Explainable and High-Performance ECG-Informed Machine Learning and Deep Learning Framework for Cardiovascular Risk Prediction

Predictive models used for the assessment of cardiovascular disease (CVD) risk must not only be highly accurate but also explainable and reliable for clinical decision-making. In this paper, we present a methodological framework for explainable machine learning and deep learning approaches to CVD risk prediction that uses clinically derived electrocardiogram (ECG) features and conventional CVD risk factors. A large synthetic dataset containing 30,000 patient cases was created according to a clinically validated distribution to provide reproducible and privacy-preserving benchmarking. A series of interpretability tests were performed on multiple models of machine learning and deep learning categories within a unified experimental setup. Out of the tested models, XGBoost and BiLSTM provided the highest discrimination capabilities in the machine learning and deep learning categories, respectively. Explaining the predictions via the SHapley Additive explanations (SHAP) approach, calibration of probability estimations, uncertainty quantification, and ablation studies were applied to assess the explainability, reliability, and robustness of the models. We show that the presented framework allows for the production of consistent and interpretable predictions along with the evaluation of the models’ explainability and predictive reliability. Instead of focusing on the single-feature importance, this paper presents a reproducible methodological framework for the explainable modeling of CVD risk.