Key result
Machine learning models detect heart disease with up to ~100% accuracy but lack external validation.
While machine and deep learning models report exceptionally high benchmark accuracies (up to 99.7%) for heart disease detection, their clinical readiness is severely limited by a reliance on small, single-center datasets and a lack of external validation.
High benchmark accuracies do not support clinical use; leaves open whether AI models can improve heart disease detection after external validation.
Cardiovascular diseases (CVDs) remain the foremost cause of mortality worldwide, claiming approximately 17.9 million lives annually and representing 32% of all global deaths. Timely and accurate detection of heart disease is therefore of paramount importance. This paper presents a narrative survey, synthesising 57 peer-reviewed studies (2000–2025, concentrated in 2020–2025) of computational methods for heart disease detection, encompassing classical machine learning, deep learning, hybrid architectures, explainable artificial intelligence, and emerging paradigms such as federated learning and IoT-integrated systems. We reviewed studies covering diverse data modalities (clinical tabular data, ECG signals, echocardiograms, cardiac MRI, chest X-rays, and wearable sensor streams), benchmark datasets, feature engineering and selection strategies, performance evaluation metrics, and clinical deployment considerations. The survey finds that ensemble methods such as XGBoost and Random Forest report accuracies ranging from the low 80 s to as high as 98% on structured clinical datasets, and that convolutional neural network (CNN), long short-term memory (LSTM), and transformer-based models report comparably wide ranges (roughly 86–99.7%, by AUC and accuracy respectively) on ECG and imaging benchmarks. Critically, we find that the studies reporting the highest figures are almost exclusively evaluated on small, single-centre benchmark datasets without external or prospective validation, so these numbers should be read as a ceiling on benchmark performance rather than as evidence of clinical-grade accuracy; several of the reviewed studies show an explicit training-to-test accuracy gap consistent with overfitting. We further discuss explainability techniques such as SHAP, LIME, and Grad-CAM that are considered important for clinical trust, and identify open challenges including class imbalance, data scarcity, limited external validation, reproducibility, multi-modal fusion, model generalisability, and fairness.
No takes yet. Share an insight, caveat, or question.
Md. Siam Ansary (2026) conducted a review in Heart disease (n=57). Machine learning and deep learning models was evaluated on Model accuracy and AUC for heart disease detection. Machine learning and deep learning models achieved high reported accuracies of 80% to 99.7% on benchmark datasets for heart disease detection, though these results are limited by a lack of external validation.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: