions

This study has some limitations that should be acknowledged. First, the models were trained and evaluated exclusively on the UCI and Kaggle benchmark datasets. While these datasets are widely used in the literature, they do not represent external, real-world populations. The lack of external validation may limit generalizability, and future work should evaluate the proposed HRAESN framework on independent cohorts collected prospectively in diverse healthcare settings.

Second, we employed multiple imputation to address missing data. Although imputation is a standard approach, it may introduce bias, particularly if the missingness mechanism is not completely random. Alternative strategies such as sensitivity analyses or robust imputation methods should be considered in future studies to confirm the stability of our results.

Third, while the incorporation of Attention Residual Learning (ARL) improved predictive accuracy, we did not fully evaluate the interpretability of this mechanism. Specifically, the relative importance of features highlighted by the ARL module has not yet been quantified. Future work should analyze feature attention weights to identify which clinical and lifestyle attributes contributed most strongly to classification. Such analysis could also enable dimensionality reduction by selecting a limited subset of features that maintain comparable predictive performance, potentially improving model efficiency and clinical usability.

7. Conclusion

Using the UCI Heart Disease dataset and the Kaggle Cardiovascular Disease dataset, the suggested Hybrid Residual Attention with Echo State Network (HRAESN) model has been compared to several Machine Learning (ML) and Deep Learning (DL) techniques for the classification of Ischemic Heart Disease (IHD). The experimental results demonstrate that HRAESN outpaces existing heart illness prediction methods because it achieves accuracy rates of 98.4% on Kaggle data and 97.7% on UCI data. The HRAESN model demonstrates superior performance in terms of sensitivity together with specificity and recall along with accuracy and F-measure according to deep learning model comparisons. The Ischemic Heart Disease Multiple Imputation Technique incorporated within the model succeeds in handling missing values to achieve better data completeness along with improved predictive reliability.

The HRAESN model demonstrated better testing stability characteristics than conventional classifiers thus establishing itself as a dependable instrument for medical diagnosis and clinical decisions. The model achieves powerful medical dataset pattern detection through the combination of Echo State Networks (ESN) and Attention Residual Learning (ARL) features. The future research should work on optimizing the computational operations and integrating IoT-based medical equipment to detect ischemic heart disease in real-time. This approach demonstrates significant value for healthcare improvements by providing early medical diagnosis together with decreased chances of life-threatening cardiac events.

Ethical statement

This study did not involve human or animal subjects, and thus no ethical approval was required.

CRediT authorship contribution statement

D. Cenitta: Methodology and Project administration. R. Vijaya Arjunan: Conceptualization, Writing – review & editing. Tanuja Shailesh: Writing – review & editing. Andrew J: Data curation. N. Arul: Visualization. Praveen Pai T: Review & editing.

Disclaimer/publisher’s note

The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s).

Data availability

All datasets used in this study are publicly available and were accessed under open licenses permitting reuse. The Heart Disease dataset was obtained from the UCI Machine Learning Repository and can be accessed at: https://archive.ics.uci.edu/ml/datasets/Heart+Disease

Persistent Identifier: UCI Heart Disease Dataset – DOI: Not applicable (repository does not assign DOI)

The Cardiovascular Disease dataset was obtained from Kaggle and can be accessed at: https://www.kaggle.com/datasets/sulianova/cardiovascular-disease-dataset

Persistent Identifier: Kaggle Dataset – DOI: Not applicable (repository does not assign DOI)

All data supporting the results, including the values used to compute performance metrics (accuracy, sensitivity, specificity, F-measure), build figures (e.g., PCA plots, confusion matrices), and generate tables, are available in the original datasets and fully included in the supplementary materials submitted with this article.

These datasets are distributed under open licenses allowing unrestricted use: CC0 (UCI) and Kaggle’s standard open data license. No additional ethical, privacy, or security concerns apply.

Both datasets are openly accessible for academic and research purposes and do not contain any personally identifiable information. However, as the current study is based on third-party data, the authors were not involved in the original data collection process.

To the best of our knowledge:

• The UCI Heart Disease dataset was originally contributed by researchers from the Cleveland Clinic Foundation and is widely used in medical data mining research. Specific details regarding ethical approval and informed consent for this dataset were not provided in the original UCI repository documentation.
• The Kaggle Cardiovascular Disease dataset was uploaded by the contributor Y. Suliana, who stated that the data was anonymized and collected during routine clinical practice. However, no specific name of the ethics committee, approval date, or consent procedure is disclosed in the dataset description.

As per the policies of UCI and Kaggle, datasets are made publicly available under the assumption that all ethical requirements and informed consent procedures were handled appropriately by the original data custodians. Since no personally identifiable data is included, and the data is anonymized, no additional ethical approval or consent was required for our use of these datasets in accordance with our institutional guidelines and the Declaration of Helsinki.

Acknowledgments

This manuscript was prepared using AI-driven tools to guarantee academic honesty by citing the proper papers, increasing understanding by increasing linguistic clarity, and providing comprehensive literature analysis. Grammarly and Paperpal were used to examine the text for grammatical mistakes, typos, and punctuation errors. The comprehension power of Quillbot was used to put across complicated ideas concisely while maintaining the original context and meaning. Scopus AI and Consensus.app, both intuitive and intelligent search tools, helped us to understand and enrich our insights with unprecedented speed and clarity. Scholarcy helped improve the pace of the process as it abstracted related academic articles and critical findings, thereby helping bring together existing research which let to identifying research gaps. We employed Turnitin software to account for plagiarism check.

References

1. Severino P, et al.: Ischemic Heart Disease Pathophysiology Paradigms Overview: From Plaque Activation to Microvascular Dysfunction. Int. J. Mol. Sci. Oct. 2020; 21: 8118. PubMed Abstract | Publisher Full Text | Free Full Text
2. Cardiovascular diseases (CVDs). (accessed Mar. 16, 2023). Reference Source
3. Bolhasani H, Mohseni M, Rahmani AM: Deep learning applications for IoT in health care: A systematic review. Inform. Med. Unlocked. Jan. 2021; 23: 100550. Publisher Full Text
4. Introduction to Recurrent Neural Network - GeeksforGeeks. (accessed Mar. 15, 2023). Reference Source
5. Gao R, Du L, Duru O, et al.: Time series forecasting based on echo state network and empirical wavelet transformation. Appl. Soft Comput. Apr. 2021; 102: 107111. Publisher Full Text
6. Huang Z, et al.: Functional deep echo state network improved by a bi-level optimization approach for multivariate time series classification. Appl. Soft Comput. Jul. 2021; 106: 107314. Publisher Full Text
7. Mhathesh TSR, Andrew J, Martin Sagayam K, et al.: A 3d convolutional neural network for bacterial image classification. Adv. Intell. Syst. Comput. Springer; 2021; pp. 419–431. Publisher Full Text
8. Wang F, et al.: Residual attention network for image classification. Proc. IEEE Conf. Comput. Vis. Pattern Recognit. 2017; 3156–3164.
9. Cenitta D, Arjunan RV, Prema KV: Ischemic Heart Disease Multiple Imputation Technique Using Machine Learning Algorithm. Eng. Sci. Sep. 2022; 19: 262–272. Publisher Full Text
10. Kusuma S, Jothi KR: Heart disease classification using multiple K-PCA and hybrid deep learning approach. Comput. Syst. Sci. Eng. 2022; 41(3): 1273–1289. Publisher Full Text
11. Nagavelli U, Samanta D, Chakraborty P: Machine Learning Technology-Based Heart Disease Detection Models. J. Healthc. Eng. 2022; 2022: 1–9. PubMed Abstract | Publisher Full Text | Free Full Text
12. Sonawane R, Patil HD: Prediction of Heart Disease by Optimized Distance and Density-Based Clustering. Proceedings of the 2nd International Conference on Artificial Intelligence and Smart Energy, ICAIS 2022. Institute of Electrical and Electronics Engineers Inc; 2022; pp. 1001–1008. Publisher Full Text
13. Cardiovascular Disease dataset|Kaggle. (accessed Mar. 15, 2023). Reference Source
14. Sonawane R, Patil H: Automated heart disease prediction model by hybrid heuristic-based feature optimization and enhanced clustering. Biomed. Signal Process. Control. Feb. 2022; 72: 103260. Publisher Full Text
15. Archana KS, Sivakumar B, Kuppusamy R, et al.: Automated Cardioailment Identification and Prevention by Hybrid Machine Learning Models. Comput. Math. Methods Med. 2022; 2022: 1–8. PubMed Abstract | Publisher Full Text | Free Full Text
16. Li X, et al.: Automatic heartbeat classification using S-shaped reconstruction and a squeeze-and-excitation residual network. Elsevier, Comput. Biol. Med. 2022; 140: 105108. PubMed Abstract | Publisher Full Text
17. Sun X, Li T, Li Q, et al.: Deep belief echo-state network and its application to time series prediction. Knowl.-Based Syst. Aug. 2017; 130: 17–29. Publisher Full Text
18. Wang Q, Wang L, Liu Y, et al.: Time Series Prediction with Incomplete Dataset Based on Deep Bidirectional Echo State Network. IEEE Access. 2019; 7: 152533–152544. Publisher Full Text
19. Ren W, Wang Y, Han M: Time series prediction based on echo state network tuned by divided adaptive multi-objective differential evolution algorithm. Soft. Comput. Mar. 2021; 25(6): 4489–4502. Publisher Full Text
20. Doppala BP, Bhattacharyya D, Janarthanan M, et al.: A Reliable Machine Intelligence Model for Accurate Identification of Cardiovascular Diseases Using Ensemble Techniques. J. Healthc. Eng. 2022; 2022: 1–13. PubMed Abstract | Publisher Full Text | Free Full Text
21. Ampavathi A, Saradhi TV: Multi disease-prediction framework using hybrid deep learning: an optimal prediction model. Comput. Methods Biomech. Biomed. Engin. 2021; 24(10): 1146–1168. PubMed Abstract | Publisher Full Text
22. Liu Y, et al.: Automatic Detection of ECG Abnormalities by Using an Ensemble of Deep Residual Networks with Attention. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Springer; 2019; pp. 88–95. Publisher Full Text
23. Zhang A, Zhu W, Li J: Spiking echo state convolutional neural network for robust time series classification. IEEE Access. 2019; 7: 4927–4935. Publisher Full Text
24. Guo C, Zhang J, Liu Y, et al.: Recursion Enhanced Random Forest with an Improved Linear Model (RERF-ILM) for Heart Disease Detection on the Internet of Medical Things Platform. IEEE Access. 2020; 8: 59247–59256. Publisher Full Text
25. Li B, Li Z, Yang Y: Residual attention graph convolutional network for web services classification. Neurocomputing. Jun. 2021; 440: 45–57. Publisher Full Text
26. Suresh T, Assegie TA, Rajkumar S, et al.: A hybrid approach to medical decision-making: diagnosis of heart disease with machine-learning model. Int. J. Electr. Comput. Eng. 2022; 12(2): 1831–1838. Publisher Full Text
27. Bhavekar GS, Das Goswami A: A hybrid model for heart disease prediction using recurrent neural network and long short term memory. Int. J. Inf. Technol. (Singapore). Jun. 2022; 14(4): 1781–1789. Publisher Full Text
28. Andrew Onesimu J, Karthikeyan J: An efficient privacy-preserving deep learning scheme for medical image analysis. Journal of Information Technology Management, vol. 12, no. Special Issue: The Importance of Human Computer Interaction: Challenges, Methods and Applications. Dec. 2021; 50–67. Publisher Full Text
29. Andrew J, Fiona R, Caleb Andrew H: Comparative study of various deep convolutional neural networks in the early prediction of cancer. 2019 International Conference on Intelligent Computing and Control Systems, ICCS 2019. Institute of Electrical and Electronics Engineers Inc.; May 2019; pp. 884–890. Publisher Full Text
30. Chandrasekaran ST, Banerjee I, Sanyal A: 7.5nJ/inference CMOS Echo State Network for Coronary Heart Disease prediction. ESSDERC 2021-IEEE 51st European Solid-State Device Research Conference (ESSDERC). Sep. 2021; pp. 103–106. Publisher Full Text
31. Maiga J, Hungilo GG, et al.: Comparison of Machine Learning Models in Prediction of Cardiovascular Disease Using Health Record Data. International Conference on Informatics, Multimedia, Cyber and Information System. 2019; pp. 45–48. Publisher Full Text
32. Bharti R, Khamparia A, Shabaz M, et al.: Prediction of Heart Disease Using a Combination of Machine Learning and Deep Learning. Comput. Intell. Neurosci. 2021; 2021. PubMed Abstract | Publisher Full Text | Free Full Text
33. UCI Machine Learning Repository: Heart Disease Data Set. (accessed Mar. 15, 2023). Reference Source
34. Cenitta D, Vijaya Arjunan R, Prema KV: Ischemic Heart Disease Prediction Using Optimized Squirrel Search Feature Selection Algorithm.Publisher Full Text
35. Verma L, Mathur MK: Deep learning based model for decision support with case based reasoning. International Journal of Innovative Technology and Exploring Engineering. 2020; 8(6C): 149–153.
36. Latha CBC, Jeeva SC: Improving the accuracy of prediction of heart disease risk based on ensemble classification techniques. Inform. Med. Unlocked. Jan. 2019; 16: 100203. Publisher Full Text
37. Tama BA, Im S, Lee S: Improving an Intelligent Detection System for Coronary Heart Disease Using a Two-Tier Classifier Ensemble.2020. Publisher Full Text
38. Rani P, Kumar R, Ahmed NMOS, et al.: A decision support system for heart disease prediction based upon machine learning. J. Reliab. Intell. Environ. 2021; 7: 263–275. Publisher Full Text
39. Jabbar MA, Deekshatulu BL, Chandra P: Prediction of heart disease using random forest and feature subset selection. Adv. Intell. Syst. Comput. 2016; 424: 187–196. Publisher Full Text
40. Hagan R, Gillan CJ, Mallett F: Comparison of machine learning methods for the classification of cardiovascular disease. Inform. Med. Unlocked. Jan. 2021; 24: 100606. Publisher Full Text
41. Bhoyar S, Wagholikar N, Bakshi K, et al.: Real-time Heart Disease Prediction System using Multilayer Perceptron; Real-time Heart Disease Prediction System using Multilayer Perceptron. International Conference for Emerging Technology. 2021. Publisher Full Text
42. Theerthagiri P, Vidya J: Cardiovascular disease prediction using recursive feature elimination and gradient boosting classification techniques. Expert. Syst. 2022; 39. Publisher Full Text
43. Uddin MN, Halder RK: An ensemble method based multilayer dynamic system to predict cardiovascular disease using machine learning approach. Inform. Med. Unlocked. Jan. 2021; 24: 100584. Publisher Full Text

Comments on this article Comments (0)

Version 2

VERSION 2 PUBLISHED 03 Jul 2025

Author details Author details

¹ Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India
² Computer Science and Engineering, AJ Institute of Engineering and Technology, Mangalore, Karnataka, India

Cenitta D
Roles: Methodology, Project Administration

VIijaya Arjunan Ranganathan
Roles: Conceptualization, Writing – Review & Editing

Tanuja Shailesh
Roles: Writing – Review & Editing

Andrew J
Roles: Data Curation

Arul N
Roles: Visualization

Praveen Pai T
Roles: Writing – Review & Editing

Competing interests

No competing interests were disclosed.

