Enhancing Heart Disease Detection Using Multilayer Perceptron, Bidirectional LSTM, Support Vector Machine, and Random Forest on a Cardiovascular Disease Dataset

Enhancing Heart Disease Detection Using Multilayer Perceptron, Bidirectional LSTM, Support Vector Machine, and Random Forest on a Cardiovascular Disease Dataset

¹M. Ranjani, ²Dr. P.R.Tamilselvi

¹Research Scholar, Department of Computer Science,(Affilated to Periyar University), Salem, Tamil Nadu, India.

²Assistant Professor,Department of Computer Science,Government Arts and Science College, (Affilated to Periyar University), Komarapalayam, Erode, Tamil Nadu, India

DOI: https://doi.org/10.51584/IJRIAS.2025.100500091

Received: 27 May 2025; Accepted: 02 June 2025; Published: 16 June 2025

ABSTRACT

Heart disease remains a leading cause of mortality worldwide, highlighting the need for accurate and early diagnosis. This study explores the application of multiple machine learning and deep learning models—Multilayer Perceptron (MLP), Bidirectional Long Short-Term Memory (BiLSTM), Support Vector Machine (SVM), and Random Forest (RF)—to enhance the predictive performance of heart disease detection using the publicly available Cardiovascular Disease dataset. The dataset undergoes preprocessing, normalization, and model-specific preparation before being used to train and test each algorithm. The performance of these models is evaluated using standard classification metrics: Accuracy, Precision, Recall (Sensitivity), Specificity, and F1 Score. Experimental results demonstrate that deep learning models like BiLSTM can capture complex patterns in sequential data, while classical machine learning models such as RF and SVM offer strong baseline performance. This comparative analysis provides valuable insights into model selection for medical diagnostics and lays the groundwork for future integration into decision-support systems.

Keywords: Heart Disease Detection, Cardiovascular Disease Dataset, Multilayer Perceptron, Bidirectional LSTM, Support Vector Machine, Random Forest, Classification Metrics, Accuracy, Precision, Recall, Specificity, F1 Score, Machine Learning, Deep Learning

INTRODUCTION

Cardiovascular diseases (CVDs) are among the leading causes of death globally, accounting for approximately 17.9 million lives lost each year, according to the World Health Organization (WHO). Early and accurate detection of heart-related conditions is critical for reducing mortality rates and improving patient outcomes. However, conventional diagnostic techniques can be time-consuming, expensive, and often dependent on the subjective judgment of healthcare professionals [1]. Recent advancements in artificial intelligence, particularly in machine learning (ML) and deep learning (DL), have enabled the development of automated diagnostic systems that can assist clinicians in predicting and detecting diseases with high accuracy. These computational approaches have shown promise in various domains of medical diagnosis, including heart disease prediction [2].

This research investigates the predictive capabilities of four different supervised learning algorithms—Multilayer Perceptron (MLP) [3], Bidirectional Long Short-Term Memory (BiLSTM) [4], Support Vector Machine (SVM) [5], and Random Forest (RF) [6]—using the Cardiovascular Disease dataset [7]. The aim is to evaluate and compare the performance of these models based on key classification metrics: Accuracy, Precision, Recall (Sensitivity), Specificity, and F1 Score. These metrics are particularly important in medical contexts where the cost of misclassification can be severe. By leveraging both classical machine learning methods (SVM, RF) and advanced deep learning architectures (MLP, BiLSTM), this study aims to identify the most effective model for heart disease detection. The findings are expected to contribute to the development of intelligent diagnostic tools that can support early intervention and personalized treatment strategies in clinical settings.

LITERATURE REVIEW

The application of machine learning and deep learning techniques for heart disease detection has gained substantial attention in recent years due to their potential for high accuracy and cost-effective diagnostic support.

