A Hybrid Model for Predicting Heart Disease Using DNN and ML Algorithms

Abstract

Cardiovascular diseases are among the most prominent health challenges facing medical systems worldwide, with high mortality rates associated with this disease recorded annually. This is largely due to the difficulty of early detection of these diseases, as diagnosis often relies on clinical findings that appear in advanced stages of the disease. In this context, the need to employ artificial intelligence (AI) techniques, particularly machine learning and deep learning, emerges to develop accurate predictive models that contribute to the early identification of patients at risk of heart disease. This study aims to develop a hybrid model that combines traditional machine learning algorithms—Support Vector Machines (SVMs), Random Forests, XGBoost, LightGBM, and Logistic Regression—with Deep Neural Networks (DNNs), which are used to extract advanced representative features from data. This study uses a reliable medical dataset obtained from Kaggle consisting of four databases: Cleveland, Hungary, Switzerland, and Long Beach. It contains 76 features, including the predicted feature, but all published experiments report using a subset of 14 of them. The target field indicates the presence of heart disease in the patient. It is an integer value where 0 = no disease and 1 = disease. This study introduces an innovative multi-stage methodology for heart disease prediction, leveraging feature augmentation to enhance performance on a moderately sized dataset of 1,025 patient records. The process begins with preprocessing clinical data, comprising 13 features (e.g., age, sex, chest pain type, resting blood pressure, serum cholesterol, fasting blood sugar, resting electrocardiographic results, maximum heart rate, exercise induced angina, oldpeak, slope, number of major vessels, thal ), followed by initial training of traditional machine learning models (e.g., SVM, Random Forest , XGBoost, LightGBM , Logistic Regression) using these features. A Deep Neural Network (DNN) then extracts 32 high-level features from its penultimate layer, capturing complex, nonlinear patterns not evident in the original data. These DNN features are combined with the 13 clinical features to form a 45-feature set, significantly enriching the input space. A set of comprehensive evaluation indicators was used, including accuracy, confusion matrix, precision, recall, F1 coefficient, and the area under the ROC-AUC curve, to provide a comprehensive evaluation of the models before and after feature combination. The Random Forest model achieved the highest performance among classification models on the original features, with an accuracy rate of 97.80%, a high recall rate of 99.31%, and a predictive accuracy of 96.64%. In a previous study, Smith et al. (2018) employed traditional machine learning algorithms The experimental results showed that the Support Vector Machine achieved an accuracy of 85.2%. The results also showed a significant improvement in the performance of all classification algorithms after combining the original features with those extracted by deep neural networks (DNNs). This combination resulted in increased classification accuracy across all key indicators. The SVM algorithm achieved the highest AUC value of 99.90%, demonstrating its high ability to accurately distinguish between classes. The Random Forest, XGBoost, and LightGBM algorithms also achieved identical results in overall accuracy (99.63%) and other indicators. The results showed that combining the DNN-extracted features with the original features led to a significant improvement denotes consistency across models, not the numerical accuracy alone. in the prediction accuracy of all machine learning algorithms used, reflecting the effectiveness of the hybrid approach in enhancing predictive performance, especially in light of the challenges associated with class imbalance and small dataset sizes. This study confirms that combining machine learning and deep learning techniques provides a promising path for developing intelligent diagnostic tools capable of supporting medical decision-making, reducing false alarm rates, and contributing to improving early treatment opportunities, especially in resource-constrained medical settings.