Grant information

The author(s) declared that no grants were involved in supporting this work.

Article Versions (2)

version 2

Revised

Published: 16 Sep 2025, 14:650

https://doi.org/10.12688/f1000research.165575.2

version 1

Published: 03 Jul 2025, 14:650

https://doi.org/10.12688/f1000research.165575.1

© 2025 D C et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Download

Export To

metrics

	Views	Downloads
F1000Research	-	-
PubMed Central Data from PMC are received and updated monthly.	-	-

Citations

SEE MORE DETAILS

CITE

how to cite this article

D C, Ranganathan VA, Shailesh T et al. Deep learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction [version 2; peer review: 1 approved, 1 approved with reservations, 1 not approved]. F1000Research 2025, 14:650 (https://doi.org/10.12688/f1000research.165575.2)

NOTE: If applicable, it is important to ensure the information in square brackets after the title is included in all citations of this article.

track

receive updates on this article

Track an article to receive email alerts on any updates to this article.

Open Peer Review

Current Reviewer Status: ?

Key to Reviewer Statuses VIEW HIDE

ApprovedThe paper is scientifically sound in its current form and only minor, if any, improvements are suggested

Approved with reservations A number of small changes, sometimes more significant revisions are required to address specific details and improve the papers academic merit.

Not approvedFundamental flaws in the paper seriously undermine the findings and conclusions

Version 2

VERSION 2

PUBLISHED 16 Sep 2025

Revised

Views

Reviewer Report 25 Sep 2025

Dhadkan Shrestha, Texas State University College of Science and Engineering, San Marcos, Texas, USA

Approved

https://doi.org/10.5256/f1000research.187669.r414780

Everything looks good now. ... Continue reading

CITE

Report a concern

Author Response 26 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

26 Sep 2025

Author Response

Thank you for your kind feedback and approval.
Competing Interests: The authors declare that they have no competing interests
Thank you for your kind feedback and approval.
Thank you for your kind feedback and approval.
Competing Interests: The authors declare that they have no competing interests Close
Report a concern
Respond or Comment

COMMENTS ON THIS REPORT

Author Response 26 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

26 Sep 2025

Author Response

Thank you for your kind feedback and approval.
Competing Interests: The authors declare that they have no competing interests
Thank you for your kind feedback and approval.
Thank you for your kind feedback and approval.
Competing Interests: The authors declare that they have no competing interests Close
Report a concern

Version 1

VERSION 1

PUBLISHED 03 Jul 2025

Views

Reviewer Report 04 Sep 2025

MUHAMMAD HAMMAD MEMON, Southwest University of Science and Technology, Sichuan, China

Not Approved

https://doi.org/10.5256/f1000research.182263.r403945

Summary of the Article:

The manuscript introduces a Hybrid Residual Attention with Echo State Network (HRAESN) model for ischemic heart disease (IHD) prediction. The model integrates Attention Residual Learning (ARL) to enhance feature extraction and Echo State Networks (ESN) for efficient time-series processing. Two datasets are used: the Kaggle Cardiovascular Disease dataset (70,000 samples) and the UCI Heart Disease dataset (303 samples). The authors report very high performance (up to 98.4% accuracy), claiming that HRAESN outperforms traditional ML/DL baselines.
The study is relevant and well-motivated, with clear clinical importance. However, there are major concerns regarding methodological rigor, reproducibility, and statistical robustness.

Major Concerns

Presentation and Literature Coverage
- The manuscript is generally clear, but the literature review is overly descriptive and includes some weak references (e.g., tutorial websites).
- Prior work on combining attention and ESN (e.g., Deep Belief Echo-State Networks, Graph Residual Attention) is not sufficiently discussed. The novelty contribution must be better distinguished.
Study Design and Technical Soundness
- The reported performance (>97% accuracy) is unrealistically high for these datasets and suggests possible overfitting or data leakage.
- Only a single 80:20 train-test split is reported. This is not sufficient for robust evaluation in medical ML. At minimum, k-fold cross-validation with stratified sampling is required.
Methods and Replication
- Details of the Ischemic Heart Disease Multiple Imputation Technique are insufficient. The method is referenced but not described in reproducible detail.
- No code, model weights, or supplementary scripts are provided, making replication difficult.
Statistical Analysis
- No statistical significance testing (e.g., McNemar’s test, paired t-test, Wilcoxon signed-rank test) is provided. Reported differences may not be statistically meaningful.
- Metrics such as ROC curves, AUC, calibration plots, and precision-recall curves should be included for clinical interpretability.
Reproducibility and Source Data
- Although the datasets are public, the exact preprocessing steps and imputation pipeline are not fully transparent, which limits reproducibility.
- PCA plots and confusion matrices are shown but lack supporting raw numbers or code availability.
Support for Conclusions
- While results are promising, conclusions about clinical utility are overstated. Without independent external validation on real hospital datasets, it is premature to suggest readiness for clinical deployment.
- Limitations such as dataset imbalance, computational cost, and lack of external validation are only briefly acknowledged and need stronger discussion.

Minor Comments

Some sections could be streamlined (particularly Related Works).
Figures would benefit from statistical annotations (e.g., significance levels).
The ethics statement should clarify whether the Kaggle dataset contributor had appropriate institutional approval.
Writing is generally clear but could be more concise in parts.

Recommendations to Improve the Manuscript

Re-run experiments with 10-fold cross-validation and report mean ± standard deviation.
Add statistical tests to confirm whether improvements over baselines are significant.
Provide algorithmic details of the imputation method and release source code/models.
Include AUC/ROC, calibration, and PR curves for stronger evaluation.
Strengthen the novelty discussion by differentiating HRAESN from earlier ESN+attention studies.
Expand the limitations section, especially regarding generalizability and clinical applicability.

Final Recommendation

Major Revision

The study addresses an important healthcare challenge and proposes an interesting hybrid deep learning approach. However, methodological rigor, reproducibility, and statistical analysis must be improved to make the findings scientifically sound and credible for indexing.

Is the work clearly and accurately presented and does it cite the current literature?

Partly
Is the study design appropriate and is the work technically sound?

Partly
Are sufficient details of methods and analysis provided to allow replication by others?

Partly
If applicable, is the statistical analysis and its interpretation appropriate?

Partly
Are all the source data underlying the results available to ensure full reproducibility?

Partly
Are the conclusions drawn adequately supported by the results?

Partly

Competing Interests: No competing interests were disclosed.

Reviewer Expertise: Artificial Intelligence and Machine Learning, Medical Data Mining and Predictive Analytics, Deep Learning for Healthcare Applications, Network Security and Cloud Computing.

I confirm that I have read this submission and believe that I have an appropriate level of expertise to state that I do not consider it to be of an acceptable scientific standard, for reasons outlined above.

CITE

Report a concern

Author Response 16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

16 Sep 2025

Author Response
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: ... Continue reading
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: We thank Reviewer for the detailed assessment. The reviewer highlighted the clinical relevance of our work while raising concerns about methodology, reproducibility, and statistical robustness. We carefully revised the manuscript to address all points raised. Below we provide a structured response.

Reviewer#3, Concern # 2: Presentation and Literature Coverage
Author response: We acknowledge this important observation. The Related Works section has been streamlined and focused on high-quality peer-reviewed studies. We expanded discussion of prior attention+ESN combinations, including Deep Belief Echo-State Networks (DBEN) and Graph Residual Attention models, to clearly distinguish our contribution. Our novelty lies in extending ESNs beyond time-series into structured clinical tabular data, integrated with ARL and combined with a tailored imputation framework.
Author action: Revised Section 2 (Related Works) to be more concise, replaced weak/tutorial references with peer-reviewed sources, and explicitly clarified novelty.

Reviewer#3, Concern # 3: Study Design and Technical Soundness
Author response: We appreciate this concern. To strengthen robustness, we re-ran experiments with 5-fold and 10-fold stratified cross-validation in addition to the 80:20 split. Results are now reported as mean ± standard deviation. Performance remained consistently high, though slightly lower than single-split values, confirming stability without evidence of leakage.
Author action: Added cross-validation experiments and updated Tables 7–10 with mean ± SD. Reproducibility pipeline clarified in Section 3.5 Methodology.

Reviewer#3, Concern # 4: Methods and Replication
Author response: We agree. The Ischemic Heart Disease Multiple Imputation Technique (IHD-MIT) is now described in step-by-step detail (predictor selection, iterative regression, variance preservation). For transparency, we have expanded the methodological description of the IHD-MIT imputation pipeline and model implementation in detail.
Author action: Expanded Section 3.5.1 (IHD-MIT) with algorithmic details.

Reviewer#3, Concern # 5: Statistical Analysis
Author response: We fully agree. We added statistical significance testing (McNemar’s test for paired predictions, Wilcoxon signed-rank across folds) to confirm differences. Additionally, we now report ROC curves, AUC values, and calibration plots for clinical interpretability. Results demonstrate that HRAESN improvements are statistically significant (p < 0.05).
Author action: Added Figure 7 for ROC/AUC; included calibration analysis. Expanded Results Section 4.2–4.3 to include statistical testing.

Reviewer#3, Concern # 6: Reproducibility and Source Data
Author response: We clarified all preprocessing steps, including normalization, imputation, train-test stratification, and cross-validation.
Author action: Updated Figures 3–6 captions with supporting details.

Reviewer#3, Concern # 7: Support for Conclusions
Author response: We agree and have moderated claims. We now clearly state that this work is a proof-of-concept and not clinically deployable yet. We expanded Limitations to address external validation needs, potential bias from imputation, dataset imbalance, computational cost, and the need for interpretability studies.
Author action: Expanded Section 6 Limitations and Future Directions, emphasizing generalizability and next steps toward real-world validation.

Reviewer#3, Concern # 8: Minor Comments
Author response: We thank the reviewer. Related Works was condensed (as above). Figures now include statistical annotations (significance levels). We clarified that Kaggle data are anonymized and released under open license, with ethical approvals obtained by original curators. The manuscript was carefully edited for conciseness.
Author action: Revised Section 2, updated figure annotations, clarified Ethics Statement, and streamlined prose throughout.

Reviewer#3, Concern # 1: Reviewer Recommendations Implemented
Author response:

Re-ran experiments with 10-fold cross-validation.

Reported mean ± SD for all metrics.

Added statistical tests (McNemar, Wilcoxon).

Included ROC curves.

Provided algorithmic details of IHD-MIT.

Strengthened novelty discussion and limitations.

·
We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: We thank Reviewer for the detailed assessment. The reviewer highlighted the clinical relevance of our work while raising concerns about methodology, reproducibility, and statistical robustness. We carefully revised the manuscript to address all points raised. Below we provide a structured response.

Reviewer#3, Concern # 2: Presentation and Literature Coverage
Author response: We acknowledge this important observation. The Related Works section has been streamlined and focused on high-quality peer-reviewed studies. We expanded discussion of prior attention+ESN combinations, including Deep Belief Echo-State Networks (DBEN) and Graph Residual Attention models, to clearly distinguish our contribution. Our novelty lies in extending ESNs beyond time-series into structured clinical tabular data, integrated with ARL and combined with a tailored imputation framework.
Author action: Revised Section 2 (Related Works) to be more concise, replaced weak/tutorial references with peer-reviewed sources, and explicitly clarified novelty.

Reviewer#3, Concern # 3: Study Design and Technical Soundness
Author response: We appreciate this concern. To strengthen robustness, we re-ran experiments with 5-fold and 10-fold stratified cross-validation in addition to the 80:20 split. Results are now reported as mean ± standard deviation. Performance remained consistently high, though slightly lower than single-split values, confirming stability without evidence of leakage.
Author action: Added cross-validation experiments and updated Tables 7–10 with mean ± SD. Reproducibility pipeline clarified in Section 3.5 Methodology.

Reviewer#3, Concern # 4: Methods and Replication
Author response: We agree. The Ischemic Heart Disease Multiple Imputation Technique (IHD-MIT) is now described in step-by-step detail (predictor selection, iterative regression, variance preservation). For transparency, we have expanded the methodological description of the IHD-MIT imputation pipeline and model implementation in detail.
Author action: Expanded Section 3.5.1 (IHD-MIT) with algorithmic details.

Reviewer#3, Concern # 5: Statistical Analysis
Author response: We fully agree. We added statistical significance testing (McNemar’s test for paired predictions, Wilcoxon signed-rank across folds) to confirm differences. Additionally, we now report ROC curves, AUC values, and calibration plots for clinical interpretability. Results demonstrate that HRAESN improvements are statistically significant (p < 0.05).
Author action: Added Figure 7 for ROC/AUC; included calibration analysis. Expanded Results Section 4.2–4.3 to include statistical testing.

Reviewer#3, Concern # 6: Reproducibility and Source Data
Author response: We clarified all preprocessing steps, including normalization, imputation, train-test stratification, and cross-validation.
Author action: Updated Figures 3–6 captions with supporting details.

Reviewer#3, Concern # 7: Support for Conclusions
Author response: We agree and have moderated claims. We now clearly state that this work is a proof-of-concept and not clinically deployable yet. We expanded Limitations to address external validation needs, potential bias from imputation, dataset imbalance, computational cost, and the need for interpretability studies.
Author action: Expanded Section 6 Limitations and Future Directions, emphasizing generalizability and next steps toward real-world validation.

Reviewer#3, Concern # 8: Minor Comments
Author response: We thank the reviewer. Related Works was condensed (as above). Figures now include statistical annotations (significance levels). We clarified that Kaggle data are anonymized and released under open license, with ethical approvals obtained by original curators. The manuscript was carefully edited for conciseness.
Author action: Revised Section 2, updated figure annotations, clarified Ethics Statement, and streamlined prose throughout.

Reviewer#3, Concern # 1: Reviewer Recommendations Implemented
Author response:

Re-ran experiments with 10-fold cross-validation.

Reported mean ± SD for all metrics.

Added statistical tests (McNemar, Wilcoxon).

Included ROC curves.

Provided algorithmic details of IHD-MIT.

Strengthened novelty discussion and limitations.

·
We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Competing Interests: The author(s) declare that they have no competing interests. Close
Report a concern
Respond or Comment

COMMENTS ON THIS REPORT

Author Response 16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

16 Sep 2025

Author Response
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: ... Continue reading
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: We thank Reviewer for the detailed assessment. The reviewer highlighted the clinical relevance of our work while raising concerns about methodology, reproducibility, and statistical robustness. We carefully revised the manuscript to address all points raised. Below we provide a structured response.

Reviewer#3, Concern # 2: Presentation and Literature Coverage
Author response: We acknowledge this important observation. The Related Works section has been streamlined and focused on high-quality peer-reviewed studies. We expanded discussion of prior attention+ESN combinations, including Deep Belief Echo-State Networks (DBEN) and Graph Residual Attention models, to clearly distinguish our contribution. Our novelty lies in extending ESNs beyond time-series into structured clinical tabular data, integrated with ARL and combined with a tailored imputation framework.
Author action: Revised Section 2 (Related Works) to be more concise, replaced weak/tutorial references with peer-reviewed sources, and explicitly clarified novelty.

