F1000ResearchF1000Research2046-1402F1000 Research LimitedLondon, UK10.12688/f1000research.165575.2Research ArticleArticlesDeep 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]
DCenittaMethodologyProject Administrationhttps://orcid.org/0000-0003-3715-69411RanganathanVIijaya ArjunanConceptualizationWriting – Review & Editinghttps://orcid.org/0000-0002-1402-6573a1ShaileshTanujaWriting – Review & Editingb1JAndrewData Curation1NArulVisualization2TPraveen PaiWriting – Review & Editing1
Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, India
Computer Science and Engineering, AJ Institute of Engineering and Technology, Mangalore, Karnataka, Indiavijay.arjun@manipal.edutanuja.s@manipal.edu
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.
UCIKaggleHeart DiseaseImputationDeep LearningEcho State NetworkResidual Attention.The author(s) declared that no grants were involved in supporting this work.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.
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.
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?
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?
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 bidirectional
18 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 Networks
23 and deep residual attention models
8,24 and Residual Attention Graph Convolutional Networks
25 demonstrated improvements in complex classification tasks. Feature refinement approaches such as Recursion-Enhanced Random Forest
26 and SVM-based ensembles with feature elimination
27 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 methods3.1 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:
Objective Characteristics (Real patient data): Gender, Age, Height, and Weight
Features of the Examination (Medical Test Results): Blood Pressure Systolic and Diastolic, Blood Pressure Levels of Cholesterol and Glucose
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.
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 works
31,34–37 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.
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:
Echo State Networks (ESNs) for efficient time-series processing
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:
Hi(t+1)=(1+Mi(t))∗XFi(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:
Relevant features are amplified, while irrelevant features are suppressed.
Deep residual networks maintain stable performance even with hundreds of layers.
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.
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.
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
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
Feature extraction and classification: The HRAESN model applies ESNs for sequence modeling and ARL for refining feature representation.
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 DynamicsXF(t+1)=fa(Wiu(t)+WrXF(t))
Where:
XF(t+1) and
XF(t) are the feature matrices at iterations t and t + 1.
Wi is the input reservoir weight matrix derived from the input data.
Wr is the reservoir weight matrix representing internal states.
u(t+1) represents the internal states computed at iteration t.
fa(.) is the activation function applied at the reservoir.
Mi(t) is the gradient of the input feature mask at iteration t.
Hi(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:
XF1(t+1)=fa(Wiu(t)+W1XF1(t))XF2(t+1)=fa(WiXF1(t)+W2XF2(t))XFM(t+1)=fa(WiXF(M−1)(t)+WMXFM(t))
Where:
Wi=Hi(t+1) represents the attention module output.
Activation Functions and Output Computation
Final Activation Function
An=YL·sigmoid(XFM(t+1))
Where:
sigmoid(.) is the activation function applied to the final output layer.
Dynamic Echo State Network Output
PR(t+1)=ga(WoXFM(t+1))
Where:
Wo represents the output reservoir weight matrix.
ga(.) is the final activation function used at step 4.
Hybrid Residual Attention with Echo State Network (HRAESN).
Input: features data
XF, label data
YL
Output: Predicted result P
r
1:
begin
2:
for each Compute the Hidden layer of dynamic ESN
3:
XF(t+1)=fa(Wiu(t)+WrXF(t))
4:
end for
5:
for each compute the attention residual learning
6:
Hi(t+1)=(1+Mi(t))∗XFi(t)
7:
end for
8:
for x=1 to M
do:
9:
XF1(t+1)=fa(Wiu(t)+W1XF1(t))
10:
XF2(t+1)=fa(WiXF1(t)+W2XF2(t))
11: …
12:
XFM(t+1)=fa(WiXFM−1(t)+WMXFM(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.
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.
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.
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
1
3
136
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
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.
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
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.
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.
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 & Mathur
35
2019
Rows with missing values deleted
MLP
85.48
Latha & Jeeva
36
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
Comparison of HRAESN with existing methods using the Kaggle cardiovascular disease dataset.
Study
Year
Classifiers
Accuracy (%)
Maiga et al.
31
2019
RF
73
Hagan
40
2021
RF, Gradient Boosting
74
Bhoyar
41
2021
MLP
89.7
Theerthagiri
42
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.
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:
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)
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
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.
ReferencesSeverinoP:
Ischemic Heart Disease Pathophysiology Paradigms Overview: From Plaque Activation to Microvascular Dysfunction.Int. J. Mol. Sci.Oct. 2020;21:8118.
3314325610.3390/IJMS21218118PMC7663258Cardiovascular diseases (CVDs). (accessed Mar. 16, 2023).
Reference SourceBolhasaniHMohseniMRahmaniAM:
Deep learning applications for IoT in health care: A systematic review. Inform. Med. Unlocked.Jan. 2021;23:100550.
10.1016/j.imu.2021.100550Introduction to Recurrent Neural Network - GeeksforGeeks.(accessed Mar. 15, 2023).
