Distribution-Aware Federated Learning for Diabetes Prediction Using Tabular Clinical Data Under Non-IID and Class-Imbalanced Settings

Federated learning (FL) enables collaborative clinical model training without centralized data sharing, yet its deployment is hindered by statistical heterogeneity (non-IID data) and inherent class imbalance across healthcare institutions. Conventional aggregation strategies such as FedAvg and FedProx weight client updates solely by dataset size, ignoring class distributions and thereby biasing the global model toward the majority class. To address this, we propose Distribution-Aware Federated Learning (DA-FL), which introduces a minority-class amplification factor{phi} k computed as the ratio of a clients local positive class rate to the global positive class rate. Combined with class-weighted cross-entropy loss at the client level, DA-FL forms a two-level correction mechanism that mitigates imbalance without additional data sharing. Experiments on the CDC BRFSS 2021 diabetes dataset (236,378 records across five simulated clients under three non-IID levels) show that DA-FL improves F1-Macro by 18.2% and G-Mean by 26.7% over FedAvg under moderate non-IID conditions, while achieving 31-fold greater F1-Macro stability across 30 communication rounds. These findings demonstrate that DA-FL is an effective and practically deployable solution for federated clinical prediction under realistic non-IID and class-imbalanced settings.