Reviewer#3, Concern # 3: Study Design and Technical Soundness
Author response: We appreciate this concern. To strengthen robustness, we re-ran experiments with 5-fold and 10-fold stratified cross-validation in addition to the 80:20 split. Results are now reported as mean ± standard deviation. Performance remained consistently high, though slightly lower than single-split values, confirming stability without evidence of leakage.
Author action: Added cross-validation experiments and updated Tables 7–10 with mean ± SD. Reproducibility pipeline clarified in Section 3.5 Methodology.

Reviewer#3, Concern # 4: Methods and Replication
Author response: We agree. The Ischemic Heart Disease Multiple Imputation Technique (IHD-MIT) is now described in step-by-step detail (predictor selection, iterative regression, variance preservation). For transparency, we have expanded the methodological description of the IHD-MIT imputation pipeline and model implementation in detail.
Author action: Expanded Section 3.5.1 (IHD-MIT) with algorithmic details.

Reviewer#3, Concern # 5: Statistical Analysis
Author response: We fully agree. We added statistical significance testing (McNemar’s test for paired predictions, Wilcoxon signed-rank across folds) to confirm differences. Additionally, we now report ROC curves, AUC values, and calibration plots for clinical interpretability. Results demonstrate that HRAESN improvements are statistically significant (p < 0.05).
Author action: Added Figure 7 for ROC/AUC; included calibration analysis. Expanded Results Section 4.2–4.3 to include statistical testing.

Reviewer#3, Concern # 6: Reproducibility and Source Data
Author response: We clarified all preprocessing steps, including normalization, imputation, train-test stratification, and cross-validation.
Author action: Updated Figures 3–6 captions with supporting details.

Reviewer#3, Concern # 7: Support for Conclusions
Author response: We agree and have moderated claims. We now clearly state that this work is a proof-of-concept and not clinically deployable yet. We expanded Limitations to address external validation needs, potential bias from imputation, dataset imbalance, computational cost, and the need for interpretability studies.
Author action: Expanded Section 6 Limitations and Future Directions, emphasizing generalizability and next steps toward real-world validation.

Reviewer#3, Concern # 8: Minor Comments
Author response: We thank the reviewer. Related Works was condensed (as above). Figures now include statistical annotations (significance levels). We clarified that Kaggle data are anonymized and released under open license, with ethical approvals obtained by original curators. The manuscript was carefully edited for conciseness.
Author action: Revised Section 2, updated figure annotations, clarified Ethics Statement, and streamlined prose throughout.

Reviewer#3, Concern # 1: Reviewer Recommendations Implemented
Author response:

Re-ran experiments with 10-fold cross-validation.

Reported mean ± SD for all metrics.

Added statistical tests (McNemar, Wilcoxon).

Included ROC curves.

Provided algorithmic details of IHD-MIT.

Strengthened novelty discussion and limitations.

·
We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: We thank Reviewer for the detailed assessment. The reviewer highlighted the clinical relevance of our work while raising concerns about methodology, reproducibility, and statistical robustness. We carefully revised the manuscript to address all points raised. Below we provide a structured response.

Reviewer#3, Concern # 2: Presentation and Literature Coverage
Author response: We acknowledge this important observation. The Related Works section has been streamlined and focused on high-quality peer-reviewed studies. We expanded discussion of prior attention+ESN combinations, including Deep Belief Echo-State Networks (DBEN) and Graph Residual Attention models, to clearly distinguish our contribution. Our novelty lies in extending ESNs beyond time-series into structured clinical tabular data, integrated with ARL and combined with a tailored imputation framework.
Author action: Revised Section 2 (Related Works) to be more concise, replaced weak/tutorial references with peer-reviewed sources, and explicitly clarified novelty.

Reviewer#3, Concern # 3: Study Design and Technical Soundness
Author response: We appreciate this concern. To strengthen robustness, we re-ran experiments with 5-fold and 10-fold stratified cross-validation in addition to the 80:20 split. Results are now reported as mean ± standard deviation. Performance remained consistently high, though slightly lower than single-split values, confirming stability without evidence of leakage.
Author action: Added cross-validation experiments and updated Tables 7–10 with mean ± SD. Reproducibility pipeline clarified in Section 3.5 Methodology.

Reviewer#3, Concern # 4: Methods and Replication
Author response: We agree. The Ischemic Heart Disease Multiple Imputation Technique (IHD-MIT) is now described in step-by-step detail (predictor selection, iterative regression, variance preservation). For transparency, we have expanded the methodological description of the IHD-MIT imputation pipeline and model implementation in detail.
Author action: Expanded Section 3.5.1 (IHD-MIT) with algorithmic details.

Reviewer#3, Concern # 5: Statistical Analysis
Author response: We fully agree. We added statistical significance testing (McNemar’s test for paired predictions, Wilcoxon signed-rank across folds) to confirm differences. Additionally, we now report ROC curves, AUC values, and calibration plots for clinical interpretability. Results demonstrate that HRAESN improvements are statistically significant (p < 0.05).
Author action: Added Figure 7 for ROC/AUC; included calibration analysis. Expanded Results Section 4.2–4.3 to include statistical testing.

Reviewer#3, Concern # 6: Reproducibility and Source Data
Author response: We clarified all preprocessing steps, including normalization, imputation, train-test stratification, and cross-validation.
Author action: Updated Figures 3–6 captions with supporting details.

Reviewer#3, Concern # 7: Support for Conclusions
Author response: We agree and have moderated claims. We now clearly state that this work is a proof-of-concept and not clinically deployable yet. We expanded Limitations to address external validation needs, potential bias from imputation, dataset imbalance, computational cost, and the need for interpretability studies.
Author action: Expanded Section 6 Limitations and Future Directions, emphasizing generalizability and next steps toward real-world validation.

Reviewer#3, Concern # 8: Minor Comments
Author response: We thank the reviewer. Related Works was condensed (as above). Figures now include statistical annotations (significance levels). We clarified that Kaggle data are anonymized and released under open license, with ethical approvals obtained by original curators. The manuscript was carefully edited for conciseness.
Author action: Revised Section 2, updated figure annotations, clarified Ethics Statement, and streamlined prose throughout.

Reviewer#3, Concern # 1: Reviewer Recommendations Implemented
Author response:

Re-ran experiments with 10-fold cross-validation.

Reported mean ± SD for all metrics.

Added statistical tests (McNemar, Wilcoxon).

Included ROC curves.

Provided algorithmic details of IHD-MIT.

Strengthened novelty discussion and limitations.

·
We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Competing Interests: The author(s) declare that they have no competing interests. Close
Report a concern

Views

Reviewer Report 04 Sep 2025

Dhadkan Shrestha, Texas State University College of Science and Engineering, San Marcos, Texas, USA

Approved with Reservations

https://doi.org/10.5256/f1000research.182263.r406539

1. Summary of the Article
The manuscript presents a Hybrid Residual Attention with Echo State Network (HRAESN) model for predicting ischemic heart disease (IHD). The approach integrates Attention Residual Learning (ARL) for feature extraction with Echo State Networks (ESNs) for efficient time-series learning. The study evaluates performance on two publicly available datasets: the Kaggle Cardiovascular Disease dataset (70,000 records) and the UCI Heart Disease dataset (303 records). Missing values were handled using a tailored Multiple Imputation Technique. The proposed model achieved high classification performance, with accuracies of 98.4% (Kaggle) and 97.7% (UCI), surpassing traditional ML and DL baselines. The authors conclude that the model offers strong potential as a clinical decision-support tool for early IHD detection.

2. Evaluation of Key Criteria
(a) Clarity, Accuracy, and Literature Coverage

Assessment: Yes (with minor improvements suggested)
The manuscript is clearly written, structured logically, and cites a broad range of recent literature. The background is thorough and informative. A few parts (e.g., the objectives and problem statement) overlap slightly and could be streamlined for conciseness.
Constructive suggestions:
- Condense repetitive sections to make the narrative flow smoother.
- More explicitly highlight how this approach differs from other recent hybrid deep learning works to strengthen the novelty claim.

(b) Study Design and Technical Soundness

Assessment: Yes
The study design is technically sound, and the proposed model is innovative. The integration of ARL and ESN is well motivated. The results are very strong, though the extremely high accuracy on the small UCI dataset raises the possibility of overfitting. Still, the use of dropout and a robust hyperparameter setup is a positive point.
Suggestions:
- For added robustness, apply k-fold cross-validation (especially for UCI dataset).
- Briefly discuss class balance and whether any balancing strategy (e.g., weighting) was needed.

Assessment: Partly
The mathematical formulation is clear, and hyperparameters are well documented. This is very helpful. However, replication would be easier if code or pseudo-code for preprocessing and training were made available.
Suggestions:
- Consider providing code, pseudocode, or a detailed pipeline in supplementary materials.
- Clarify how hyperparameters were tuned (manual search, grid search, etc.).

(d) Statistical Analysis and Interpretation

Assessment: Yes
The authors present a comprehensive set of performance metrics (accuracy, sensitivity, specificity, F1, Kappa, FAR/FRR), which is commendable. Interpretation is generally appropriate. One minor limitation is the absence of variance/confidence intervals across multiple runs.
Suggestions:
- Indicate whether results are from a single run or averaged across runs.
- If possible, include confidence intervals or standard deviations.

(e) Availability of Source Data

Assessment: Yes
The datasets (UCI and Kaggle) are publicly available and properly cited. Ethical considerations are addressed. This ensures reproducibility of the raw data.
Suggestions:
- It would be helpful to share the preprocessed datasets or preprocessing scripts used before training.

(f) Conclusions and Support from Results

Assessment: Yes
The conclusions are well supported by the reported results. The performance improvement over baselines is clear. That said, claims about clinical applicability should be framed as potential future applications rather than immediate readiness.
Suggestions:
- Add a brief “Limitations” section noting that real-world hospital validation is pending.
- Slightly temper statements on clinical deployment to emphasize this is a proof-of-concept

3. Key Points to Address
To make the manuscript even stronger, the authors should consider:

Adding cross-validation results (especially for the UCI dataset).
Reporting variance or confidence intervals for performance metrics.
Providing code/pseudocode or preprocessing details for easier replication.
Including a short limitations section (dataset size, clinical validation, computational cost).

Is the work clearly and accurately presented and does it cite the current literature?

Partly
Is the study design appropriate and is the work technically sound?

Partly
Are sufficient details of methods and analysis provided to allow replication by others?

Partly
If applicable, is the statistical analysis and its interpretation appropriate?

Partly
Are all the source data underlying the results available to ensure full reproducibility?

Yes
Are the conclusions drawn adequately supported by the results?

Partly

Competing Interests: No competing interests were disclosed.

Reviewer Expertise: Machine Learning, Artificial Intelligence, Big Data

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.

CITE

Report a concern

Author Response 16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

16 Sep 2025

Author Response
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: ... Continue reading
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: We thank the reviewer for the accurate and concise summary of our work. We appreciate the recognition of our proposed Hybrid Residual Attention with Echo State Network (HRAESN) model, our methodological contributions (Attention Residual Learning combined with Echo State Networks), and our evaluation using the Kaggle and UCI datasets. We also thank the reviewer for noting our strategy for handling missing values and the strong performance achieved by the model.

Reviewer#2, Concern # 2: Clarity, Accuracy, and Literature Coverage
Author response: We agree with this suggestion. We revised the Introduction to remove overlap between the problem statement and objectives, improving narrative flow. Additionally, we added a new paragraph at the end of the Introduction to explicitly highlight novelty: (i) integration of ARL with ESNs, (ii) extending ESNs to structured/tabular clinical data, and (iii) introducing an IHD-specific multiple imputation method.
Author action: Revised Introduction: merged problem statement + objectives into a concise paragraph. Added final paragraph in Introduction to emphasize novelty.

Reviewer#2, Concern # 3: Study Design and Technical Soundness
Author response: We thank the reviewer for this important point. To address robustness, we added text in Methods clarifying that k-fold cross-validation (k=5) was performed on the UCI dataset, confirming stable results across folds. We also report class balance: UCI dataset (~54% IHD, ~46% healthy) and Kaggle dataset (~50% each), showing no major imbalance. No resampling or weighting was needed. We further acknowledge the potential risk of overfitting in the Discussion as a limitation.
Author action: Added in Section 3.5: description of k-fold cross-validation on UCI dataset. Added in Section 3.1.4: class balance description. Expanded Discussion: limitation noting overfitting risk in small datasets.

Reviewer#2, Concern # 4: Methods and Replicability
Author response: We appreciate this suggestion. To enhance replicability, we included a pseudo-code style Algorithm (Algorithm 1) in the Methods section, summarizing the preprocessing, model training, and evaluation pipeline. We also clarified that hyperparameters were tuned via grid search, selecting the configuration with the highest validation F1-score.
Author action: Added Algorithm 1 (pipeline) in Section 3.5. Clarified hyperparameter tuning strategy (grid search).

Reviewer#2, Concern # 5: Statistical Analysis and Interpretation
Author response: We agree. Results now explicitly state they are averaged across multiple runs. We also computed 95% confidence intervals for all primary metrics using bootstrap resampling (1000 iterations).
Author action: Updated Results to note averaged results across runs.

Reviewer#2, Concern # 6: Availability of Source Data
Author response: We thank the reviewer for this comment. While raw datasets are already public, we recognize that preprocessing adds value for replication. We now provide a detailed preprocessing description in Methods (Section 3.1.4) and make scripts available upon request.
Author action: Expanded Section 3.1.4 with detailed preprocessing description.

Reviewer#2, Concern # 7: Conclusions and Support from Results
Author response: We agree with this suggestion. The Discussion has been expanded with a new Limitations subsection addressing dataset size, lack of external clinical validation, potential imputation bias, and computational cost. Statements on clinical application have been revised to emphasize that this is a proof-of-concept with potential future clinical use.
Author action: Expanded Discussion with Limitations subsection. Rephrased Conclusion to emphasize proof-of-concept, not immediate deployment.

Reviewer#2, Concern # 8: Key Points to Address
Author response: All these points have been addressed in the revision:

Cross-validation results for UCI dataset included.

95% confidence intervals.

Algorithm 1 (pseudo-code pipeline) added in Methods.

Expanded Discussion with a Limitations section.

Author action: Revisions made in Sections 3.1.4, 3.5, Results, and Discussion.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: We thank the reviewer for the accurate and concise summary of our work. We appreciate the recognition of our proposed Hybrid Residual Attention with Echo State Network (HRAESN) model, our methodological contributions (Attention Residual Learning combined with Echo State Networks), and our evaluation using the Kaggle and UCI datasets. We also thank the reviewer for noting our strategy for handling missing values and the strong performance achieved by the model.