Reference SourceGaoRDuLDuruO:
Time series forecasting based on echo state network and empirical wavelet transformation.Appl. Soft Comput.Apr. 2021;102:107111.
10.1016/j.asoc.2021.107111HuangZ:
Functional deep echo state network improved by a bi-level optimization approach for multivariate time series classification.Appl. Soft Comput.Jul. 2021;106:107314.
10.1016/j.asoc.2021.107314MhatheshTSRAndrewJMartin SagayamK:
A 3d convolutional neural network for bacterial image classification.Adv. Intell. Syst. Comput.Springer;2021; pp.419–431.
10.1007/978-981-15-5285-4_42WangF:
Residual attention network for image classification.Proc. IEEE Conf. Comput. Vis. Pattern Recognit.2017;3156–3164.CenittaDArjunanRVPremaKV:
Ischemic Heart Disease Multiple Imputation Technique Using Machine Learning Algorithm.Eng. Sci.Sep. 2022;19:262–272.
10.30919/ES8D681KusumaSJothiKR:
Heart disease classification using multiple K-PCA and hybrid deep learning approach.Comput. Syst. Sci. Eng.2022;41(3):1273–1289.
10.32604/csse.2022.021741NagavelliUSamantaDChakrabortyP:
Machine Learning Technology-Based Heart Disease Detection Models.J. Healthc. Eng.2022;2022:1–9.
3526530310.1155/2022/7351061PMC8898839SonawaneRPatilHD:
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.
10.1109/ICAIS53314.2022.9742885Cardiovascular Disease dataset|Kaggle. (accessed Mar. 15, 2023).
Reference SourceSonawaneRPatilH:
Automated heart disease prediction model by hybrid heuristic-based feature optimization and enhanced clustering.Biomed. Signal Process. Control.Feb. 2022;72:103260.
10.1016/j.bspc.2021.103260ArchanaKSSivakumarBKuppusamyR:
Automated Cardioailment Identification and Prevention by Hybrid Machine Learning Models.Comput. Math. Methods Med.2022;2022:1–8.
3521119010.1155/2022/9797844PMC8863449LiX:
Automatic heartbeat classification using S-shaped reconstruction and a squeeze-and-excitation residual network.Elsevier, Comput. Biol. Med.2022;140:105108.
3487541010.1016/j.compbiomed.2021.105108SunXLiTLiQ:
Deep belief echo-state network and its application to time series prediction.Knowl.-Based Syst.Aug. 2017;130:17–29.
10.1016/j.knosys.2017.05.022WangQWangLLiuY:
Time Series Prediction with Incomplete Dataset Based on Deep Bidirectional Echo State Network.IEEE Access.2019;7:152533–152544.
10.1109/ACCESS.2019.2948367RenWWangYHanM:
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.
10.1007/s00500-020-05457-8DoppalaBPBhattacharyyaDJanarthananM:
A Reliable Machine Intelligence Model for Accurate Identification of Cardiovascular Diseases Using Ensemble Techniques.J. Healthc. Eng.2022;2022:1–13.
3529968610.1155/2022/2585235PMC8923755AmpavathiASaradhiTV:
Multi disease-prediction framework using hybrid deep learning: an optimal prediction model.Comput. Methods Biomech. Biomed. Engin.2021;24(10):1146–1168.
3342748010.1080/10255842.2020.1869726LiuY:
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.
10.1007/978-3-030-33327-0_11ZhangAZhuWLiJ:
Spiking echo state convolutional neural network for robust time series classification.IEEE Access.2019;7:4927–4935.
10.1109/ACCESS.2018.2887354GuoCZhangJLiuY:
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.
10.1109/ACCESS.2020.2981159LiBLiZYangY:
Residual attention graph convolutional network for web services classification.Neurocomputing.Jun. 2021;440:45–57.
10.1016/j.neucom.2021.01.089SureshTAssegieTARajkumarS:
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.
10.11591/ijece.v12i2.pp1831-1838BhavekarGSDas GoswamiA:
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.
10.1007/s41870-022-00896-yAndrew OnesimuJKarthikeyanJ:
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.
10.22059/jitm.2020.79191AndrewJFionaRCaleb AndrewH:
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.
10.1109/ICCS45141.2019.9065445ChandrasekaranSTBanerjeeISanyalA:
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.
10.1109/ESSCIRC53450.2021.9567753MaigaJHungiloGG:
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.
10.1109/ICIMCIS48181.2019.8985205BhartiRKhampariaAShabazM:
Prediction of Heart Disease Using a Combination of Machine Learning and Deep Learning.Comput. Intell. Neurosci.2021;2021.
3430605610.1155/2021/8387680PMC8266441UCI Machine Learning Repository: Heart Disease Data Set.(accessed Mar. 15, 2023).
