Adaptive Federated Learning with Hierarchical Optimization and Dynamic Data Representation for Enhanced Healthcare Analytics

Privacy and data variability concerns present significant challenges in developing predictive healthcare models, mainly when exchanging sensitive medical information among various healthcare centers. This study addresses these issues by proposing a Hierarchical Adaptive Federated Learning System (HAFL), a two-tier aggregation framework designed to facilitate secure, scalable, and efficient model training across distributed medical data. The system integrates federated learning, neural networks, and ensemble techniques by employing local aggregation at hospitals and global aggregation in the cloud. This approach enhances privacy while reducing communication overhead. The proposed Federated Enhanced Neural Classifier (FENC) algorithm demonstrates impressive performance, achieving an accuracy rate of 91.14% in predicting heart disease and 87.35% for diabetes. In contrast, the Federated Random Forest (FRF) model shows superior performance, achieving an accuracy of 92.26% for heart disease data and 92.02% for diabetes data sets. These results illustrate the effectiveness of the proposed system in enabling privacy-preserving, high-performance healthcare analytics, addressing challenges related to heterogeneous data and model scalability.