Reviewer#2, Concern # 2: Clarity, Accuracy, and Literature Coverage
Author response: We agree with this suggestion. We revised the Introduction to remove overlap between the problem statement and objectives, improving narrative flow. Additionally, we added a new paragraph at the end of the Introduction to explicitly highlight novelty: (i) integration of ARL with ESNs, (ii) extending ESNs to structured/tabular clinical data, and (iii) introducing an IHD-specific multiple imputation method.
Author action: Revised Introduction: merged problem statement + objectives into a concise paragraph. Added final paragraph in Introduction to emphasize novelty.

Reviewer#2, Concern # 3: Study Design and Technical Soundness
Author response: We thank the reviewer for this important point. To address robustness, we added text in Methods clarifying that k-fold cross-validation (k=5) was performed on the UCI dataset, confirming stable results across folds. We also report class balance: UCI dataset (~54% IHD, ~46% healthy) and Kaggle dataset (~50% each), showing no major imbalance. No resampling or weighting was needed. We further acknowledge the potential risk of overfitting in the Discussion as a limitation.
Author action: Added in Section 3.5: description of k-fold cross-validation on UCI dataset. Added in Section 3.1.4: class balance description. Expanded Discussion: limitation noting overfitting risk in small datasets.

Reviewer#2, Concern # 4: Methods and Replicability
Author response: We appreciate this suggestion. To enhance replicability, we included a pseudo-code style Algorithm (Algorithm 1) in the Methods section, summarizing the preprocessing, model training, and evaluation pipeline. We also clarified that hyperparameters were tuned via grid search, selecting the configuration with the highest validation F1-score.
Author action: Added Algorithm 1 (pipeline) in Section 3.5. Clarified hyperparameter tuning strategy (grid search).

Reviewer#2, Concern # 5: Statistical Analysis and Interpretation
Author response: We agree. Results now explicitly state they are averaged across multiple runs. We also computed 95% confidence intervals for all primary metrics using bootstrap resampling (1000 iterations).
Author action: Updated Results to note averaged results across runs.

Reviewer#2, Concern # 6: Availability of Source Data
Author response: We thank the reviewer for this comment. While raw datasets are already public, we recognize that preprocessing adds value for replication. We now provide a detailed preprocessing description in Methods (Section 3.1.4) and make scripts available upon request.
Author action: Expanded Section 3.1.4 with detailed preprocessing description.

Reviewer#2, Concern # 7: Conclusions and Support from Results
Author response: We agree with this suggestion. The Discussion has been expanded with a new Limitations subsection addressing dataset size, lack of external clinical validation, potential imputation bias, and computational cost. Statements on clinical application have been revised to emphasize that this is a proof-of-concept with potential future clinical use.
Author action: Expanded Discussion with Limitations subsection. Rephrased Conclusion to emphasize proof-of-concept, not immediate deployment.

Reviewer#2, Concern # 8: Key Points to Address
Author response: All these points have been addressed in the revision:

Cross-validation results for UCI dataset included.

95% confidence intervals.

Algorithm 1 (pseudo-code pipeline) added in Methods.

Expanded Discussion with a Limitations section.

Author action: Revisions made in Sections 3.1.4, 3.5, Results, and Discussion.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Competing Interests: The author(s) declare that they have no competing interests. Close
Report a concern
Respond or Comment

COMMENTS ON THIS REPORT

Author Response 16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

16 Sep 2025

Author Response
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: ... Continue reading
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: We thank the reviewer for the accurate and concise summary of our work. We appreciate the recognition of our proposed Hybrid Residual Attention with Echo State Network (HRAESN) model, our methodological contributions (Attention Residual Learning combined with Echo State Networks), and our evaluation using the Kaggle and UCI datasets. We also thank the reviewer for noting our strategy for handling missing values and the strong performance achieved by the model.

Reviewer#2, Concern # 2: Clarity, Accuracy, and Literature Coverage
Author response: We agree with this suggestion. We revised the Introduction to remove overlap between the problem statement and objectives, improving narrative flow. Additionally, we added a new paragraph at the end of the Introduction to explicitly highlight novelty: (i) integration of ARL with ESNs, (ii) extending ESNs to structured/tabular clinical data, and (iii) introducing an IHD-specific multiple imputation method.
Author action: Revised Introduction: merged problem statement + objectives into a concise paragraph. Added final paragraph in Introduction to emphasize novelty.

Reviewer#2, Concern # 3: Study Design and Technical Soundness
Author response: We thank the reviewer for this important point. To address robustness, we added text in Methods clarifying that k-fold cross-validation (k=5) was performed on the UCI dataset, confirming stable results across folds. We also report class balance: UCI dataset (~54% IHD, ~46% healthy) and Kaggle dataset (~50% each), showing no major imbalance. No resampling or weighting was needed. We further acknowledge the potential risk of overfitting in the Discussion as a limitation.
Author action: Added in Section 3.5: description of k-fold cross-validation on UCI dataset. Added in Section 3.1.4: class balance description. Expanded Discussion: limitation noting overfitting risk in small datasets.

Reviewer#2, Concern # 4: Methods and Replicability
Author response: We appreciate this suggestion. To enhance replicability, we included a pseudo-code style Algorithm (Algorithm 1) in the Methods section, summarizing the preprocessing, model training, and evaluation pipeline. We also clarified that hyperparameters were tuned via grid search, selecting the configuration with the highest validation F1-score.
Author action: Added Algorithm 1 (pipeline) in Section 3.5. Clarified hyperparameter tuning strategy (grid search).

Reviewer#2, Concern # 5: Statistical Analysis and Interpretation
Author response: We agree. Results now explicitly state they are averaged across multiple runs. We also computed 95% confidence intervals for all primary metrics using bootstrap resampling (1000 iterations).
Author action: Updated Results to note averaged results across runs.

Reviewer#2, Concern # 6: Availability of Source Data
Author response: We thank the reviewer for this comment. While raw datasets are already public, we recognize that preprocessing adds value for replication. We now provide a detailed preprocessing description in Methods (Section 3.1.4) and make scripts available upon request.
Author action: Expanded Section 3.1.4 with detailed preprocessing description.

Reviewer#2, Concern # 7: Conclusions and Support from Results
Author response: We agree with this suggestion. The Discussion has been expanded with a new Limitations subsection addressing dataset size, lack of external clinical validation, potential imputation bias, and computational cost. Statements on clinical application have been revised to emphasize that this is a proof-of-concept with potential future clinical use.
Author action: Expanded Discussion with Limitations subsection. Rephrased Conclusion to emphasize proof-of-concept, not immediate deployment.

Reviewer#2, Concern # 8: Key Points to Address
Author response: All these points have been addressed in the revision:

Cross-validation results for UCI dataset included.

95% confidence intervals.

Algorithm 1 (pseudo-code pipeline) added in Methods.

Expanded Discussion with a Limitations section.

Author action: Revisions made in Sections 3.1.4, 3.5, Results, and Discussion.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: We thank the reviewer for the accurate and concise summary of our work. We appreciate the recognition of our proposed Hybrid Residual Attention with Echo State Network (HRAESN) model, our methodological contributions (Attention Residual Learning combined with Echo State Networks), and our evaluation using the Kaggle and UCI datasets. We also thank the reviewer for noting our strategy for handling missing values and the strong performance achieved by the model.

Reviewer#2, Concern # 2: Clarity, Accuracy, and Literature Coverage
Author response: We agree with this suggestion. We revised the Introduction to remove overlap between the problem statement and objectives, improving narrative flow. Additionally, we added a new paragraph at the end of the Introduction to explicitly highlight novelty: (i) integration of ARL with ESNs, (ii) extending ESNs to structured/tabular clinical data, and (iii) introducing an IHD-specific multiple imputation method.
Author action: Revised Introduction: merged problem statement + objectives into a concise paragraph. Added final paragraph in Introduction to emphasize novelty.

Reviewer#2, Concern # 3: Study Design and Technical Soundness
Author response: We thank the reviewer for this important point. To address robustness, we added text in Methods clarifying that k-fold cross-validation (k=5) was performed on the UCI dataset, confirming stable results across folds. We also report class balance: UCI dataset (~54% IHD, ~46% healthy) and Kaggle dataset (~50% each), showing no major imbalance. No resampling or weighting was needed. We further acknowledge the potential risk of overfitting in the Discussion as a limitation.
Author action: Added in Section 3.5: description of k-fold cross-validation on UCI dataset. Added in Section 3.1.4: class balance description. Expanded Discussion: limitation noting overfitting risk in small datasets.

Reviewer#2, Concern # 4: Methods and Replicability
Author response: We appreciate this suggestion. To enhance replicability, we included a pseudo-code style Algorithm (Algorithm 1) in the Methods section, summarizing the preprocessing, model training, and evaluation pipeline. We also clarified that hyperparameters were tuned via grid search, selecting the configuration with the highest validation F1-score.
Author action: Added Algorithm 1 (pipeline) in Section 3.5. Clarified hyperparameter tuning strategy (grid search).

Reviewer#2, Concern # 5: Statistical Analysis and Interpretation
Author response: We agree. Results now explicitly state they are averaged across multiple runs. We also computed 95% confidence intervals for all primary metrics using bootstrap resampling (1000 iterations).
Author action: Updated Results to note averaged results across runs.

Reviewer#2, Concern # 6: Availability of Source Data
Author response: We thank the reviewer for this comment. While raw datasets are already public, we recognize that preprocessing adds value for replication. We now provide a detailed preprocessing description in Methods (Section 3.1.4) and make scripts available upon request.
Author action: Expanded Section 3.1.4 with detailed preprocessing description.

Reviewer#2, Concern # 7: Conclusions and Support from Results
Author response: We agree with this suggestion. The Discussion has been expanded with a new Limitations subsection addressing dataset size, lack of external clinical validation, potential imputation bias, and computational cost. Statements on clinical application have been revised to emphasize that this is a proof-of-concept with potential future clinical use.
Author action: Expanded Discussion with Limitations subsection. Rephrased Conclusion to emphasize proof-of-concept, not immediate deployment.

Reviewer#2, Concern # 8: Key Points to Address
Author response: All these points have been addressed in the revision:

Cross-validation results for UCI dataset included.

95% confidence intervals.

Algorithm 1 (pseudo-code pipeline) added in Methods.

Expanded Discussion with a Limitations section.

Author action: Revisions made in Sections 3.1.4, 3.5, Results, and Discussion.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Competing Interests: The author(s) declare that they have no competing interests. Close
Report a concern

Views

Reviewer Report 25 Aug 2025

Amalie Dahl Haue, University of Copenhagen, Copenhagen, Denmark

Approved with Reservations

https://doi.org/10.5256/f1000research.182263.r401540

The research article by Ranganathan et al. presents a deep learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction derived from analysis of the Kaggle Cardiovascular Disease dataset and the UCI Heart Disease dataset. Their model (HRAEN) demonstrates superior perfomance with accuracy rating between 97.7% and 98.4%.

Introduction
The very first paragraph is could benefit from being rewritten to ensure a better flow and updated to align with current practise. For example, neither stress test, nor Holter monitoring are used routinely to detect ischemic heart disease (IHD). Rather cardiac CT, RbPET and invasive examinations such as coronary arteriography are being used to assess degree of IHD

Related works
This section would benefit greatly from a more condensed presentation of the literature.

Materials and methods
It is not clear how heart disease (presence or absence) was defined in the two cohorts, i.e. which diagnostic tests were used.
Figure 1 and 2 are not detailed enough, i.e. were all entries (observations) in the two datasets included in the study, what was the degree of missingness, and (again) how was IHD assessed?
It is not clear how the Echo State Networkds (ESNs) were applied to the data at hand since not time-series data is introduced.

Results and analysis
The different classes are not annotated consistently. That is, is "Class 1" "heart disease" (as listed in Materials and methods) or "ischemic heart disease" (as listed in Results and analysis)?
New metrics, such as Kappa score/coefficient and Jaccard coefficient are introduced in this section. They ought to be introduced in Materials and methods.
Figure 4: Does the figure display the performance of the models on a particular dataset or a combined version?
For the performance metrics, it would be beneficial to include confidence intervals for assessment of statistical significance.
Figure 6: The authors state that it converts that the proposed HRAESN model outperforms traditional classifiers in multiple performance aspects. However, only the HRAESN evaluated on the UCI and Kaggle dataset are reported in this figure.
Were the test and training sets similar? It would be nice with a table that provides an overview of the baseline characteristics in the different populations.
Table 6 and 7: The authors ought to argue that the HRAESN is comparable to the existing methods. For example, it is not clear why HRAESN on the UCI Heart Disease Dataset and the Kaggle Cardiovascular Disease dataset are being compared to different existing methods. Further, were the existing methods trained to perform a similar classification task as HRAESN. Again the definition of heart disease/IHD as how it was diagnosed is crucial here, but unfortunately lacking from this version of the manuscript.

Discussion
This section appears to be incomplete. For example, the lack of external validation is not addressed. Further, the use of imputation and the potential bias of the results is not discussed. Finally, there is not evaluation of the impact of the Attention Residual Learning (ARL), i.e., which features were most important in the classification when ARL was performed? And, could this strategy be used to identify a limited set of features that could obtain similar performance metrics?

Is the work clearly and accurately presented and does it cite the current literature?

Yes
Is the study design appropriate and is the work technically sound?

Partly
Are sufficient details of methods and analysis provided to allow replication by others?

Partly
If applicable, is the statistical analysis and its interpretation appropriate?

I cannot comment. A qualified statistician is required.
Are all the source data underlying the results available to ensure full reproducibility?

Yes
Are the conclusions drawn adequately supported by the results?

Yes

Competing Interests: No competing interests were disclosed.

Reviewer Expertise: Cardiology resident

CITE

Report a concern

Author Response 16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

16 Sep 2025

Author Response
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is ... Continue reading
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is could benefit from being rewritten to ensure a better flow and updated to align with current practise. For example, neither stress test, nor Holter monitoring are used routinely to detect ischemic heart disease (IHD). Rather cardiac CT, RbPET and invasive examinations such as coronary arteriography are being used to assess degree of IHD
Author response: We thank the reviewer for this valuable suggestion. We have revised the introductory paragraph to better reflect current clinical practices, replacing outdated references (stress test, Holter monitoring) with contemporary modalities such as cardiac CT, RbPET, and coronary angiography.
Author action: The Introduction now begins with a discussion of ischemic heart disease pathophysiology and updated diagnostic modalities.

Reviewer#1, Concern # 2: Related works
This section would benefit greatly from a more condensed presentation of the literature.
Author response: We appreciate this suggestion. We revised the Related Works section to streamline the narrative, grouping studies under thematic categories (traditional ML, deep learning, hybrid models, ESN-based, and attention-based methods).
Author action: Section 2 was restructured for conciseness while retaining comprehensiveness.

Reviewer#1, Concern # 3: Materials and methods
It is not clear how heart disease (presence or absence) was defined in the two cohorts, i.e. which diagnostic tests were used.
Author response: We acknowledge this concern. We now clearly define the target variables in both datasets:

UCI: angiography-based “num” variable, binarized (0 = absence, 1–4 = presence of disease).

Kaggle: “cardio” variable defined by combined clinical assessments (blood pressure, cholesterol, ECG).

Author action: Added Section 3.1.4 Definition of Heart Disease in the Datasets.