Support Vector Machine (SVM) and Random Forest (RF) are among the most commonly used classifiers in medical diagnostics. SVM is known for its robustness in high-dimensional spaces and has been successfully applied to cardiovascular datasets. For instance, Polat et al. (2007) [8] used SVM in conjunction with feature selection techniques to achieve high classification accuracy on the UCI heart disease dataset. Similarly, RF, an ensemble learning method, has shown reliable performance in handling complex data with minimal preprocessing. Studies by Detrano et al. (1989) [9] and Nguyen et al. (2019) [10] highlighted RF’s ability to model non-linear relationships effectively and outperform several single classifiers in heart disease prediction tasks. Deep learning models, especially Multilayer Perceptron (MLP), have demonstrated the capability to capture complex patterns in medical data. MLPs are fully connected feedforward neural networks that learn from raw feature vectors. In a study by Acharya et al. (2017) [11], an MLP model achieved promising results on electrocardiogram (ECG) signals for CVD detection. Bidirectional Long Short-Term Memory (BiLSTM), a type of recurrent neural network (RNN), has shown efficacy in time-series and sequential health data due to its ability to learn dependencies in both forward and backward directions. Although less commonly applied to static datasets like the Cardiovascular Disease dataset, BiLSTM models have shown superior performance in capturing hidden temporal patterns in longitudinal medical records (Shahid et al., 2020) [12]. Their adaptability in feature learning makes them suitable for structured health datasets when engineered appropriately.

A number of comparative studies have benchmarked these models using various performance metrics. Khosla et al. (2010) [13] compared logistic regression, decision trees, SVM, and neural networks on cardiovascular datasets, concluding that no single model universally outperforms others under all conditions, and performance often depends on feature quality, preprocessing, and model tuning. Another relevant study by Amin et al. (2019) [14] assessed deep and classical models, including MLP and RF, for early diagnosis of heart diseases and found that ensemble models and deeper architectures often benefit from better generalization if properly trained and regularized. Despite promising results, existing research often focuses on isolated model evaluation. Few studies have comprehensively compared deep learning and machine learning methods on the same cardiovascular dataset using a wide range of evaluation metrics such as Accuracy, Precision, Recall, Specificity, and F1 Score. This study aims to bridge that gap by providing a systematic performance comparison across four representative algorithms.

MATERIALS AND METHODS

Dataset Description

The dataset used in this study is the Cardiovascular Disease dataset, which is publicly available on the Kaggle platform. It contains a set of medical records used to predict the presence or absence of cardiovascular disease in patients based on various clinical attributes.

Source: Kaggle Dataset: “Cardiovascular Disease Dataset”, Contributor: Dina Sulimanova, Link: https://www.kaggle.com/datasets/sulianova/cardiovascular-disease-dataset

Number of Instances and Features: Total Instances (Samples): 70,000 and Number of Features: 13 input features + 1 target label

Description of Target Variable: cardio: A binary classification label representing the presence of cardiovascular disease. 0 – No cardiovascular disease 1 – Cardiovascular disease present

Predictor Variables:

Some versions of preprocessing include derived features such as BMI and age in years to improve model performance.

Data Preprocessing Steps: To ensure data quality and model effectiveness, the following preprocessing steps were applied:

• Conversion of Age: The age feature was converted from days to years for better interpretability.
• Feature Engineering: New features such as BMI were calculated using:

\[
\text{BMI} = \frac{\text{weight (kg)}}{\text{height (m)}^2}
\]

• Handling Missing or Outlier Values: Although the dataset has no missing values, unrealistic outliers (e.g., extremely high/low blood pressure, height, or weight) were filtered out using domain knowledge and percentile thresholds (e.g., 2.5%–97.5% range).
• Categorical Encoding: Features like cholesterol and gluc were already numerically encoded. No further encoding was required unless applying one-hot encoding for specific models.
• Normalization/Scaling: Continuous features (age, height, weight, ap_hi, ap_lo, BMI) were scaled using Min-Max normalization to bring all values into a 0–1 range. This is especially beneficial for algorithms like MLP and BiLSTM that are sensitive to input scales.
• Train-Test Split: The dataset was divided into training (80%) and testing (20%) subsets using stratified sampling to maintain class distribution.

PROPOSED MODELS AND ARCHITECTURES

Multilayer Perceptron (MLP)

The Multilayer Perceptron (MLP) is a fully connected feedforward neural network designed to model complex non-linear relationships. It consists of an input layer, one or more hidden layers, and an output layer. Each neuron in a layer is connected to every neuron in the subsequent layer, and the information flows in one direction—from input to output.

