Predicting heart disease was once done with convention machine learning algorithms, before deep learning knocked them off the top. Naïve Bayes, Support Vector Machines (SVM), Random Forests (RF) and Logistic Regression (LR) were often applied to the UCI Heart Disease dataset. On data sets describing heart illness, Jabbar et al. found that using RF with feature selection resulted in a much better prediction outcome. ⁸ To determine the most pertinent predictors of IHD, Shah et al. used SVM with embedded feature selection, obtaining comparatively good accuracy with less model complexity. ⁹ These conventional approaches show limits when handling high-dimensional, nonlinear, and temporally dependent health data, notwithstanding their early success. Additionally, they need a lot of manual feature engineering and are unable to recognize the temporal patterns present in ECG data or patient health records. ¹⁰

DL techniques have demonstrated significant potential as medical data and computing power become more accessible. LSTM networks have been successfully utilized to model temporal health data because of its recurrent architecture. By identifying long-term relationships in patient vitals and historical data, Bhavekar and Goswami’s hybrid RNN-LSTM model demonstrated improved accuracy on datasets related to heart disease. ¹¹ For the identification of cardiovascular disease, Doppala et al. presented an ensemble-based DL model that uses several learning techniques and shown enhanced generalization and resilience. ¹²

Furthermore, hybrid models that blend DL with traditional ML or soft computing techniques have drawn interest. For the diagnosis of IHD, Suresh et al. suggested a hybrid SVM-RF model that optimizes model inputs through recursive feature removal. ¹³ To optimize predictions for a variety of biomedical datasets, Ampavathi and Saradhi presented a multi-disease prediction system built on a hybrid deep learning architecture. ¹⁴ Although these models outperform conventional machine learning techniques, most of them are not interpretable or generalizable to other patient groups. Furthermore, the issue of model bias and fairness is not sufficiently addressed by many.

Attention mechanisms and residual learning have recently been incorporated into DL designs, especially for time-series data like as ECG signals. By focusing on important segments of the input sequence, attention enables models to increase interpretability and accuracy. Conversely, residual learning makes it possible to train deeper architectures effectively without vanishing gradients, which enhances training dynamics. For the purpose of identifying ECG anomalies, Liu et al. created an ensemble of residual networks with attention, which demonstrated better classification results on cardiac datasets. ¹⁵ With a 97.7% accuracy rate, Cenitta et al.’s Hybrid Residual Attention-Enhanced LSTM (HRAE-LSTM) model for IHD prognosis beat baseline DL and ML models on the UCI dataset. ¹⁶ These models demonstrate the effectiveness of integrating attention mechanisms with residual connections. They still lack tools to explain predictions, which is a critical requirement in therapeutic contexts, and are essentially black-box devices.

To make medical AI more understandable, Explainable AI approaches have become very important tools. Such methods comprise integrated gradients, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) and are often applied to explain the importance of features in DL models. Using SHAP and a two-tier ensemble model, Tama et al. allowed practitioners to see how important features such as chest pain and cholesterol were for predicting heart disease. ¹⁷ The approach introduced by Andrew and Karthikeyan uses privacy-protecting XAI to combine how private data should be and how a model is explained for image analysis. ¹⁸ According to the results, it is more important now to use models that are clear and practical. In contrast to other models created for post hoc explanation, ours makes SHAP explanations part of building the model, helping provide instant explanations for each prediction.

In healthcare applications, transfer learning has emerged as a successful tactic for enhancing model generalization and overcoming data shortage. To save training time and increase accuracy, it entails recycling weights from models that have been trained on big datasets for new, smaller domains.

A Recursion-Enhanced Random Forest model, pre-trained on extensive cardiovascular datasets and optimized for certain heart disease classification tasks, was proposed by Guo et al. ¹⁹ For better arrhythmia classification, Li et al. used a squeeze-and-excitation residual network pretrained on generic heartbeat datasets. ²⁰ Performance on target datasets that would otherwise be too small to train sophisticated DL models was greatly enhanced in both situations by transfer learning. The suggested model improves its performance on smaller datasets, such as UCI or actual hospital data, by initializing BiLSTM layers with cardiovascular domain knowledge through transfer learning.

Concerns regarding algorithmic bias have grown in significance as AI systems become more and more integrated into healthcare decision-making. Research has demonstrated that a number of models operate differently for different genders, ages, and ethnicities, which results in unfair treatment outcomes. ²¹ Although fairness was not specifically assessed, Sonawane and Patil suggested a hybrid heuristic-based clustering technique to balance performance across demographic categories. ²² To guarantee equitable healthcare delivery, Rani et al. demanded that fairness criteria be incorporated into machine learning algorithms used for clinical risk prediction. ²³ By using demographic reweighting during training, the model directly addresses this problem with the goal of lowering prediction bias and enhancing equity across a range of demographics.

Using a range of ML and DL approaches, the body of existing work shows excellent progress in IHD prognosis. Nonetheless, several restrictions still exist:

Attention-residual models continue to lack integrated interpretability. •

Traditional machine learning models lack scalability and temporal awareness

Our suggested model fills these shortcomings by integrating bias mitigation techniques, SHAP-based interpretability, transfer learning, and residual attention-enhanced BiLSTM into a single framework, creating a novel and comprehensive method for IHD diagnosis.

The UCI Heart Disease Dataset, a popular benchmark dataset in cardiovascular research that is openly accessible via the UCI Machine Learning Repository, is used in this study. 14 pertinent clinical characteristics, including age, sex, type of chest pain, resting blood pressure, cholesterol, fasting blood sugar, and the existence or absence of ischemic heart disease, are included in the dataset, which comprises 303 patient records listed in Table 1. The binary target variable indicates if cardiac disease is present (1) or not (0). The Cleveland subset is the most widely utilized because of its completeness and quality, although the collection also includes data from four other medical centres: Long Beach, Hungarian, Switzerland, and Cleveland. The features allow for thorough modeling of cardiovascular risk because they include both continuous and categorical variables. Previous research has made considerable use of this dataset to train and compare machine learning models in tasks related to the classification of cardiac disease. ²⁴

3.1.1 Ethical considerations

This study was conducted using the publicly available UCI Heart Disease dataset, ²⁴ which comprises anonymized and de-identified data. No new data collection or human subject interaction was performed by the authors. Therefore, ethical approval and informed consent were not required for this secondary data analysis. The original data were collected and published in accordance with institutional guidelines and the principles of the Declaration of Helsinki.

Preprocessing is essential for guaranteeing data quality, particularly in medical datasets where mixed data types, missing values, and inconsistencies are prevalent. Preprocessing was meticulously planned for this project to get the UCI Heart Disease dataset ²⁴ ready for deep learning model training.

3.2.1 Missing value handling

Many medical records are incomplete because patients have left the study, forgotten to record results or given inconsistent information. There are no measurements for thal (thalassemia) and ca (number of coloured main vessels seen by fluoroscopy) in several of the records in the UCI Heart Disease dataset. ²⁴ Since removing these entries or working on them with simple strategies like mean or mode imputation is risky, try using a fuzzy-based multiple imputation strategy. Fuzzy logic helps this approach find several legitimate responses for each missing data point which are then averaged to improve the accuracy of the estimate. Using fuzzy sets, multiple imputation handles uncertainty better than plain deterministic methods. It removes unfair bias, allows the model to use data in ways it is trained for and keeps the main statistical qualities of the dataset. This design assists clinicians because losing critical patient information can be catastrophic if data is cleared. ²⁵

3.2.2 Normalization

There are clinical characteristics in the data such as age, cholesterol and resting blood pressure that researchers can plot as numbers on charts. Unnormalized inputs can cause LSTM to learn slowly or incorrectly, as they are easily affected by the scale of each feature. For this reason, figures were input as numbers between 0 and 1 following the following formula: X scaled = X − X min X max − X min

This method guarantees that every feature contributes equally to model learning while maintaining the original distribution’s form. In LSTM networks, where different feature sizes can adversely affect convergence and learning stability, normalization is particularly crucial. ²⁶

3.2.3 Categorical encoding

The dataset contains several categorical important variables, such as thalassemia (thal), the slope of the ST segment (slope), and the type of chest pain (cp). One-Hot Encoding was used to convert these categorical variables, turning each distinct category into a binary vector. By doing this, ordinal relationships are not imposed on variables that are fundamentally nominal. Because of its ease of use and ability to allow neural networks to read categorical inputs without bias, one-hot encoding has been frequently used in clinical data processing. ²⁷ The feature space becomes entirely compatible with the downstream neural network design after encoding, although its dimensionality grows.

To effectively capture temporal dependencies, highlight important features through attention, encourage model generalization through transfer learning, and guarantee interpretability through Explainable AI (XAI) techniques like SHAP, the suggested model, X-TLRABiLSTM (Explainable Transfer Learning-based Residual Attention Bidirectional LSTM), was created. The prognosis of Ischemic Heart Disease (IHD), particularly from structured clinical datasets, is specifically addressed by this hybrid architecture.

3.3.1 Architecture overview

The Figure 1 indicates the X-TLRABiLSTM model incorporates the following critical components: 1.

Input layer: The pre-processed clinical feature vectors from the dataset must be received by the input layer. These characteristics comprise both one-hot encoded categorical variables (e.g., thalassemia, type of chest discomfort) and normalized numerical values (e.g., age, blood pressure, cholesterol). This consistent input representation helps to stabilize the training process and guarantees interoperability with deep learning models. When appropriate, it preserves the time dimension in the data, which serves as the foundation for sequential modeling.

Where M ( t ) is the attention mask, and ⊙ denotes element-wise multiplication. 5.

Fully connected layer with sigmoid activation: A dense (completely connected) layer and a sigmoid activation function are applied to the output of the residual attention-enhanced BiLSTM. By doing this, the high-dimensional latent vector is converted into a single probability value that represents the patient’s chance of developing ischemic heart disease: y ̂ = σ ( W o · H ( T ) + b o )

Where σ is the sigmoid function, y ̂ ∈ [ 0 , 1 ] is, the predicted probability, and H ( T ) is the final hidden state. 6.

Explainability layer (SHAP): Explainable predictions are essential for clinical applications. Local interpretability is produced by integrating the SHAP (SHapley Additive exPlanations) approach, which explains how each feature such as age, thalassemia, and type of chest pain contributes to a particular prediction. ³⁰ Based on game theory, SHAP rates each feature according to its marginal contribution to the model’s output. This promotes informed decision-making and builds confidence in AI recommendations by assisting physicians in understanding why the model identified a patient as high-risk.

Where w i reflects demographic importance. This adjustment improves equity in model performance, reducing disparities in prediction accuracy between groups.

3.3.2 Mathematical formulation

Let: •

X ( t ) ∈ ℝ n denote the input feature vector at time t

Equation 1: BiLSTM Representation h t = [ h f ( t ) ; h b ( t ) ]

Where [ · ; · ] denotes vector concatenation.

Equation 2: Residual Feature Update X F ( t + 1 ) = f a ( W i X ( t ) + W r X F ( t ) ) + λ ( X F ( t ) − f a ( W i X ( t ) + W r X F ( t ) ) )

Here, f a is the activation function (e.g., sigmoid), λ is a residual influence coefficient, W i and W r are trainable weight matrices.

Equation 3: Attention Mask Update H ( t + 1 ) = ( 1 + M ( t ) ) ⊙ X F ( t ) + Residual ( t )

Where M ( t ) is the attention mask matrix and ⊙ is element-wise multiplication.