Reviewer#1, Concern # 4: Figure 1 and 2 are not detailed enough, i.e. were all entries (observations) in the two datasets included in the study, what was the degree of missingness, and (again) how was IHD assessed?
Author response: We appreciate this important comment. We have clarified in Section 3.1.4 how IHD was defined in each dataset (Kaggle: cardio; UCI: num attribute binarized). Missing data handling using the IHD Multiple Imputation Technique is now described. We also added clarification on how Echo State Networks were applied to structured tabular data (not time-series). Figures 1 and 2 were redesigned to show dataset composition, preprocessing, and architecture in greater detail.
Author action: Section 3.1.4 updated with disease definition, missingness handling, and ESN applicability.
Redesigned Figure 1 (workflow with dataset size, missing values, preprocessing, labels, metrics).
Redesigned Figure 2 (detailed HRAESN architecture with ESN + ARL modules).

Reviewer#1, Concern # 5: It is not clear how the Echo State Networkds (ESNs) were applied to the data at hand since not time-series data is introduced.
Author response: We agree this required clarification. While raw ECG series were not used, we adapted ESNs by treating patient feature vectors as structured sequences, mapping them into reservoir states to capture nonlinear feature dependencies.
Author action: Added explanation in Section 3.5 Methodology – Application of ESNs to Tabular Data.

Reviewer#1, Concern # 6: Results and analysis
The different classes are not annotated consistently. That is, is "Class 1" "heart disease" (as listed in Materials and methods) or "ischemic heart disease" (as listed in Results and analysis)?

Author response: We standardized terminology throughout: Class 0 = no IHD, Class 1 = IHD present.
Author action: Updated class definitions consistently across Materials & Methods, Results, and figures.

Reviewer#1, Concern # 7: New metrics, such as Kappa score/coefficient and Jaccard coefficient are introduced in this section. They ought to be introduced in Materials and methods.
Author response: We thank the reviewer. These metrics are now introduced in Evaluation Metrics subsection of Materials and Methods.
Author action: Section 3.5 includes definitions of Kappa coefficient and Jaccard index.

Reviewer#1, Concern # 8: Figure 4: Does the figure display the performance of the models on a particular dataset or a combined version?
Author response: We have clarified the figure captions to indicate that Figure 4 reports performance metrics separately for both UCI and Kaggle datasets.
Author action: Updated Figure 4 caption as suggested

Reviewer#1, Concern # 9: For the performance metrics, it would be beneficial to include confidence intervals for assessment of statistical significance.
Author response: We have now reported 95% confidence intervals using bootstrap resampling (1000 iterations) for all major performance metrics.
Author action: Confidence intervals are included in tables as suggested.

Reviewer#1, Concern # 10: Figure 6: The authors state that it converts that the proposed HRAESN model outperforms traditional classifiers in multiple performance aspects. However, only the HRAESN evaluated on the UCI and Kaggle dataset are reported in this figure.
Author response: We agree this was ambiguous. Figure 6 is intended to illustrate HRAESN error rates across datasets, while comparative results with baselines are in Tables 8–9.
Author action: Figure 6 caption updated to clarify scope

Reviewer#1, Concern # 11: Were the test and training sets similar? It would be nice with a table that provides an overview of the baseline characteristics in the different populations.
Author response: We now provide a table of baseline characteristics (age, sex, cholesterol, blood pressure) for training and test subsets.
Author action: Added Table 3: Baseline Characteristics.

Reviewer#1, Concern # 12: Table 6 and 7: The authors ought to argue that the HRAESN is comparable to the existing methods. For example, it is not clear why HRAESN on the UCI Heart Disease Dataset and the Kaggle Cardiovascular Disease dataset are being compared to different existing methods. Further, were the existing methods trained to perform a similar classification task as HRAESN. Again the definition of heart disease/IHD as how it was diagnosed is crucial here, but unfortunately lacking from this version of the manuscript.

Author response: We appreciate this comment. Tables 6 and 7 were updated/clarified with consistent captions and explanations of dataset comparisons. We expanded the Discussion to address:

Lack of external validation and need for future hospital-based datasets.

Potential imputation bias and future use of sensitivity analyses.

Importance of ARL interpretability and plans to evaluate feature contributions.

Author action: Tables 6–7 revised with rationale for comparing UCI vs Kaggle against different baselines. Expanded Section 6 Limitations and Future Directions.

Reviewer#1, Concern # 13: Discussion
This section appears to be incomplete. For example, the lack of external validation is not addressed. Further, the use of imputation and the potential bias of the results is not discussed. Finally, there is not evaluation of the impact of the Attention Residual Learning (ARL), i.e., which features were most important in the classification when ARL was performed? And, could this strategy be used to identify a limited set of features that could obtain similar performance metrics?
Author response: We agree with this comment and have expanded the Discussion to address these limitations. We now discuss the absence of external validation, the potential bias introduced by imputation, and the interpretability of ARL. We also comment on future work to explore feature importance and whether a smaller subset of features could achieve comparable accuracy.
Author action: Discussion section expanded with subsections on limitations, imputation bias, and ARL interpretability.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is could benefit from being rewritten to ensure a better flow and updated to align with current practise. For example, neither stress test, nor Holter monitoring are used routinely to detect ischemic heart disease (IHD). Rather cardiac CT, RbPET and invasive examinations such as coronary arteriography are being used to assess degree of IHD
Author response: We thank the reviewer for this valuable suggestion. We have revised the introductory paragraph to better reflect current clinical practices, replacing outdated references (stress test, Holter monitoring) with contemporary modalities such as cardiac CT, RbPET, and coronary angiography.
Author action: The Introduction now begins with a discussion of ischemic heart disease pathophysiology and updated diagnostic modalities.

Reviewer#1, Concern # 2: Related works
This section would benefit greatly from a more condensed presentation of the literature.
Author response: We appreciate this suggestion. We revised the Related Works section to streamline the narrative, grouping studies under thematic categories (traditional ML, deep learning, hybrid models, ESN-based, and attention-based methods).
Author action: Section 2 was restructured for conciseness while retaining comprehensiveness.

Reviewer#1, Concern # 3: Materials and methods
It is not clear how heart disease (presence or absence) was defined in the two cohorts, i.e. which diagnostic tests were used.
Author response: We acknowledge this concern. We now clearly define the target variables in both datasets:

UCI: angiography-based “num” variable, binarized (0 = absence, 1–4 = presence of disease).

Kaggle: “cardio” variable defined by combined clinical assessments (blood pressure, cholesterol, ECG).

Author action: Added Section 3.1.4 Definition of Heart Disease in the Datasets.

Reviewer#1, Concern # 4: Figure 1 and 2 are not detailed enough, i.e. were all entries (observations) in the two datasets included in the study, what was the degree of missingness, and (again) how was IHD assessed?
Author response: We appreciate this important comment. We have clarified in Section 3.1.4 how IHD was defined in each dataset (Kaggle: cardio; UCI: num attribute binarized). Missing data handling using the IHD Multiple Imputation Technique is now described. We also added clarification on how Echo State Networks were applied to structured tabular data (not time-series). Figures 1 and 2 were redesigned to show dataset composition, preprocessing, and architecture in greater detail.
Author action: Section 3.1.4 updated with disease definition, missingness handling, and ESN applicability.
Redesigned Figure 1 (workflow with dataset size, missing values, preprocessing, labels, metrics).
Redesigned Figure 2 (detailed HRAESN architecture with ESN + ARL modules).

Reviewer#1, Concern # 5: It is not clear how the Echo State Networkds (ESNs) were applied to the data at hand since not time-series data is introduced.
Author response: We agree this required clarification. While raw ECG series were not used, we adapted ESNs by treating patient feature vectors as structured sequences, mapping them into reservoir states to capture nonlinear feature dependencies.
Author action: Added explanation in Section 3.5 Methodology – Application of ESNs to Tabular Data.

Reviewer#1, Concern # 6: Results and analysis
The different classes are not annotated consistently. That is, is "Class 1" "heart disease" (as listed in Materials and methods) or "ischemic heart disease" (as listed in Results and analysis)?

Author response: We standardized terminology throughout: Class 0 = no IHD, Class 1 = IHD present.
Author action: Updated class definitions consistently across Materials & Methods, Results, and figures.

Reviewer#1, Concern # 7: New metrics, such as Kappa score/coefficient and Jaccard coefficient are introduced in this section. They ought to be introduced in Materials and methods.
Author response: We thank the reviewer. These metrics are now introduced in Evaluation Metrics subsection of Materials and Methods.
Author action: Section 3.5 includes definitions of Kappa coefficient and Jaccard index.

Reviewer#1, Concern # 8: Figure 4: Does the figure display the performance of the models on a particular dataset or a combined version?
Author response: We have clarified the figure captions to indicate that Figure 4 reports performance metrics separately for both UCI and Kaggle datasets.
Author action: Updated Figure 4 caption as suggested

Reviewer#1, Concern # 9: For the performance metrics, it would be beneficial to include confidence intervals for assessment of statistical significance.
Author response: We have now reported 95% confidence intervals using bootstrap resampling (1000 iterations) for all major performance metrics.
Author action: Confidence intervals are included in tables as suggested.

Reviewer#1, Concern # 10: Figure 6: The authors state that it converts that the proposed HRAESN model outperforms traditional classifiers in multiple performance aspects. However, only the HRAESN evaluated on the UCI and Kaggle dataset are reported in this figure.
Author response: We agree this was ambiguous. Figure 6 is intended to illustrate HRAESN error rates across datasets, while comparative results with baselines are in Tables 8–9.
Author action: Figure 6 caption updated to clarify scope

Reviewer#1, Concern # 11: Were the test and training sets similar? It would be nice with a table that provides an overview of the baseline characteristics in the different populations.
Author response: We now provide a table of baseline characteristics (age, sex, cholesterol, blood pressure) for training and test subsets.
Author action: Added Table 3: Baseline Characteristics.

Reviewer#1, Concern # 12: Table 6 and 7: The authors ought to argue that the HRAESN is comparable to the existing methods. For example, it is not clear why HRAESN on the UCI Heart Disease Dataset and the Kaggle Cardiovascular Disease dataset are being compared to different existing methods. Further, were the existing methods trained to perform a similar classification task as HRAESN. Again the definition of heart disease/IHD as how it was diagnosed is crucial here, but unfortunately lacking from this version of the manuscript.

Author response: We appreciate this comment. Tables 6 and 7 were updated/clarified with consistent captions and explanations of dataset comparisons. We expanded the Discussion to address:

Lack of external validation and need for future hospital-based datasets.

Potential imputation bias and future use of sensitivity analyses.

Importance of ARL interpretability and plans to evaluate feature contributions.

Author action: Tables 6–7 revised with rationale for comparing UCI vs Kaggle against different baselines. Expanded Section 6 Limitations and Future Directions.

Reviewer#1, Concern # 13: Discussion
This section appears to be incomplete. For example, the lack of external validation is not addressed. Further, the use of imputation and the potential bias of the results is not discussed. Finally, there is not evaluation of the impact of the Attention Residual Learning (ARL), i.e., which features were most important in the classification when ARL was performed? And, could this strategy be used to identify a limited set of features that could obtain similar performance metrics?
Author response: We agree with this comment and have expanded the Discussion to address these limitations. We now discuss the absence of external validation, the potential bias introduced by imputation, and the interpretability of ARL. We also comment on future work to explore feature importance and whether a smaller subset of features could achieve comparable accuracy.
Author action: Discussion section expanded with subsections on limitations, imputation bias, and ARL interpretability.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Competing Interests: The author(s) declare that they have no competing interests. Close
Report a concern
Respond or Comment

COMMENTS ON THIS REPORT

Author Response 16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

16 Sep 2025

Author Response
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is ... Continue reading
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is could benefit from being rewritten to ensure a better flow and updated to align with current practise. For example, neither stress test, nor Holter monitoring are used routinely to detect ischemic heart disease (IHD). Rather cardiac CT, RbPET and invasive examinations such as coronary arteriography are being used to assess degree of IHD
Author response: We thank the reviewer for this valuable suggestion. We have revised the introductory paragraph to better reflect current clinical practices, replacing outdated references (stress test, Holter monitoring) with contemporary modalities such as cardiac CT, RbPET, and coronary angiography.
Author action: The Introduction now begins with a discussion of ischemic heart disease pathophysiology and updated diagnostic modalities.

Reviewer#1, Concern # 2: Related works
This section would benefit greatly from a more condensed presentation of the literature.
Author response: We appreciate this suggestion. We revised the Related Works section to streamline the narrative, grouping studies under thematic categories (traditional ML, deep learning, hybrid models, ESN-based, and attention-based methods).
Author action: Section 2 was restructured for conciseness while retaining comprehensiveness.

Reviewer#1, Concern # 3: Materials and methods
It is not clear how heart disease (presence or absence) was defined in the two cohorts, i.e. which diagnostic tests were used.
Author response: We acknowledge this concern. We now clearly define the target variables in both datasets:

UCI: angiography-based “num” variable, binarized (0 = absence, 1–4 = presence of disease).

Kaggle: “cardio” variable defined by combined clinical assessments (blood pressure, cholesterol, ECG).

Author action: Added Section 3.1.4 Definition of Heart Disease in the Datasets.

Reviewer#1, Concern # 4: Figure 1 and 2 are not detailed enough, i.e. were all entries (observations) in the two datasets included in the study, what was the degree of missingness, and (again) how was IHD assessed?
Author response: We appreciate this important comment. We have clarified in Section 3.1.4 how IHD was defined in each dataset (Kaggle: cardio; UCI: num attribute binarized). Missing data handling using the IHD Multiple Imputation Technique is now described. We also added clarification on how Echo State Networks were applied to structured tabular data (not time-series). Figures 1 and 2 were redesigned to show dataset composition, preprocessing, and architecture in greater detail.
Author action: Section 3.1.4 updated with disease definition, missingness handling, and ESN applicability.
Redesigned Figure 1 (workflow with dataset size, missing values, preprocessing, labels, metrics).
Redesigned Figure 2 (detailed HRAESN architecture with ESN + ARL modules).

Reviewer#1, Concern # 5: It is not clear how the Echo State Networkds (ESNs) were applied to the data at hand since not time-series data is introduced.
Author response: We agree this required clarification. While raw ECG series were not used, we adapted ESNs by treating patient feature vectors as structured sequences, mapping them into reservoir states to capture nonlinear feature dependencies.
Author action: Added explanation in Section 3.5 Methodology – Application of ESNs to Tabular Data.

Reviewer#1, Concern # 6: Results and analysis
The different classes are not annotated consistently. That is, is "Class 1" "heart disease" (as listed in Materials and methods) or "ischemic heart disease" (as listed in Results and analysis)?

Author response: We standardized terminology throughout: Class 0 = no IHD, Class 1 = IHD present.
Author action: Updated class definitions consistently across Materials & Methods, Results, and figures.

Reviewer#1, Concern # 7: New metrics, such as Kappa score/coefficient and Jaccard coefficient are introduced in this section. They ought to be introduced in Materials and methods.
Author response: We thank the reviewer. These metrics are now introduced in Evaluation Metrics subsection of Materials and Methods.
Author action: Section 3.5 includes definitions of Kappa coefficient and Jaccard index.