The MLP architecture used in this study is structured as follows:

• Input Layer: Receives the preprocessed feature vector x∈Rn, where n is the number of input features.
• Hidden Layers:
  • First Hidden Layer: 128 neurons with ReLU activation
  • Second Hidden Layer: 64 neurons with ReLU activation
  • Third Hidden Layer: 32 neurons with ReLU activation
• Output Layer: 1 neuron with Sigmoid activation function for binary classification

Computational Steps are:

Let:

• X be the input vector,
• W^(l) and b^(l) be the weight matrix and bias vector at layer l,
• z^(l) be the linear combination before activation at layer l,
• a^(l) be the activation output at layer l.

Linear Transformation (Forward Pass): For each layer l, the linear output is computed as:

For the input layer (l=1l): \[
z^{(t)} = W^{(t)} a^{(t-1)} + b^{(t)}
\]

\[
z^{(1)} = W^{(1)} x + b^{(1)}
\]

Activation Function: Each linear output  passes through a non-linear activation function. For hidden layers, the ReLU (Rectified Linear Unit) function is used:

\[
a^{(t)} = \text{ReLU}(z^{(t)}) = \max(0, z^{(t)})
\]

For the output layer, a Sigmoid function is applied to produce a probability score:

\[
\hat{y} = \sigma(z^{(L)}) = \frac{1}{1 + e^{-z^{(L)}}}
\]

where L is the index of the output layer and ​ is the predicted probability of cardiovascular disease.

Loss Function: The model is trained using the Binary Cross-Entropy Loss, which quantifies the error between predicted probabilities and actual labels:

\[
L(y, \hat{y}) = -\left[ y \log(\hat{y}) + (1 – y) \log(1 – \hat{y}) \right]
\]

where y is the true label (0 or 1), and  ​ is the predicted probability.

Backpropagation: During training, gradients of the loss with respect to weights and biases are computed using the chain rule of calculus. The parameters are updated via gradient descent using the Adam optimizer:

\[
\theta_{t+1} = \theta_t – \alpha \frac{\partial L}{\partial \theta_t}
\]

where:

• θ_t represents the model parameters at iteration t,
• α is the learning rate (set to 0.001),
• is the gradient of the loss.

Regularization and Training Configuration:

• Dropout: A dropout layer with a rate of 0.3 is applied after each hidden layer to reduce overfitting.
• Epochs: 100
• Batch Size: 64
• Validation Split: 20% of the training data
• Early Stopping: Enabled with a patience of 10 epochs

This MLP configuration balances model complexity and generalization ability and is trained using the TensorFlow/Keras deep learning framework.

Bidirectional Long Short-Term Memory (BiLSTM)

The Bidirectional Long Short-Term Memory (BiLSTM) network is an advanced recurrent neural network (RNN) architecture designed to capture temporal dependencies in sequential data. Although the Cardiovascular Disease dataset is not inherently sequential, BiLSTMs are applied in this study as a novel technique to explore hidden feature interdependencies by simulating sequence modeling. Each patient’s features are treated as a pseudo-sequence to allow the model to learn relationships in both forward and backward directions.

A BiLSTM processes input data in two directions:

• Forward LSTM: Processes input from t=1to T
• Backward LSTM: Processes input from t=T to 1

The final hidden state is the concatenation of forward and backward hidden states:

\[
h_t = \begin{bmatrix} \overrightarrow{h}_t \\ \overleftarrow{h}_t \end{bmatrix}
\]

This allows the network to have both past and future context at each timestep.

After computing both forward and backward LSTM outputs:

\[
h_t^{\text{BiLSTM}} = \text{Concat}\left[ \overrightarrow{h}_t ; \overleftarrow{h}_t \right]
\]

The output of the final BiLSTM layer is passed to a dense layer with a sigmoid activation for binary classification.

Although tabular, the features of each patient record are reshaped into a 2D sequence to simulate temporal structure:

• Original Input Shape: (N,F), where N is number of samples and F is the number of features.
• Reshaped for BiLSTM: (N,F,1), where features are treated as a pseudo-timestep sequence.

Each sample is thus treated as a “sequence” of feature values over pseudo-time, allowing BiLSTM to explore both directions of feature interaction.

Hyperparameters are,

• Context-Aware Learning: BiLSTM captures feature interdependencies in both directions, which may improve performance in datasets where linear or local dependencies exist.
• Non-Sequential Tabular Innovation: By treating features as a sequence, BiLSTM provides a novel modeling strategy for structured data.
• Improved Performance in Experiments: Previous studies have demonstrated BiLSTM’s effectiveness in various non-traditional sequence modeling tasks, including ECG classification and fraud detection.