Equation 4: Final Output Calculation y ̂ = σ ( W o · H ( T ) + b o )

Where σ is the sigmoid activation, and y ̂ ∈ [ 0 , 1 ] denotes the probability of heart disease.

3.3.3 Algorithm: X-TLRABiLSTM training

3.3.4 Explainability and fairness

The model incorporates SHAP (SHapley Additive exPlanations) to explain the prediction for everyone by attributing the prediction to input features. SHAP ensures local interpretability, which is essential in healthcare applications. ³⁰ Furthermore, demographic reweighting is applied to the loss function to address class imbalance across subgroups such as gender and age. Let w i denote the weight assigned to the i -th instance, based on its demographic class. The reweighted binary cross-entropy loss becomes:

Equation 5: Fairness-Aware Loss L = − ∑ i = 1 N w i [ y i log ( y ̂ i ) + ( 1 − y i ) log ( 1 − y ̂ i ) ]

Ensuring fairness in model predictions is essential in healthcare applications, especially when working with unbalanced demographic groups. Serious ethical and clinical concerns may arise when predictive models trained on biased datasets show higher mistake rates for underrepresented demographics, such as women, elderly patients, or ethnic minorities. ³³ used a demographic reweighting technique during model training to solve this. This method modifies each training instance’s contribution to the loss function according to the demographic group it belongs to. Let G denote the set of demographic groups (e.g., gender, age category), and p ( g ) be the proportion of group g in the dataset. The weight w i for a sample x i belonging to group g i is defined as: w i = 1 p ( g i )

In order to balance the model’s learning and lessen unequal performance across populations, this inverse-frequency weighting makes sure that minority groups are given greater priority during training. ³⁴ The ultimate training loss is as follows: L = − ∑ i = 1 N w i [ y i log ( y ̂ i ) + ( 1 − y i ) log ( 1 − y ̂ i ) ]

In clinical decision support systems, where fairness and trust are crucial, this fairness-aware loss motivates the model to produce more equitable predictions. ³⁵

As seen in Table 2, carried out an extensive grid search across a variety of hyperparameters to attain the best model performance. Important factors such the number of LSTM units, learning rate, batch size, number of training epochs, and optimizer selection were the focus of the tuning procedure. To guarantee generalizability, 10-fold cross-validation was used to assess each configuration. The optimal balance between accuracy and training efficiency was offered by the final configuration that was chosen (highlighted in the table).

This systematic tuning process is consistent with recent best practices in medical deep learning. ³⁶ To employ a variety of classification and fairness metrics, such as accuracy, precision, recall, specificity, F1-score, AUC-ROC, and fairness gap, which are very pertinent in medical AI systems, to thoroughly assess the model’s performance. ^{37, 38} Ten-fold cross-validation was used to average all metrics to guarantee statistical robustness and dependability. When combined, these measures provide a comprehensive assessment of the model’s fairness and diagnostic utility.

The accuracy, interpretability, generalizability, and fairness of ischemic heart disease (IHD) prognostic systems are all improved by the various changes introduced by the suggested X-TLRABiLSTM model. The following are the main improvements: •

Integration of transfer learning: The model makes use of information from a BiLSTM model that has already been trained on a sizable cardiovascular dataset. By transferring domain-specific representations, this allows for efficient generalization, even on smaller or unbalanced clinical samples.

The classification performance of X-TLRABiLSTM is contrasted with several benchmark models in Table 3, such as Residual Attention BiLSTM (RA-BiLSTM), Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and conventional LSTM. Tenfold cross-validation was used to assess performance.

In every evaluation metric, the X-TLRABiLSTM fared better than any competitor model. Its remarkable 98.2% accuracy and 99.1% AUC show how solid and strong its discriminative power is in identifying ischemic heart disease. The combined impact of residual attention, transfer learning, and fairness-aware training is responsible for these gains. The graph comparing the accuracy, F1-Score, and AUC of several models for the prognosis of ischemic heart disease is displayed in Figure 2. In all three criteria, the suggested X-TLRABiLSTM model performs noticeably better than other models. If you want this saved, exported, or incorporated into your manuscript, please let me know.