Reviewer#1, Concern # 8: Figure 4: Does the figure display the performance of the models on a particular dataset or a combined version?
Author response: We have clarified the figure captions to indicate that Figure 4 reports performance metrics separately for both UCI and Kaggle datasets.
Author action: Updated Figure 4 caption as suggested

Reviewer#1, Concern # 9: For the performance metrics, it would be beneficial to include confidence intervals for assessment of statistical significance.
Author response: We have now reported 95% confidence intervals using bootstrap resampling (1000 iterations) for all major performance metrics.
Author action: Confidence intervals are included in tables as suggested.

Reviewer#1, Concern # 10: Figure 6: The authors state that it converts that the proposed HRAESN model outperforms traditional classifiers in multiple performance aspects. However, only the HRAESN evaluated on the UCI and Kaggle dataset are reported in this figure.
Author response: We agree this was ambiguous. Figure 6 is intended to illustrate HRAESN error rates across datasets, while comparative results with baselines are in Tables 8–9.
Author action: Figure 6 caption updated to clarify scope

Reviewer#1, Concern # 11: Were the test and training sets similar? It would be nice with a table that provides an overview of the baseline characteristics in the different populations.
Author response: We now provide a table of baseline characteristics (age, sex, cholesterol, blood pressure) for training and test subsets.
Author action: Added Table 3: Baseline Characteristics.

Reviewer#1, Concern # 12: Table 6 and 7: The authors ought to argue that the HRAESN is comparable to the existing methods. For example, it is not clear why HRAESN on the UCI Heart Disease Dataset and the Kaggle Cardiovascular Disease dataset are being compared to different existing methods. Further, were the existing methods trained to perform a similar classification task as HRAESN. Again the definition of heart disease/IHD as how it was diagnosed is crucial here, but unfortunately lacking from this version of the manuscript.

Author response: We appreciate this comment. Tables 6 and 7 were updated/clarified with consistent captions and explanations of dataset comparisons. We expanded the Discussion to address:

Lack of external validation and need for future hospital-based datasets.

Potential imputation bias and future use of sensitivity analyses.

Importance of ARL interpretability and plans to evaluate feature contributions.

Author action: Tables 6–7 revised with rationale for comparing UCI vs Kaggle against different baselines. Expanded Section 6 Limitations and Future Directions.

Reviewer#1, Concern # 13: Discussion
This section appears to be incomplete. For example, the lack of external validation is not addressed. Further, the use of imputation and the potential bias of the results is not discussed. Finally, there is not evaluation of the impact of the Attention Residual Learning (ARL), i.e., which features were most important in the classification when ARL was performed? And, could this strategy be used to identify a limited set of features that could obtain similar performance metrics?
Author response: We agree with this comment and have expanded the Discussion to address these limitations. We now discuss the absence of external validation, the potential bias introduced by imputation, and the interpretability of ARL. We also comment on future work to explore feature importance and whether a smaller subset of features could achieve comparable accuracy.
Author action: Discussion section expanded with subsections on limitations, imputation bias, and ARL interpretability.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is could benefit from being rewritten to ensure a better flow and updated to align with current practise. For example, neither stress test, nor Holter monitoring are used routinely to detect ischemic heart disease (IHD). Rather cardiac CT, RbPET and invasive examinations such as coronary arteriography are being used to assess degree of IHD
Author response: We thank the reviewer for this valuable suggestion. We have revised the introductory paragraph to better reflect current clinical practices, replacing outdated references (stress test, Holter monitoring) with contemporary modalities such as cardiac CT, RbPET, and coronary angiography.
Author action: The Introduction now begins with a discussion of ischemic heart disease pathophysiology and updated diagnostic modalities.

Reviewer#1, Concern # 2: Related works
This section would benefit greatly from a more condensed presentation of the literature.
Author response: We appreciate this suggestion. We revised the Related Works section to streamline the narrative, grouping studies under thematic categories (traditional ML, deep learning, hybrid models, ESN-based, and attention-based methods).
Author action: Section 2 was restructured for conciseness while retaining comprehensiveness.

Reviewer#1, Concern # 3: Materials and methods
It is not clear how heart disease (presence or absence) was defined in the two cohorts, i.e. which diagnostic tests were used.
Author response: We acknowledge this concern. We now clearly define the target variables in both datasets:

UCI: angiography-based “num” variable, binarized (0 = absence, 1–4 = presence of disease).

Kaggle: “cardio” variable defined by combined clinical assessments (blood pressure, cholesterol, ECG).

Author action: Added Section 3.1.4 Definition of Heart Disease in the Datasets.

Reviewer#1, Concern # 4: Figure 1 and 2 are not detailed enough, i.e. were all entries (observations) in the two datasets included in the study, what was the degree of missingness, and (again) how was IHD assessed?
Author response: We appreciate this important comment. We have clarified in Section 3.1.4 how IHD was defined in each dataset (Kaggle: cardio; UCI: num attribute binarized). Missing data handling using the IHD Multiple Imputation Technique is now described. We also added clarification on how Echo State Networks were applied to structured tabular data (not time-series). Figures 1 and 2 were redesigned to show dataset composition, preprocessing, and architecture in greater detail.
Author action: Section 3.1.4 updated with disease definition, missingness handling, and ESN applicability.
Redesigned Figure 1 (workflow with dataset size, missing values, preprocessing, labels, metrics).
Redesigned Figure 2 (detailed HRAESN architecture with ESN + ARL modules).

Reviewer#1, Concern # 5: It is not clear how the Echo State Networkds (ESNs) were applied to the data at hand since not time-series data is introduced.
Author response: We agree this required clarification. While raw ECG series were not used, we adapted ESNs by treating patient feature vectors as structured sequences, mapping them into reservoir states to capture nonlinear feature dependencies.
Author action: Added explanation in Section 3.5 Methodology – Application of ESNs to Tabular Data.

Reviewer#1, Concern # 6: Results and analysis
The different classes are not annotated consistently. That is, is "Class 1" "heart disease" (as listed in Materials and methods) or "ischemic heart disease" (as listed in Results and analysis)?

Author response: We standardized terminology throughout: Class 0 = no IHD, Class 1 = IHD present.
Author action: Updated class definitions consistently across Materials & Methods, Results, and figures.

Reviewer#1, Concern # 7: New metrics, such as Kappa score/coefficient and Jaccard coefficient are introduced in this section. They ought to be introduced in Materials and methods.
Author response: We thank the reviewer. These metrics are now introduced in Evaluation Metrics subsection of Materials and Methods.
Author action: Section 3.5 includes definitions of Kappa coefficient and Jaccard index.

Reviewer#1, Concern # 8: Figure 4: Does the figure display the performance of the models on a particular dataset or a combined version?
Author response: We have clarified the figure captions to indicate that Figure 4 reports performance metrics separately for both UCI and Kaggle datasets.
Author action: Updated Figure 4 caption as suggested

Reviewer#1, Concern # 9: For the performance metrics, it would be beneficial to include confidence intervals for assessment of statistical significance.
Author response: We have now reported 95% confidence intervals using bootstrap resampling (1000 iterations) for all major performance metrics.
Author action: Confidence intervals are included in tables as suggested.

Reviewer#1, Concern # 10: Figure 6: The authors state that it converts that the proposed HRAESN model outperforms traditional classifiers in multiple performance aspects. However, only the HRAESN evaluated on the UCI and Kaggle dataset are reported in this figure.
Author response: We agree this was ambiguous. Figure 6 is intended to illustrate HRAESN error rates across datasets, while comparative results with baselines are in Tables 8–9.
Author action: Figure 6 caption updated to clarify scope

Reviewer#1, Concern # 11: Were the test and training sets similar? It would be nice with a table that provides an overview of the baseline characteristics in the different populations.
Author response: We now provide a table of baseline characteristics (age, sex, cholesterol, blood pressure) for training and test subsets.
Author action: Added Table 3: Baseline Characteristics.

Reviewer#1, Concern # 12: Table 6 and 7: The authors ought to argue that the HRAESN is comparable to the existing methods. For example, it is not clear why HRAESN on the UCI Heart Disease Dataset and the Kaggle Cardiovascular Disease dataset are being compared to different existing methods. Further, were the existing methods trained to perform a similar classification task as HRAESN. Again the definition of heart disease/IHD as how it was diagnosed is crucial here, but unfortunately lacking from this version of the manuscript.

Author response: We appreciate this comment. Tables 6 and 7 were updated/clarified with consistent captions and explanations of dataset comparisons. We expanded the Discussion to address:

Lack of external validation and need for future hospital-based datasets.

Potential imputation bias and future use of sensitivity analyses.

Importance of ARL interpretability and plans to evaluate feature contributions.

Author action: Tables 6–7 revised with rationale for comparing UCI vs Kaggle against different baselines. Expanded Section 6 Limitations and Future Directions.

Reviewer#1, Concern # 13: Discussion
This section appears to be incomplete. For example, the lack of external validation is not addressed. Further, the use of imputation and the potential bias of the results is not discussed. Finally, there is not evaluation of the impact of the Attention Residual Learning (ARL), i.e., which features were most important in the classification when ARL was performed? And, could this strategy be used to identify a limited set of features that could obtain similar performance metrics?
Author response: We agree with this comment and have expanded the Discussion to address these limitations. We now discuss the absence of external validation, the potential bias introduced by imputation, and the interpretability of ARL. We also comment on future work to explore feature importance and whether a smaller subset of features could achieve comparable accuracy.
Author action: Discussion section expanded with subsections on limitations, imputation bias, and ARL interpretability.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.
Competing Interests: The author(s) declare that they have no competing interests. Close
Report a concern

Comments on this article Comments (0)

Version 2

VERSION 2 PUBLISHED 03 Jul 2025

Open Peer Review

Reviewer Status

Reviewer Reports

	Invited Reviewers
	1	2	3
Version 2 (revision) 16 Sep 25		read
Version 1 03 Jul 25	read	read	read

Amalie Dahl Haue, University of Copenhagen, Copenhagen, Denmark
Dhadkan Shrestha, Texas State University College of Science and Engineering, San Marcos, USA
MUHAMMAD HAMMAD MEMON, Southwest University of Science and Technology, Sichuan, China

Comments on this article

All Comments(0)

Add a comment

Browse by related subjects

Back to all reports

Reviewer Report

6 Views

25 Sep 2025 | for Version 2

Dhadkan Shrestha, Texas State University College of Science and Engineering, San Marcos, Texas, USA

6 Views Cite this report Responses(1)

Approved

Everything looks good now. Thank you for revising.

Competing Interests

No competing interests were disclosed.

Reviewer Expertise

Machine Learning, Artificial Intelligence, Big Data

I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.

Respond to this report

Responses (1)

Back to all reports

Reviewer Report

14 Views

04 Sep 2025 | for Version 1

MUHAMMAD HAMMAD MEMON, Southwest University of Science and Technology, Sichuan, China

14 Views Cite this report Responses(1)

Not Approved

Presentation and Literature Coverage
- The manuscript is generally clear, but the literature review is overly descriptive and includes some weak references (e.g., tutorial websites).
- Prior work on combining attention and ESN (e.g., Deep Belief Echo-State Networks, Graph Residual Attention) is not sufficiently discussed. The novelty contribution must be better distinguished.
Study Design and Technical Soundness
- The reported performance (>97% accuracy) is unrealistically high for these datasets and suggests possible overfitting or data leakage.
- Only a single 80:20 train-test split is reported. This is not sufficient for robust evaluation in medical ML. At minimum, k-fold cross-validation with stratified sampling is required.
Methods and Replication
- Details of the Ischemic Heart Disease Multiple Imputation Technique are insufficient. The method is referenced but not described in reproducible detail.
- No code, model weights, or supplementary scripts are provided, making replication difficult.
Statistical Analysis
- No statistical significance testing (e.g., McNemar’s test, paired t-test, Wilcoxon signed-rank test) is provided. Reported differences may not be statistically meaningful.
- Metrics such as ROC curves, AUC, calibration plots, and precision-recall curves should be included for clinical interpretability.
Reproducibility and Source Data
- Although the datasets are public, the exact preprocessing steps and imputation pipeline are not fully transparent, which limits reproducibility.
- PCA plots and confusion matrices are shown but lack supporting raw numbers or code availability.
Support for Conclusions
- While results are promising, conclusions about clinical utility are overstated. Without independent external validation on real hospital datasets, it is premature to suggest readiness for clinical deployment.
- Limitations such as dataset imbalance, computational cost, and lack of external validation are only briefly acknowledged and need stronger discussion.

Minor Comments

Some sections could be streamlined (particularly Related Works).
Figures would benefit from statistical annotations (e.g., significance levels).
The ethics statement should clarify whether the Kaggle dataset contributor had appropriate institutional approval.
Writing is generally clear but could be more concise in parts.

Recommendations to Improve the Manuscript

Re-run experiments with 10-fold cross-validation and report mean ± standard deviation.
Add statistical tests to confirm whether improvements over baselines are significant.
Provide algorithmic details of the imputation method and release source code/models.
Include AUC/ROC, calibration, and PR curves for stronger evaluation.
Strengthen the novelty discussion by differentiating HRAESN from earlier ESN+attention studies.
Expand the limitations section, especially regarding generalizability and clinical applicability.

Is the work clearly and accurately presented and does it cite the current literature?

Partly
Is the study design appropriate and is the work technically sound?

Partly
Are sufficient details of methods and analysis provided to allow replication by others?

Partly
If applicable, is the statistical analysis and its interpretation appropriate?

Partly
Are all the source data underlying the results available to ensure full reproducibility?

Partly
Are the conclusions drawn adequately supported by the results?

Partly

Competing Interests

No competing interests were disclosed.

Reviewer Expertise

Artificial Intelligence and Machine Learning, Medical Data Mining and Predictive Analytics, Deep Learning for Healthcare Applications, Network Security and Cloud Computing.

Respond to this report

Responses (1)

Author Response

16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#3, Concern # 1: Summary of the Article:
Author response: We thank Reviewer for the detailed assessment. The reviewer highlighted the clinical relevance of our work while raising concerns about methodology, reproducibility, and statistical robustness. We carefully revised the manuscript to address all points raised. Below we provide a structured response.

Reviewer#3, Concern # 2: Presentation and Literature Coverage
Author response: We acknowledge this important observation. The Related Works section has been streamlined and focused on high-quality peer-reviewed studies. We expanded discussion of prior attention+ESN combinations, including Deep Belief Echo-State Networks (DBEN) and Graph Residual Attention models, to clearly distinguish our contribution. Our novelty lies in extending ESNs beyond time-series into structured clinical tabular data, integrated with ARL and combined with a tailored imputation framework.
Author action: Revised Section 2 (Related Works) to be more concise, replaced weak/tutorial references with peer-reviewed sources, and explicitly clarified novelty.

Reviewer#3, Concern # 3: Study Design and Technical Soundness
Author response: We appreciate this concern. To strengthen robustness, we re-ran experiments with 5-fold and 10-fold stratified cross-validation in addition to the 80:20 split. Results are now reported as mean ± standard deviation. Performance remained consistently high, though slightly lower than single-split values, confirming stability without evidence of leakage.
Author action: Added cross-validation experiments and updated Tables 7–10 with mean ± SD. Reproducibility pipeline clarified in Section 3.5 Methodology.