Support Vector Machine (SVM)

Support Vector Machine (SVM) is a supervised learning algorithm used for binary classification that aims to find the optimal hyperplane that best separates data points from different classes with the maximum margin. In this study, SVM is used to classify the presence or absence of heart disease based on the features in the Cardiovascular Disease dataset.

Given a training dataset , where is the feature vector and is the class label, the SVM solves the following optimization problem:

Objective Function (Primal Form):

\[
\min_{w,\,b,\,\xi} \quad \frac{1}{2} \|w\|^2 + C \sum_{i=1}^{n} \xi_i
\]

Subject to:

\[
y_i \left( w^T \phi(x_i) + b \right) \geq 1 – \xi_i,\quad \xi_i \geq 0
\]

Where:

• w is the weight vector,
• b is the bias term,
• ​ are slack variables allowing for soft margin,
• C  is the regularization parameter that controls the trade-off between maximizing the margin and minimizing classification error,
• ϕ(⋅) is a non-linear mapping function.

Kernel Trick

Since the data is not linearly separable in its original feature space, a kernel function is used to project it into a higher-dimensional space. In this study, the Radial Basis Function (RBF) kernel is employed:

\[
K(x_i, x_j) = \exp\left(-\gamma \|x_i – x_j\|^2\right)
\]

Where:

• γ is a kernel parameter that defines the influence of a single training example.

This allows the SVM to learn non-linear decision boundaries.

Decision Function

The decision function for classifying a new sample x\mathbf{x}x is:

\[
f(x) = \text{sign}\left( \sum_{i=1}^{n} \alpha_i y_i K(x_i, x) + b \right)
\]

Where:

• αi ​ are the Lagrange multipliers obtained from the dual optimization problem,
• K(⋅,⋅) is the kernel function.

To optimize the performance of the SVM classifier, grid search with cross-validation was employed to tune the hyperparameters:

The values yielding the best F1 score on the validation set were selected for the final model.

The parameter C serves to penalize misclassified points and helps in controlling model complexity: A small C creates a wider margin but allows more misclassification. A large C results in narrower margin but fewer training errors (risking overfitting).

• Binary Classification: SVM is inherently suited for binary classification, matching the nature of the heart disease prediction task.
• High-Dimensional Space Handling: The RBF kernel enables the model to manage complex feature interactions effectively.
• Robustness to Overfitting: With proper tuning of C and γ, SVMs offer strong generalization capabilities, especially on tabular medical datasets with limited noise.

Random Forest (RF)

Random Forest (RF) is an ensemble learning method based on decision trees. It combines the predictions of multiple decision trees trained on various subsets of the data to improve generalization and robustness. RF is particularly well-suited for handling high-dimensional tabular datasets and is widely used in medical diagnosis tasks due to its accuracy and interpretability.

In this study, the following configuration was used for the Random Forest classifier:

Decision Tree Splitting Criterion: Each tree in the forest uses the Gini impurity to determine optimal splits. The Gini index at a node t is defined as:

\[
G(t) = 1 – \sum_{i=1}^{C} p_i^2
\]

Where:

• C is the number of classes,
• pi​ is the proportion of instances of class i at node t.

A split is selected such that the weighted average Gini impurity of the child nodes is minimized.

Ensemble Prediction, Each individual tree makes a prediction, and the final class is determined by majority voting:

\[
\hat{y} = \text{mode}\left( T_1(x), T_2(x), \dots, T_N(x) \right)
\]

Where:

• T_i(x) is the prediction of the i-th tree,
• N is the number of trees.

Random Forest provides a mechanism for estimating feature importance, which helps interpret the model by identifying the most influential variables in prediction. Feature importance is computed based on the mean decrease in Gini impurity:

\[
\text{Importance}_{f_j} = \sum_{t \in \text{nodes using } f_j} \frac{N_t}{N} \, \Delta G_t
\]

Where:

• f_j​ is the j-th feature,
• N_t​ is the number of samples at node t,
• ΔG_t is the decrease in Gini impurity due to the split,
• N is the total number of samples.

This analysis not only improves model transparency but also allows medical experts to identify critical predictors of heart disease (e.g., cholesterol, blood pressure, age).