Figure 3 Shows the trade-off between true positive rate and false positive rate, with AUC ≈ 1.00 for the X-TLRABiLSTM model.

Figure 4 illustrates the exceptional classification performance of the confusion matrix for the proposed X-TLRABiLSTM model in ischemic heart disease (IHD) identification. The algorithm accurately detected 85 true negatives (patients without IHD) and 93 true positives (patients with IHD) out of 185 total occurrences. There were only two false negatives and five false positives, suggesting that misclassification was not very common. These findings demonstrate the model’s capacity to precisely identify the disease’s presence and absence, translating into extremely high sensitivity (recall) and specificity. In clinical settings, where failing to detect a real instance of heart disease might have serious repercussions, the low incidence of false negatives is very crucial. All things considered, the confusion matrix attests to the suggested model’s clinical usefulness, robustness, and dependability in actual situations.

The most significant features were identified by computing SHAP values for each prediction to verify the explainability of the model. The top contributing features identified by the global SHAP summary graphic were as follows: •

Chest Pain Type (cp)

The SHAP summary is depicted in Figure 5, where a low value is indicated by blue and a high feature value by red. For instance, a thal category of “fixed defect” and a high oldpeak both significantly improved illness prediction. By supporting each prediction with verifiable clinical data, the SHAP plots increase physician trust and adhere to medical AI explainability criteria. ³⁰

F1-score parity between several demographic categories (male vs. female, age < 50 vs. ≥ 50) will be used to assess fairness. To compute the Fairness Gap (ΔF1), the absolute difference in F1-score between subgroups was used.

The suggested model achieves nearly equal predictive accuracy across the gender and age subgroups displayed in Table 4, according to the Fairness Gap values (≤ 0.6). Demographic reweighting during the training phase, which addresses biases seen in previous publications, ^{31, 33} is directly responsible for this. Figure 6 shows the F1-scores across demographic groups (Male, Female, Age < 50, Age ≥ 50), demonstrating balanced performance.

Across several performance criteria, the suggested X-TLRABiLSTM model performs better than several cutting-edge methods in the prognosis of Ischemic Heart Disease (IHD). Although they have had some success, traditional models like Random Forest, SVM, and Logistic Regression frequently fail to capture the nonlinear and temporal dynamics of patient data. Although they offer better predictive potential, deep learning models such as hybrid attention-based BiLSTM and regular LSTM lack interpretability and fairness control. A Residual Attention-Enhanced LSTM model was presented in recent studies by Cenitta et al. ¹⁶ and demonstrated an outstanding accuracy of 97.7% and an AUC of 98.6%. Further improving accuracy to 98.2%, F1-score to 98.1%, and AUC to 99.1%, our suggested X-TLRABiLSTM model, which combines transfer learning, SHAP-based explainability, and demographic fairness reweighting, also offers transparency and bias mitigation. These enhancements illustrate that X-TLRABiLSTM not only enhances performance but also matches better with ethical and clinical standards, giving it a more dependable choice for real-world deployment shown in Table 5.

Our tests verify that on limited data, incorporating transfer-learned BiLSTM weights improves generalization and speeds up convergence. Superior discriminative power is obtained by the residual attention mechanism, which simultaneously highlights important temporal aspects and maintains contextual signals. Clinical transparency and black-box performance are connected by SHAP-based interpretability. Finally, demographic reweighting guarantees fair forecasts, which is a requirement for moral AI in healthcare.

Limitations & future work •

Evaluation on larger, multi-institutional cohorts (e.g., MIMIC-IV) is needed to further validate generalizability.

Overall, X-TLRABiLSTM provides a reliable, comprehensible, and equitable approach for IHD prognosis that is ready to be included into useful clinical decision-support tools.