ADAPTIVE DRIFT-AWARE FEDERATED LEARNING FRAMEWORK FOR HEART DISEASE PREDICTION

Abstract

Heart disease develops gradually through variations in physiological parameters such as heart rate and oxygen saturation, often remaining undetected until a cardiac event occurs. Most prediction models in current use depend on centralized data collection, which creates privacy risks and limits deployment across distributed healthcare institutions. Additionally, the statistical properties of the physiological data evolve over time due to changes in patients’ health conditions and behavioural patterns, resulting in concept drift that progressively degrades model performance. This paper proposes the Drift-Aware Federated HeartCare (DFHC) framework, a unified federated learning architecture that simultaneously addresses privacy preservation and adaptive concept drift handling for heart disease prediction. Within the proposed framework client devices train models on local data and share only parameter updates, ensuring raw data never leaves the source device. A dual-stage drift detection component monitors physiological signals and triggers adaptive model migration when concept drift is detected. The proposed framework integrates ADWIN-DDM-based drift detection with cosine-similarity-driven adaptive model migration. Experimental results demonstrate that the proposed DFHC framework achieves a classification accuracy of 94.3% and an F1-score of 0.92, outperforming standard FedAvg, FedProx, and Per-FedAvg, while converging in fewer communication rounds