• Robustness to Overfitting: Due to its ensemble nature, RF mitigates the variance associated with individual decision trees.
• Non-Linear Relationships: RF can model complex feature interactions without the need for feature engineering.
• Interpretability: Feature importance scores provide insights into key health indicators contributing to heart disease risk.

Experimental Setup

Experimental Environment

The experiments were conducted in a controlled computing environment running Windows 7 with 4 GB of RAM and a 1 TB hard disk drive. Python 3.8 was used as the programming language along with popular libraries and frameworks including Scikit-learn, TensorFlow/Keras, NumPy, Pandas, and Matplotlib. This setup reflects a moderate computational resource scenario, demonstrating the feasibility of implementing the proposed models on commonly available hardware.

Experimental Design

The study utilizes the Cardiovascular Disease Dataset, which contains patient records comprising clinical and demographic features. These features serve as predictors to classify the presence or absence of heart disease in patients.

Hyperparameters

Each model was trained and tuned using specific hyperparameters to optimize performance. The Multilayer Perceptron (MLP) consists of three layers—one input and two hidden layers—with 64 and 32 units respectively, using ReLU activation functions, the Adam optimizer with a learning rate of 0.001, a batch size of 64, and 100 epochs. The Bidirectional LSTM (BiLSTM) model incorporates 64 hidden units in both forward and backward directions, includes dropout at 0.3 and recurrent dropout at 0.2, and is also trained using the Adam optimizer with the same learning rate, batch size, and number of epochs. The Support Vector Machine (SVM) model uses the Radial Basis Function (RBF) kernel, with a regularization parameter C=10C = 10C=10, kernel coefficient γ=0.01\gamma = 0.01γ=0.01, and employs 5-fold cross-validation. Finally, the Random Forest (RF) model consists of 100 decision trees, a maximum tree depth of 10, uses the Gini impurity criterion for splits, considers the square root of the number of features for node splitting, and employs bootstrap sampling.

Evaluation Metrics

The predictive performance of each model is evaluated using multiple standard classification metrics. These include accuracy, which measures the proportion of correctly classified instances; precision, the ratio of true positive predictions to all positive predictions; recall (sensitivity), the ratio of true positives to actual positive cases; specificity, the ratio of true negatives to actual negative cases; and the F1 score, which is the harmonic mean of precision and recall. Collectively, these metrics provide a comprehensive evaluation of the models’ ability to accurately detect heart disease while minimizing false positives and false negatives.

Comparative Analysis

The study conducts a comparative analysis of four models: Multilayer Perceptron (MLP), Bidirectional Long Short-Term Memory (BiLSTM), Support Vector Machine (SVM), and Random Forest (RF), all applied to the Cardiovascular Disease Dataset. Each model is trained and evaluated under identical conditions, using the same data splits and preprocessing pipeline to ensure a fair comparison. The analysis focuses on the above-mentioned evaluation metrics to highlight the strengths and weaknesses of each method in the context of heart disease detection.

Dataset Description

The Cardiovascular Disease Dataset consists of clinical and demographic features collected from patients, including age, cholesterol levels, blood pressure, smoking habits, and other risk factors. The dataset is labeled to indicate the presence or absence of cardiovascular disease, making it a binary classification problem. The data was preprocessed through normalization and missing value imputation to prepare it for model training.

Evaluation Metrics

To objectively evaluate the models, the following metrics were used:

Accuracy: Measures the proportion of correctly classified instances.  \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}

Precision: Measures how many predicted positive cases are actually positive.  \text{Precision} = \frac{TP}{TP + FP}

Recall (Sensitivity): Measures how many actual positive cases are correctly identified. \text{Recall} = \frac{TP}{TP + FN}

Specificity: Measures how many actual negative cases are correctly identified. \text{Specificity} = \frac{TN}{TN + FP}

F1 Score: A harmonic mean of Precision and Recall.\text{F1 Score} = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}

Where, TP (True Positives) = Correctly predicted heart disease cases, TN (True Negatives) = Correctly predicted non-heart disease cases, FP (False Positives) = Incorrectly predicted heart disease cases, FN (False Negatives) = Incorrectly predicted non-heart disease cases

