Explainable Hybrid Deep Learning with Evolutionary Optimization for ECG-based Cardiovascular Risk Prediction

Abstract

Cardiovascular disease remains one of the leading causes of mortality worldwide, making early and reliable diagnosis essential for effective clinical intervention and patient management. Electrocardiogram (ECG) signals provide a non-invasive and efficient approach for monitoring cardiac activity and detecting abnormal heart conditions. However, accurate cardiovascular risk prediction using ECG signals remains challenging due to complex waveform morphology, inter-patient variability, and temporal dependencies present in heartbeat sequences. Traditional ML methods often rely on fixed feature representations and may fail to capture complex nonlinear ECG characteristics effectively. Recent advances in deep learning have demonstrated significant potential in automatically learning discriminative representations directly from ECG signals for improved prediction performance. In this work, an explainable deep learning framework optimized via evolutionary algorithms for ECG-based cardiovascular risk prediction is presented. The proposed framework evaluates multiple ML and DL models, including Logistic Regression, SVM, KNN, Random Forest, Gradient Boosting, CNN, LSTM, and a hybrid CNN-LSTM architecture. To further enhance predictive performance, a Genetic Algorithm-based optimization strategy is incorporated for hyperparameter tuning of the proposed CNN-LSTM framework. In addition, SHAP-based explainability analysis is integrated to improve transparency and interpretability of prediction decisions. Experimental results demonstrate that the proposed GA-optimized CNN-LSTM model achieves superior cardiovascular risk prediction performance compared to baseline ML and DL approaches. The findings confirm that combining hybrid deep learning, evolutionary optimization, and explainable AI provides a reliable and effective framework for ECG-based cardiovascular risk prediction.