Reviewer#3, Concern # 4: Methods and Replication
Author response: We agree. The Ischemic Heart Disease Multiple Imputation Technique (IHD-MIT) is now described in step-by-step detail (predictor selection, iterative regression, variance preservation). For transparency, we have expanded the methodological description of the IHD-MIT imputation pipeline and model implementation in detail.
Author action: Expanded Section 3.5.1 (IHD-MIT) with algorithmic details.

Reviewer#3, Concern # 5: Statistical Analysis
Author response: We fully agree. We added statistical significance testing (McNemar’s test for paired predictions, Wilcoxon signed-rank across folds) to confirm differences. Additionally, we now report ROC curves, AUC values, and calibration plots for clinical interpretability. Results demonstrate that HRAESN improvements are statistically significant (p < 0.05).
Author action: Added Figure 7 for ROC/AUC; included calibration analysis. Expanded Results Section 4.2–4.3 to include statistical testing.

Reviewer#3, Concern # 6: Reproducibility and Source Data
Author response: We clarified all preprocessing steps, including normalization, imputation, train-test stratification, and cross-validation.
Author action: Updated Figures 3–6 captions with supporting details.

Reviewer#3, Concern # 7: Support for Conclusions
Author response: We agree and have moderated claims. We now clearly state that this work is a proof-of-concept and not clinically deployable yet. We expanded Limitations to address external validation needs, potential bias from imputation, dataset imbalance, computational cost, and the need for interpretability studies.
Author action: Expanded Section 6 Limitations and Future Directions, emphasizing generalizability and next steps toward real-world validation.

Reviewer#3, Concern # 8: Minor Comments
Author response: We thank the reviewer. Related Works was condensed (as above). Figures now include statistical annotations (significance levels). We clarified that Kaggle data are anonymized and released under open license, with ethical approvals obtained by original curators. The manuscript was carefully edited for conciseness.
Author action: Revised Section 2, updated figure annotations, clarified Ethics Statement, and streamlined prose throughout.

Reviewer#3, Concern # 1: Reviewer Recommendations Implemented
Author response:

Re-ran experiments with 10-fold cross-validation.
Reported mean ± SD for all metrics.
Added statistical tests (McNemar, Wilcoxon).
Included ROC curves.
Provided algorithmic details of IHD-MIT.
Strengthened novelty discussion and limitations.

·
We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.

View more View less

Competing Interests

The author(s) declare that they have no competing interests.

Back to all reports

Reviewer Report

7 Views

04 Sep 2025 | for Version 1

Dhadkan Shrestha, Texas State University College of Science and Engineering, San Marcos, Texas, USA

7 Views Cite this report Responses(1)

Approved With Reservations

Assessment: Yes (with minor improvements suggested)
The manuscript is clearly written, structured logically, and cites a broad range of recent literature. The background is thorough and informative. A few parts (e.g., the objectives and problem statement) overlap slightly and could be streamlined for conciseness.
Constructive suggestions:
- Condense repetitive sections to make the narrative flow smoother.
- More explicitly highlight how this approach differs from other recent hybrid deep learning works to strengthen the novelty claim.

(b) Study Design and Technical Soundness

Assessment: Yes
The study design is technically sound, and the proposed model is innovative. The integration of ARL and ESN is well motivated. The results are very strong, though the extremely high accuracy on the small UCI dataset raises the possibility of overfitting. Still, the use of dropout and a robust hyperparameter setup is a positive point.
Suggestions:
- For added robustness, apply k-fold cross-validation (especially for UCI dataset).
- Briefly discuss class balance and whether any balancing strategy (e.g., weighting) was needed.

Assessment: Partly
The mathematical formulation is clear, and hyperparameters are well documented. This is very helpful. However, replication would be easier if code or pseudo-code for preprocessing and training were made available.
Suggestions:
- Consider providing code, pseudocode, or a detailed pipeline in supplementary materials.
- Clarify how hyperparameters were tuned (manual search, grid search, etc.).

(d) Statistical Analysis and Interpretation

Assessment: Yes
The authors present a comprehensive set of performance metrics (accuracy, sensitivity, specificity, F1, Kappa, FAR/FRR), which is commendable. Interpretation is generally appropriate. One minor limitation is the absence of variance/confidence intervals across multiple runs.
Suggestions:
- Indicate whether results are from a single run or averaged across runs.
- If possible, include confidence intervals or standard deviations.

(e) Availability of Source Data

Assessment: Yes
The datasets (UCI and Kaggle) are publicly available and properly cited. Ethical considerations are addressed. This ensures reproducibility of the raw data.
Suggestions:
- It would be helpful to share the preprocessed datasets or preprocessing scripts used before training.

(f) Conclusions and Support from Results

Assessment: Yes
The conclusions are well supported by the reported results. The performance improvement over baselines is clear. That said, claims about clinical applicability should be framed as potential future applications rather than immediate readiness.
Suggestions:
- Add a brief “Limitations” section noting that real-world hospital validation is pending.
- Slightly temper statements on clinical deployment to emphasize this is a proof-of-concept

3. Key Points to Address
To make the manuscript even stronger, the authors should consider:

Adding cross-validation results (especially for the UCI dataset).
Reporting variance or confidence intervals for performance metrics.
Providing code/pseudocode or preprocessing details for easier replication.
Including a short limitations section (dataset size, clinical validation, computational cost).

Is the work clearly and accurately presented and does it cite the current literature?

Partly
Is the study design appropriate and is the work technically sound?

Partly
Are sufficient details of methods and analysis provided to allow replication by others?

Partly
If applicable, is the statistical analysis and its interpretation appropriate?

Partly
Are all the source data underlying the results available to ensure full reproducibility?

Yes
Are the conclusions drawn adequately supported by the results?

Partly

Competing Interests

No competing interests were disclosed.

Reviewer Expertise

Machine Learning, Artificial Intelligence, Big Data

Respond to this report

Responses (1)

Author Response

16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#2, Concern # 1: Summary of the Article
Author response: We thank the reviewer for the accurate and concise summary of our work. We appreciate the recognition of our proposed Hybrid Residual Attention with Echo State Network (HRAESN) model, our methodological contributions (Attention Residual Learning combined with Echo State Networks), and our evaluation using the Kaggle and UCI datasets. We also thank the reviewer for noting our strategy for handling missing values and the strong performance achieved by the model.

Reviewer#2, Concern # 2: Clarity, Accuracy, and Literature Coverage
Author response: We agree with this suggestion. We revised the Introduction to remove overlap between the problem statement and objectives, improving narrative flow. Additionally, we added a new paragraph at the end of the Introduction to explicitly highlight novelty: (i) integration of ARL with ESNs, (ii) extending ESNs to structured/tabular clinical data, and (iii) introducing an IHD-specific multiple imputation method.
Author action: Revised Introduction: merged problem statement + objectives into a concise paragraph. Added final paragraph in Introduction to emphasize novelty.

Reviewer#2, Concern # 3: Study Design and Technical Soundness
Author response: We thank the reviewer for this important point. To address robustness, we added text in Methods clarifying that k-fold cross-validation (k=5) was performed on the UCI dataset, confirming stable results across folds. We also report class balance: UCI dataset (~54% IHD, ~46% healthy) and Kaggle dataset (~50% each), showing no major imbalance. No resampling or weighting was needed. We further acknowledge the potential risk of overfitting in the Discussion as a limitation.
Author action: Added in Section 3.5: description of k-fold cross-validation on UCI dataset. Added in Section 3.1.4: class balance description. Expanded Discussion: limitation noting overfitting risk in small datasets.

Reviewer#2, Concern # 4: Methods and Replicability
Author response: We appreciate this suggestion. To enhance replicability, we included a pseudo-code style Algorithm (Algorithm 1) in the Methods section, summarizing the preprocessing, model training, and evaluation pipeline. We also clarified that hyperparameters were tuned via grid search, selecting the configuration with the highest validation F1-score.
Author action: Added Algorithm 1 (pipeline) in Section 3.5. Clarified hyperparameter tuning strategy (grid search).

Reviewer#2, Concern # 5: Statistical Analysis and Interpretation
Author response: We agree. Results now explicitly state they are averaged across multiple runs. We also computed 95% confidence intervals for all primary metrics using bootstrap resampling (1000 iterations).
Author action: Updated Results to note averaged results across runs.

Reviewer#2, Concern # 6: Availability of Source Data
Author response: We thank the reviewer for this comment. While raw datasets are already public, we recognize that preprocessing adds value for replication. We now provide a detailed preprocessing description in Methods (Section 3.1.4) and make scripts available upon request.
Author action: Expanded Section 3.1.4 with detailed preprocessing description.

Reviewer#2, Concern # 7: Conclusions and Support from Results
Author response: We agree with this suggestion. The Discussion has been expanded with a new Limitations subsection addressing dataset size, lack of external clinical validation, potential imputation bias, and computational cost. Statements on clinical application have been revised to emphasize that this is a proof-of-concept with potential future clinical use.
Author action: Expanded Discussion with Limitations subsection. Rephrased Conclusion to emphasize proof-of-concept, not immediate deployment.

Reviewer#2, Concern # 8: Key Points to Address
Author response: All these points have been addressed in the revision:

Cross-validation results for UCI dataset included.
95% confidence intervals.
Algorithm 1 (pseudo-code pipeline) added in Methods.
Expanded Discussion with a Limitations section.

Author action: Revisions made in Sections 3.1.4, 3.5, Results, and Discussion.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.

View more View less

Competing Interests

The author(s) declare that they have no competing interests.

Back to all reports

Reviewer Report

29 Views

25 Aug 2025 | for Version 1

Amalie Dahl Haue, University of Copenhagen, Copenhagen, Denmark

29 Views Cite this report Responses(1)

Approved With Reservations

Is the work clearly and accurately presented and does it cite the current literature?

Yes
Is the study design appropriate and is the work technically sound?

Partly
Are sufficient details of methods and analysis provided to allow replication by others?

Partly
If applicable, is the statistical analysis and its interpretation appropriate?

I cannot comment. A qualified statistician is required.
Are all the source data underlying the results available to ensure full reproducibility?

Yes
Are the conclusions drawn adequately supported by the results?

Yes

Competing Interests

No competing interests were disclosed.

Reviewer Expertise

Cardiology resident

Respond to this report

Responses (1)

Author Response

16 Sep 2025

VIJAYA ARJUNAN RANGANATHAN, Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India

Deep Learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction
Responses to peer review reports
Reviewer#1, Concern # 1: Introduction
The very first paragraph is could benefit from being rewritten to ensure a better flow and updated to align with current practise. For example, neither stress test, nor Holter monitoring are used routinely to detect ischemic heart disease (IHD). Rather cardiac CT, RbPET and invasive examinations such as coronary arteriography are being used to assess degree of IHD
Author response: We thank the reviewer for this valuable suggestion. We have revised the introductory paragraph to better reflect current clinical practices, replacing outdated references (stress test, Holter monitoring) with contemporary modalities such as cardiac CT, RbPET, and coronary angiography.
Author action: The Introduction now begins with a discussion of ischemic heart disease pathophysiology and updated diagnostic modalities.

Reviewer#1, Concern # 2: Related works
This section would benefit greatly from a more condensed presentation of the literature.
Author response: We appreciate this suggestion. We revised the Related Works section to streamline the narrative, grouping studies under thematic categories (traditional ML, deep learning, hybrid models, ESN-based, and attention-based methods).
Author action: Section 2 was restructured for conciseness while retaining comprehensiveness.

Reviewer#1, Concern # 3: Materials and methods
It is not clear how heart disease (presence or absence) was defined in the two cohorts, i.e. which diagnostic tests were used.
Author response: We acknowledge this concern. We now clearly define the target variables in both datasets:

UCI: angiography-based “num” variable, binarized (0 = absence, 1–4 = presence of disease).
Kaggle: “cardio” variable defined by combined clinical assessments (blood pressure, cholesterol, ECG).

Author action: Added Section 3.1.4 Definition of Heart Disease in the Datasets.

Reviewer#1, Concern # 4: Figure 1 and 2 are not detailed enough, i.e. were all entries (observations) in the two datasets included in the study, what was the degree of missingness, and (again) how was IHD assessed?
Author response: We appreciate this important comment. We have clarified in Section 3.1.4 how IHD was defined in each dataset (Kaggle: cardio; UCI: num attribute binarized). Missing data handling using the IHD Multiple Imputation Technique is now described. We also added clarification on how Echo State Networks were applied to structured tabular data (not time-series). Figures 1 and 2 were redesigned to show dataset composition, preprocessing, and architecture in greater detail.
Author action: Section 3.1.4 updated with disease definition, missingness handling, and ESN applicability.
Redesigned Figure 1 (workflow with dataset size, missing values, preprocessing, labels, metrics).
Redesigned Figure 2 (detailed HRAESN architecture with ESN + ARL modules).

Reviewer#1, Concern # 5: It is not clear how the Echo State Networkds (ESNs) were applied to the data at hand since not time-series data is introduced.
Author response: We agree this required clarification. While raw ECG series were not used, we adapted ESNs by treating patient feature vectors as structured sequences, mapping them into reservoir states to capture nonlinear feature dependencies.
Author action: Added explanation in Section 3.5 Methodology – Application of ESNs to Tabular Data.

Reviewer#1, Concern # 6: Results and analysis
The different classes are not annotated consistently. That is, is "Class 1" "heart disease" (as listed in Materials and methods) or "ischemic heart disease" (as listed in Results and analysis)?

Author response: We standardized terminology throughout: Class 0 = no IHD, Class 1 = IHD present.
Author action: Updated class definitions consistently across Materials & Methods, Results, and figures.

Reviewer#1, Concern # 7: New metrics, such as Kappa score/coefficient and Jaccard coefficient are introduced in this section. They ought to be introduced in Materials and methods.
Author response: We thank the reviewer. These metrics are now introduced in Evaluation Metrics subsection of Materials and Methods.
Author action: Section 3.5 includes definitions of Kappa coefficient and Jaccard index.

Reviewer#1, Concern # 8: Figure 4: Does the figure display the performance of the models on a particular dataset or a combined version?
Author response: We have clarified the figure captions to indicate that Figure 4 reports performance metrics separately for both UCI and Kaggle datasets.
Author action: Updated Figure 4 caption as suggested

Reviewer#1, Concern # 9: For the performance metrics, it would be beneficial to include confidence intervals for assessment of statistical significance.
Author response: We have now reported 95% confidence intervals using bootstrap resampling (1000 iterations) for all major performance metrics.
Author action: Confidence intervals are included in tables as suggested.

Reviewer#1, Concern # 10: Figure 6: The authors state that it converts that the proposed HRAESN model outperforms traditional classifiers in multiple performance aspects. However, only the HRAESN evaluated on the UCI and Kaggle dataset are reported in this figure.
Author response: We agree this was ambiguous. Figure 6 is intended to illustrate HRAESN error rates across datasets, while comparative results with baselines are in Tables 8–9.
Author action: Figure 6 caption updated to clarify scope

Reviewer#1, Concern # 11: Were the test and training sets similar? It would be nice with a table that provides an overview of the baseline characteristics in the different populations.
Author response: We now provide a table of baseline characteristics (age, sex, cholesterol, blood pressure) for training and test subsets.
Author action: Added Table 3: Baseline Characteristics.