Reference SourceCenittaDVijaya ArjunanRPremaKV:
Ischemic Heart Disease Prediction Using Optimized Squirrel Search Feature Selection Algorithm.10.1109/ACCESS.2022.3223429VermaLMathurMK:
Deep learning based model for decision support with case based reasoning.International Journal of Innovative Technology and Exploring Engineering.2020;8(6C):149–153.LathaCBCJeevaSC:
Improving the accuracy of prediction of heart disease risk based on ensemble classification techniques.Inform. Med. Unlocked.Jan. 2019;16:100203.
10.1016/J.IMU.2019.100203TamaBAImSLeeS:
Improving an Intelligent Detection System for Coronary Heart Disease Using a Two-Tier Classifier Ensemble.2020.
10.1155/2020/9816142RaniPKumarRAhmedNMOS:
A decision support system for heart disease prediction based upon machine learning.J. Reliab. Intell. Environ.2021;7:263–275.
10.1007/s40860-021-00133-6JabbarMADeekshatuluBLChandraP:
Prediction of heart disease using random forest and feature subset selection.Adv. Intell. Syst. Comput.2016;424:187–196.
10.1007/978-3-319-28031-8_16/FIGURES/4HaganRGillanCJMallettF:
Comparison of machine learning methods for the classification of cardiovascular disease.Inform. Med. Unlocked.Jan. 2021;24:100606.
10.1016/J.IMU.2021.100606BhoyarSWagholikarNBakshiK:
Real-time Heart Disease Prediction System using Multilayer Perceptron; Real-time Heart Disease Prediction System using Multilayer Perceptron.International Conference for Emerging Technology.2021.
10.1109/INCET51464.2021.9456389TheerthagiriPVidyaJ:
Cardiovascular disease prediction using recursive feature elimination and gradient boosting classification techniques.Expert. Syst.2022;39.
10.1111/exsy.13064UddinMNHalderRK:
An ensemble method based multilayer dynamic system to predict cardiovascular disease using machine learning approach.Inform. Med. Unlocked.Jan. 2021;24:100584.
10.1016/J.IMU.2021.10058410.5256/f1000research.187669.r414780Reviewer response for version 2ShresthaDhadkan1Refereehttps://orcid.org/0009-0007-2588-3734
Texas State University College of Science and Engineering, San Marcos, Texas, USA
Competing interests: No competing interests were disclosed.
Everything looks good now. Thank you for revising.
Is the work clearly and accurately presented and does it cite the current literature?
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
Is the study design appropriate and is the work technically sound?
Partly
Are the conclusions drawn adequately supported by the results?
Partly
Are sufficient details of methods and analysis provided to allow replication by others?
Partly
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.
RANGANATHANVIJAYA ARJUNANSchool of Computer Engineering, Manipal Academy of Higher Education, Manipal, Karnataka, India
Competing interests: The authors declare that they have no competing interests
2592025
Thank you for your kind feedback and approval.
10.5256/f1000research.182263.r403945Reviewer response for version 1MEMONMUHAMMAD HAMMAD1Refereehttps://orcid.org/0000-0002-8680-1831
Southwest University of Science and Technology, Sichuan, China
Competing interests: No competing interests were disclosed.
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
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
Is the study design appropriate and is the work technically sound?
Partly
Are the conclusions drawn adequately supported by the results?
Partly
Are sufficient details of methods and analysis provided to allow replication by others?
Partly
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.
RANGANATHANVIJAYA ARJUNANSchool of Computer Engineering, Manipal Academy of Higher Education, Manipal, Karnataka, India
Competing interests: The author(s) declare that they have no competing interests.
1092025
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.
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.
10.5256/f1000research.182263.r406539Reviewer response for version 1ShresthaDhadkan1Refereehttps://orcid.org/0009-0007-2588-3734
Texas State University College of Science and Engineering, San Marcos, Texas, USA
Competing interests: No competing interests were disclosed.
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.
(c) Methods and Replicability
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
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
Is the study design appropriate and is the work technically sound?
Partly
Are the conclusions drawn adequately supported by the results?
Partly
Are sufficient details of methods and analysis provided to allow replication by others?
Partly
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.
RANGANATHANVIJAYA ARJUNANSchool of Computer Engineering, Manipal Academy of Higher Education, Manipal, Karnataka, India
Competing interests: The author(s) declare that they have no competing interests.
1092025
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.
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.
10.5256/f1000research.182263.r401540Reviewer response for version 1HaueAmalie Dahl1Refereehttps://orcid.org/0000-0001-7656-7976
University of Copenhagen, Copenhagen, Denmark
Competing interests: No competing interests were disclosed.
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
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
Is the study design appropriate and is the work technically sound?
Partly
Are the conclusions drawn adequately supported by the results?
Yes
Are sufficient details of methods and analysis provided to allow replication by others?
Partly
Reviewer Expertise:
Cardiology resident
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.
RANGANATHANVIJAYA ARJUNANSchool of Computer Engineering, Manipal Academy of Higher Education, Manipal, Karnataka, India
Competing interests: The author(s) declare that they have no competing interests.
1092025
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:
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.
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.
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.