Federated, Explainable Artificial Intelligence for Heart Disease Prediction: A Literature Synthesis and a Proposed SHAP-Based Framework
Abstract
Heart disease is still the world's biggest killer, responsible for an estimated 19.8 million deaths in 2022 up from the 17.9 million figures for 2019 that gets cited so often in older papers. Machine learning models have gotten quite good at predicting cardiovascular risk, but two problems keep them out of the clinic: nobody can easily explain why a model flagged a given patient, and hospitals cannot legally pool their data to train better models in the first place. This paper looks at how federated learning (FL) and explainable AI (XAI) are being combined to deal with both problems at once, drawing on a review of twenty-eight peer-reviewed studies published between 2022 and 2026 that apply FL, ensemble learning, or XAI methods such as SHAP and LIME to cardiovascular data. Building on the gaps that review turns up, we propose FedX-Heart, a framework pairing a federated ensemble of Random Forest, XGBoost, and LSTM models with layered privacy protection (differential privacy, secure aggregation, k-anonymity) and dual SHAP/LIME explanations. The studies reviewed here report that federated training can match centralized accuracy (92.3% versus 90.1%), that stacking ensembles beat individual base learners (91–98% accuracy), that a transfer-learning ensemble reaches 98.2% accuracy with SHAP interpretability built in, and that adding k-anonymity costs only a few points of accuracy (83.21%). A synthesis of feature-importance findings across six SHAP-based studies turns up solid agreement on some predictors and real disagreement on others, and lining FedX-Heart up against its three closest predecessors shows that none of them covers federation, accuracy, privacy, and generalizability all at once. Taken together, it shows that relationship mastery and explanatory AI can blend out and deliver a lot of accuracy, though open questions remain around how privacy noise affects causes, how badly non-IID medical laboratory data hurts things in practice, and whether those protections penalize small patient subgroups. Even the FedX-Heart is a proposed design and not something we have tried, yet all need empirical validation before it should be clinically agreed upon.