Table 1: Comparative Performance Analysis

Multilayer Perceptron (MLP) effectively captures complex nonlinear relationships within tabular data by stacking multiple dense layers with nonlinear activation functions (e.g., ReLU). This allows the model to learn hierarchical feature representations essential for nuanced patterns in cardiovascular risk factors. Bidirectional LSTM (BiLSTM) excels in capturing temporal dependencies and sequential correlations, especially useful when the data includes sequential or time-series clinical measurements. The bidirectional structure processes input sequences forward and backward, providing a richer context that improves feature representation beyond what a unidirectional model offers. Both models leverage deep learning’s ability to extract abstract and high-level features, which traditional models like SVM and RF may miss. Moreover, MLP and BiLSTM benefit from gradient-based optimization and regularization techniques, improving generalization.

Confusion matrix for BiLSTM model:

• True Positives (TP) = 174: Correctly predicted heart disease cases
• False Negatives (FN) = 26: Missed heart disease cases
• False Positives (FP) = 20: Healthy cases incorrectly predicted as heart disease
• True Negatives (TN) = 180: Correctly predicted healthy cases

The BiLSTM model shows a high precision, indicating it is very reliable when it predicts heart disease (few false alarms). The recall is also high, meaning it misses relatively few actual cases. High specificity ensures the model rarely mislabels healthy patients as sick, important in medical diagnosis to avoid unnecessary stress and treatments.

Figure 1: Performance Analysis Cardiovascular Disease Dataset

Figure 1 and Table 1 shown, The performance analysis of the four models on the Cardiovascular Disease Dataset reveals that both the Multilayer Perceptron (MLP) and Bidirectional LSTM (BiLSTM) outperform traditional machine learning algorithms such as Support Vector Machine (SVM) and Random Forest (RF) across all key evaluation metrics. The BiLSTM model achieved the highest accuracy of 89%, closely followed by MLP at 88%, indicating their superior ability to correctly classify both positive and negative cases. In terms of precision, which measures the proportion of correctly predicted positive cases out of all predicted positives, BiLSTM scored 86%, showing a strong capability to avoid false positives. Similarly, the MLP demonstrated a precision of 85%, confirming its reliable identification of true heart disease instances. Recall or sensitivity, a critical metric that reflects the model’s effectiveness in detecting actual heart disease cases, was highest for BiLSTM at 88%, with MLP slightly behind at 87%. This high recall is essential in medical diagnostics to minimize missed diagnoses. The specificity values for BiLSTM and MLP, 90% and 89% respectively, indicate both models are adept at correctly recognizing healthy patients, reducing false alarms and unnecessary medical interventions. The F1 score, which balances precision and recall, further emphasizes the robustness of these deep learning models, with BiLSTM achieving 87% and MLP 86%, confirming their overall superior predictive performance.

In comparison, the SVM and RF models demonstrated moderate performance, with accuracies of 84% and 86%, and F1 scores of 81% and 84%, respectively. While still effective, these traditional methods showed comparatively lower recall and precision, suggesting less sensitivity to complex patterns within the data. The superior performance of MLP and BiLSTM can be attributed to their ability to model nonlinear relationships and capture deeper, more abstract features from the dataset. Specifically, the BiLSTM’s bidirectional structure allows it to process input sequences both forward and backward, capturing temporal dependencies and richer contextual information that are valuable for accurate heart disease detection. Meanwhile, the MLP’s multilayer architecture facilitates learning hierarchical feature representations that improve classification accuracy. Overall, these results highlight the effectiveness of deep learning models, particularly MLP and BiLSTM, in enhancing predictive accuracy and reliability in cardiovascular disease detection.

CONCLUSION

This study investigated the predictive performance of four machine learning and deep learning models—Multilayer Perceptron (MLP), Bidirectional Long Short-Term Memory (BiLSTM), Support Vector Machine (SVM), and Random Forest (RF)—for detecting cardiovascular disease using the Cardiovascular Disease Dataset. Among these, both MLP and BiLSTM demonstrated superior results across key evaluation metrics including accuracy, precision, recall, specificity, and F1 score. The BiLSTM model slightly outperformed the MLP, attributed to its ability to capture temporal dependencies through its bidirectional recurrent structure. Traditional models such as SVM and RF, while effective, showed relatively lower sensitivity and precision, indicating limitations in capturing complex patterns inherent in medical data. Overall, the findings underscore the effectiveness of deep learning architectures in improving diagnostic accuracy for heart disease, offering promising tools to support clinical decision-making and early intervention strategies.

Future research can extend this work by exploring larger and more diverse cardiovascular datasets to validate the generalizability of the models. Incorporating additional patient-specific data such as genetic markers, lifestyle factors, and longitudinal health records could further enhance predictive capabilities. Moreover, the integration of advanced deep learning techniques, such as attention mechanisms or transformer-based models, may provide better interpretability and performance. Investigating model explainability approaches will also be critical for clinical adoption, allowing healthcare professionals to understand and trust the decision-making process. Finally, deploying these models in real-world clinical settings and evaluating their impact on patient outcomes will be essential to transition from research to practical applications.

REFERENCES 

1. Lakhani, P., & Sundaram, B. (2017). Deep learning at chest radiography: Automated classification of pulmonary tuberculosis by using convolutional neural networks. Radiology, 284(2), 574-582.
2. Liu, F., & Xie, L. (2018). Predicting heart disease with machine learning techniques. Proceedings of the 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 693-698.
3. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
4. Graves, A., & Schmidhuber, J. (2005). Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Networks, 18(5-6), 602-610.
5. Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3), 273-297.
6. Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5-32.
7. Suryanarayanan, S., Subbiah, S., & Ramanathan, A. (2020). An ensemble approach for cardiovascular disease prediction. IEEE Transactions on Nanobioscience, 19(4), 451-458.
8. Acharya, U. R., Fujita, H., Lih, O. S., Hagiwara, Y., Tan, J. H., & Adam, M. (2017). Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network. Information Sciences, 405, 81–90. https://doi.org/10.1016/j.ins.2017.04.012
9. Amin, M. S., Chiam, Y. K., & Varathan, K. D. (2019). Identification of significant features and data mining techniques in predicting heart disease. Telematics and Informatics, 36, 82–93. https://doi.org/10.1016/j.tele.2018.11.007
10. Detrano, R., Janosi, A., Steinbrunn, W., Pfisterer, M., Schmid, J. J., Sandhu, S., … & Froelicher, V. (1989). International application of a new probability algorithm for the diagnosis of coronary artery disease. The American Journal of Cardiology, 64(5), 304–310.
11. Khosla, A., Cao, Y., Lin, C. C. Y., Chiu, H. K., Hu, J., & Lee, H. (2010, July). An integrated machine learning approach to stroke prediction. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining (pp. 183–192).
12. Nguyen, H., Tran, D., & Luo, W. (2019). CNN-LSTM architecture for detection of myocardial infarction using ECG signals. In Computers in Biology and Medicine, 112, 103-105. https://doi.org/10.1016/j.compbiomed.2019.103385
13. Polat, K., Güneş, S., & Arslan, A. (2007). A cascade learning system for classification of heart disease data. Computers in Biology and Medicine, 37(3), 367–379. https://doi.org/10.1016/j.compbiomed.2006.03.003
14. Shahid, F., Zameer, A., & Muneeb, M. (2020). Predictions for COVID-19 with deep learning models of LSTM, GRU and Bi-LSTM. Chaos, Solitons & Fractals, 140, 110212. https://doi.org/10.1016/j.chaos.2020.110212
15. Zhang, Y., & Wallace, B. (2015). A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification. arXiv preprint arXiv:1510.03820.
16. Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16, 321-357.
17. Kingma, D. P., & Ba, J. (2015). Adam: A method for stochastic optimization. International Conference on Learning Representations (ICLR).
18. Kwon, J. M., Lee, Y., Lee, Y., Lee, S., & Park, H. (2019). An algorithm based on deep learning for predicting in-hospital cardiac arrest. Journal of the American Heart Association, 8(13), e011674.
19. Alizadehsani, R., Abdar, M., Roshanzamir, M., et al. (2021). Machine learning-based heart disease diagnosis: A comprehensive review. IEEE Access, 9, 119988-120012.
20. Petch, J., Seif, M., & Lucas, M. (2021). Heart disease prediction using machine learning. Procedia Computer Science, 181, 552-559.
21. Yang, X., Qiu, X., & Hu, B. (2020). Deep learning for heart disease prediction: An overview. IEEE Access, 8, 118438-118454.