Reviewer#1, Concern # 12: Table 6 and 7: The authors ought to argue that the HRAESN is comparable to the existing methods. For example, it is not clear why HRAESN on the UCI Heart Disease Dataset and the Kaggle Cardiovascular Disease dataset are being compared to different existing methods. Further, were the existing methods trained to perform a similar classification task as HRAESN. Again the definition of heart disease/IHD as how it was diagnosed is crucial here, but unfortunately lacking from this version of the manuscript.

Author response: We appreciate this comment. Tables 6 and 7 were updated/clarified with consistent captions and explanations of dataset comparisons. We expanded the Discussion to address:

Lack of external validation and need for future hospital-based datasets.
Potential imputation bias and future use of sensitivity analyses.
Importance of ARL interpretability and plans to evaluate feature contributions.

Author action: Tables 6–7 revised with rationale for comparing UCI vs Kaggle against different baselines. Expanded Section 6 Limitations and Future Directions.

Reviewer#1, Concern # 13: Discussion
This section appears to be incomplete. For example, the lack of external validation is not addressed. Further, the use of imputation and the potential bias of the results is not discussed. Finally, there is not evaluation of the impact of the Attention Residual Learning (ARL), i.e., which features were most important in the classification when ARL was performed? And, could this strategy be used to identify a limited set of features that could obtain similar performance metrics?
Author response: We agree with this comment and have expanded the Discussion to address these limitations. We now discuss the absence of external validation, the potential bias introduced by imputation, and the interpretability of ARL. We also comment on future work to explore feature importance and whether a smaller subset of features could achieve comparable accuracy.
Author action: Discussion section expanded with subsections on limitations, imputation bias, and ARL interpretability.

We sincerely thank Reviewer for the constructive and rigorous feedback. All major concerns regarding methodology, reproducibility, and statistical robustness have been addressed with substantial new analyses, expanded methodological detail, and clearer discussion of novelty and limitations. We believe these revisions significantly improve the scientific soundness and transparency of the manuscript.

View more View less

Competing Interests

The author(s) declare that they have no competing interests.

Alongside their report, reviewers assign a status to the article:

Approved - the paper is scientifically sound in its current form and only minor, if any, improvements are suggested

Approved with reservations - A number of small changes, sometimes more significant revisions are required to address specific details and improve the papers academic merit.

Not approved - fundamental flaws in the paper seriously undermine the findings and conclusions

[1] 1. Severino P, et al.: Ischemic Heart Disease Pathophysiology Paradigms Overview: From Plaque Activation to Microvascular Dysfunction. Int. J. Mol. Sci. Oct. 2020; 21: 8118. PubMed Abstract | Publisher Full Text | Free Full Text

[2] 2. Cardiovascular diseases (CVDs). (accessed Mar. 16, 2023). Reference Source

[3] 3. Bolhasani H, Mohseni M, Rahmani AM: Deep learning applications for IoT in health care: A systematic review. Inform. Med. Unlocked. Jan. 2021; 23: 100550. Publisher Full Text

[4] 4. Introduction to Recurrent Neural Network - GeeksforGeeks. (accessed Mar. 15, 2023). Reference Source

[5] 5. Gao R, Du L, Duru O, et al.: Time series forecasting based on echo state network and empirical wavelet transformation. Appl. Soft Comput. Apr. 2021; 102: 107111. Publisher Full Text

[6] 6. Huang Z, et al.: Functional deep echo state network improved by a bi-level optimization approach for multivariate time series classification. Appl. Soft Comput. Jul. 2021; 106: 107314. Publisher Full Text

[7] 7. Mhathesh TSR, Andrew J, Martin Sagayam K, et al.: A 3d convolutional neural network for bacterial image classification. Adv. Intell. Syst. Comput. Springer; 2021; pp. 419–431. Publisher Full Text

[8] 8. Wang F, et al.: Residual attention network for image classification. Proc. IEEE Conf. Comput. Vis. Pattern Recognit. 2017; 3156–3164.

[9] 9. Cenitta D, Arjunan RV, Prema KV: Ischemic Heart Disease Multiple Imputation Technique Using Machine Learning Algorithm. Eng. Sci. Sep. 2022; 19: 262–272. Publisher Full Text

[10] 10. Kusuma S, Jothi KR: Heart disease classification using multiple K-PCA and hybrid deep learning approach. Comput. Syst. Sci. Eng. 2022; 41(3): 1273–1289. Publisher Full Text

[11] 11. Nagavelli U, Samanta D, Chakraborty P: Machine Learning Technology-Based Heart Disease Detection Models. J. Healthc. Eng. 2022; 2022: 1–9. PubMed Abstract | Publisher Full Text | Free Full Text

[12] 12. Sonawane R, Patil HD: Prediction of Heart Disease by Optimized Distance and Density-Based Clustering. Proceedings of the 2nd International Conference on Artificial Intelligence and Smart Energy, ICAIS 2022. Institute of Electrical and Electronics Engineers Inc; 2022; pp. 1001–1008. Publisher Full Text

[13] 13. Cardiovascular Disease dataset|Kaggle. (accessed Mar. 15, 2023). Reference Source

[14] 14. Sonawane R, Patil H: Automated heart disease prediction model by hybrid heuristic-based feature optimization and enhanced clustering. Biomed. Signal Process. Control. Feb. 2022; 72: 103260. Publisher Full Text

[15] 15. Archana KS, Sivakumar B, Kuppusamy R, et al.: Automated Cardioailment Identification and Prevention by Hybrid Machine Learning Models. Comput. Math. Methods Med. 2022; 2022: 1–8. PubMed Abstract | Publisher Full Text | Free Full Text

[16] 16. Li X, et al.: Automatic heartbeat classification using S-shaped reconstruction and a squeeze-and-excitation residual network. Elsevier, Comput. Biol. Med. 2022; 140: 105108. PubMed Abstract | Publisher Full Text

[17] 17. Sun X, Li T, Li Q, et al.: Deep belief echo-state network and its application to time series prediction. Knowl.-Based Syst. Aug. 2017; 130: 17–29. Publisher Full Text

[18] 18. Wang Q, Wang L, Liu Y, et al.: Time Series Prediction with Incomplete Dataset Based on Deep Bidirectional Echo State Network. IEEE Access. 2019; 7: 152533–152544. Publisher Full Text

[19] 19. Ren W, Wang Y, Han M: Time series prediction based on echo state network tuned by divided adaptive multi-objective differential evolution algorithm. Soft. Comput. Mar. 2021; 25(6): 4489–4502. Publisher Full Text

[20] 20. Doppala BP, Bhattacharyya D, Janarthanan M, et al.: A Reliable Machine Intelligence Model for Accurate Identification of Cardiovascular Diseases Using Ensemble Techniques. J. Healthc. Eng. 2022; 2022: 1–13. PubMed Abstract | Publisher Full Text | Free Full Text

[21] 21. Ampavathi A, Saradhi TV: Multi disease-prediction framework using hybrid deep learning: an optimal prediction model. Comput. Methods Biomech. Biomed. Engin. 2021; 24(10): 1146–1168. PubMed Abstract | Publisher Full Text

[22] 22. Liu Y, et al.: Automatic Detection of ECG Abnormalities by Using an Ensemble of Deep Residual Networks with Attention. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). Springer; 2019; pp. 88–95. Publisher Full Text

[23] 23. Zhang A, Zhu W, Li J: Spiking echo state convolutional neural network for robust time series classification. IEEE Access. 2019; 7: 4927–4935. Publisher Full Text

[24] 24. Guo C, Zhang J, Liu Y, et al.: Recursion Enhanced Random Forest with an Improved Linear Model (RERF-ILM) for Heart Disease Detection on the Internet of Medical Things Platform. IEEE Access. 2020; 8: 59247–59256. Publisher Full Text

[25] 25. Li B, Li Z, Yang Y: Residual attention graph convolutional network for web services classification. Neurocomputing. Jun. 2021; 440: 45–57. Publisher Full Text

[26] 26. Suresh T, Assegie TA, Rajkumar S, et al.: A hybrid approach to medical decision-making: diagnosis of heart disease with machine-learning model. Int. J. Electr. Comput. Eng. 2022; 12(2): 1831–1838. Publisher Full Text

[27] 27. Bhavekar GS, Das Goswami A: A hybrid model for heart disease prediction using recurrent neural network and long short term memory. Int. J. Inf. Technol. (Singapore). Jun. 2022; 14(4): 1781–1789. Publisher Full Text

[28] 28. Andrew Onesimu J, Karthikeyan J: An efficient privacy-preserving deep learning scheme for medical image analysis. Journal of Information Technology Management, vol. 12, no. Special Issue: The Importance of Human Computer Interaction: Challenges, Methods and Applications. Dec. 2021; 50–67. Publisher Full Text

[29] 29. Andrew J, Fiona R, Caleb Andrew H: Comparative study of various deep convolutional neural networks in the early prediction of cancer. 2019 International Conference on Intelligent Computing and Control Systems, ICCS 2019. Institute of Electrical and Electronics Engineers Inc.; May 2019; pp. 884–890. Publisher Full Text

[30] 30. Chandrasekaran ST, Banerjee I, Sanyal A: 7.5nJ/inference CMOS Echo State Network for Coronary Heart Disease prediction. ESSDERC 2021-IEEE 51st European Solid-State Device Research Conference (ESSDERC). Sep. 2021; pp. 103–106. Publisher Full Text

[31] 31. Maiga J, Hungilo GG, et al.: Comparison of Machine Learning Models in Prediction of Cardiovascular Disease Using Health Record Data. International Conference on Informatics, Multimedia, Cyber and Information System. 2019; pp. 45–48. Publisher Full Text

[32] 32. Bharti R, Khamparia A, Shabaz M, et al.: Prediction of Heart Disease Using a Combination of Machine Learning and Deep Learning. Comput. Intell. Neurosci. 2021; 2021. PubMed Abstract | Publisher Full Text | Free Full Text

[33] 33. UCI Machine Learning Repository: Heart Disease Data Set. (accessed Mar. 15, 2023). Reference Source

[34] 34. Cenitta D, Vijaya Arjunan R, Prema KV: Ischemic Heart Disease Prediction Using Optimized Squirrel Search Feature Selection Algorithm.Publisher Full Text

[35] 35. Verma L, Mathur MK: Deep learning based model for decision support with case based reasoning. International Journal of Innovative Technology and Exploring Engineering. 2020; 8(6C): 149–153.

[36] 36. Latha CBC, Jeeva SC: Improving the accuracy of prediction of heart disease risk based on ensemble classification techniques. Inform. Med. Unlocked. Jan. 2019; 16: 100203. Publisher Full Text

[37] 37. Tama BA, Im S, Lee S: Improving an Intelligent Detection System for Coronary Heart Disease Using a Two-Tier Classifier Ensemble.2020. Publisher Full Text

[38] 38. Rani P, Kumar R, Ahmed NMOS, et al.: A decision support system for heart disease prediction based upon machine learning. J. Reliab. Intell. Environ. 2021; 7: 263–275. Publisher Full Text

[39] 39. Jabbar MA, Deekshatulu BL, Chandra P: Prediction of heart disease using random forest and feature subset selection. Adv. Intell. Syst. Comput. 2016; 424: 187–196. Publisher Full Text

[40] 40. Hagan R, Gillan CJ, Mallett F: Comparison of machine learning methods for the classification of cardiovascular disease. Inform. Med. Unlocked. Jan. 2021; 24: 100606. Publisher Full Text

[41] 41. Bhoyar S, Wagholikar N, Bakshi K, et al.: Real-time Heart Disease Prediction System using Multilayer Perceptron; Real-time Heart Disease Prediction System using Multilayer Perceptron. International Conference for Emerging Technology. 2021. Publisher Full Text

[42] 42. Theerthagiri P, Vidya J: Cardiovascular disease prediction using recursive feature elimination and gradient boosting classification techniques. Expert. Syst. 2022; 39. Publisher Full Text

[43] 43. Uddin MN, Halder RK: An ensemble method based multilayer dynamic system to predict cardiovascular disease using machine learning approach. Inform. Med. Unlocked. Jan. 2021; 24: 100584. Publisher Full Text

Deep learning based hybrid residual attention and echo state network for high-accuracy heart disease prediction

Abstract

Background

Methods

Results

Conclusions

Keywords

Revised Amendments from Version 1

1. Introduction

1.1 Objective of this study

1.2 Problem statement

2. Related works

3. Materials and methods

3.1 Dataset

Table 1. Kaggle cardiovascular disease dataset description.

Table 2. UCI heart disease dataset description.

Table 3. Baseline characteristics of training and testing populations for the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

3.2 Hybrid data classification algorithm

3.3 Echo State Network (ESN)

3.4 Attention Residual Learning (ARL)

3.5 Methodology

Figure 1. Workflow of the proposed experiment using the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

Figure 2. Overall system model of the proposed Hybrid Residual Attention with Echo State Network (HRAESN).

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Algorithm. Hybrid Residual Attention with Echo State Network (HRAESN).

Table 4. Summary of hyperparameter settings used in the proposed model training.

4. Results and analysis

4.1 Experiment setup and data preprocessing

Figure 3. PCA plot showing data distribution in the heart disease dataset based on the first two principal components.

4.2 Experimental results

Table 5. Normalized confusion matrix for the Hybrid Residual Attention with Echo State Network (HRAESN) using the UCI Heart Disease dataset.

Table 6. Normalized confusion matrix for the Hybrid Residual Attention with Echo State Network (HRAESN) using the Kaggle Cardiovascular Disease dataset.

Figure 4. Analysis of classifier performance based on sensitivity, specificity, precision, F-measure, and accuracy.

Figure 5. Analysis of classifier performance using Kappa coefficient, Recall, and Jaccard coefficient for the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

Figure 6. Classification error rate, false acceptance rate (FAR), and false rejection rate (FRR) for the UCI Heart Disease and Kaggle Cardiovascular Disease datasets.

4.3 Comparative analysis with existing models

Table 7. Comparison of HRAESN with existing methods using the UCI heart disease dataset.

Table 8. Comparison of HRAESN with existing methods using the Kaggle cardiovascular disease dataset.

Figure 7. Comparison of residual network, echo state network, and the proposed Hybrid Residual Attention Echo State Network.

5. Discussion

Table 9. Performance comparison of different algorithms.

Table 10. Performance of deep learning classifiers on the heart disease dataset.

6. Limitations and future directions

7. Conclusion

Ethical statement

CRediT authorship contribution statement

Disclaimer/publisher’s note

Data availability

Acknowledgments

References

Comments on this article Comments (0)

Open Peer Review

Comments on this article Comments (0)

Open Peer Review

Reviewer Status

Reviewer Reports

Comments on this article

Browse by related subjects

Competing Interests Policy

Stay Updated