A Systematic Review of Machine Learning and Deep Learning Approaches for Heart Disease Prediction

Abstract

Machine learning (ML) and deep learning (DL) approaches have shown increasing promise for heart disease prediction, but comprehensive evidence regarding their effectiveness compared to traditional clinical methods, implementation challenges, and real-world deployment readiness remains fragmented. Healthcare systems require systematic evaluation of these approaches to make informed decisions about clinical adoption, yet current literature lacks comprehensive synthesis of performance outcomes, generalizability challenges, and practical implementation considerations. This systematic review evaluates studies published between 2020 and 2025 to: (1) evaluate ML and DL approaches applied to heart disease prediction and assess their effectiveness compared to traditional clinical methods; (2) examine commonly used datasets and evaluation metrics; (3) compare algorithm performance and generalization capabilities between ML and DL approaches; (4) investigate how interpretability and explainability influence model selection and clinical adoption; and (5) identify challenges and gaps remaining for real-world deployment. A systematic search was conducted across multiple databases for peer-reviewed studies published between 2020–2025. Studies were included if they applied ML/DL algorithms to heart disease prediction using clinically relevant data, included comparative analysis with traditional methods, reported quantitative performance metrics, and provided sufficient methodological details. Data extraction focused on ML/DL approaches, performance results, validation methods, study limitations, and effectiveness findings using structured extraction forms. Thirty-one studies met inclusion criteria, representing diverse global applications. Traditional ML methods appeared in 19 studies (Random Forest n=10, XGBoost n=8, Logistic Regression n=8, overlap present). Deep learning appeared in 16 studies, with CNNs dominant in imaging-based prediction. ML/DL approaches demonstrated superior performance in 17 of 31 studies, while an additional 7 studies reported non-inferior or equivalent performance, indicating that most applications performed at least as well or better than clinical baselines. AUC improvements ranged from 0.01–0.21 across validated models. XGBoost excelled in structured tabular data, while deep learning dominated imaging-based tasks. However, only ~50% of studies performed external validation, with several reporting performance degradation on external datasets. Critical gaps included limited interpretability evaluation, persistent algorithmic bias, incomplete prospective validation, and minimal reporting of real-world clinical integration — reflecting challenges that must be addressed before widespread deployment is feasible.