A TRUST-AWARE FEDERATED LEARNING FRAMEWORK WITH CONTEXT-AWARE DYNAMIC GRADIENT PRESERVATION FOR EARLY CARDIOVASCULAR RISK PREDICTION

Abstract

Cardiovascular diseases (CVD) is a major cause of death worldwide and require early and privacy guaranteeing risk prediction models. Current federated learning models employ gradient pruning methods such as Adaptive Gradient Pruning Optimization (AGPO) which can carelessly remove clinically important but low-magnitude features. This results in degraded convergence and compromised model fairness. To overcome these challenges, this paper introduces FedCure-X, a new trust-aware federated learning system that incorporates Context-Aware Dynamic Gradient Preservation Optimization (CD-GPO). In contrast to traditional pruning techniques, CD-GPO uses medical context-informed filters, convergence-aware scheduling and fairness regularization for maintaining semantically meaningful gradients even in heterogeneous and non-IID healthcare datasets. Secure aggregation, patient privacy and client trust are guaranteed by the framework along with enhanced predictive accuracy and convergence rate. The proposed framework is tested against benchmark datasets Framingham and MIMIC-III. FedCure-X shows better performance than baseline models with an accuracy rate of 96.8% and enhanced robustness at distributed clinical nodes. The proposed framework is ethical, scalable and smart in early detection systems for CVD risk prediction.