FEDERATED DEEP LEARNING FOR DIABETES PREDICTION IN INDIA: ENHANCING PRIVACY AND INTERPRETABILITY WITH DP-SGD AND LIME

Diabetes prevalence in South Asia is increasing, and automated risk prediction is necessary to enable early intervention. This work proposes FedHybNet, a privacy-preserving federated hybrid architecture that combines a TabTransformer backbone with per-client personalization, applies differentially private stochastic gradient descent (DP-SGD) for formal Differential Privacy (DP) guarantees, and uses Local Interpretable Model-agnostic Explanations (LIME) for instance-level interpretability. FedHybNet protects client updates via DP-SGD while preserving predictive utility through a hybrid aggregation scheme and lightweight on-device personalization. Performance and interpretability were evaluated on a recent public South-Asia tabular benchmark (DiaBD, 5,288 records, 2025) and on a larger synthetic India cohort (15,000 records) designed to match national demographics and covariate distributions. Models were compared to XGBoost (centralized classical baseline) and an advanced federated baseline (FedDL). Experiments used stratified 5-fold cross-validation and reported AUROC, AUPRC, F1-macro, expected calibration error (ECE), computational cost (GFLOPs), training time, and privacy budget (ϵ,δ). FedHybNet achieved AUROC 0.89 and AUPRC 0.69 at ϵ=2.0, improving AUROC by 0.01–0.03 (≈1.1–3.5% relative) over baselines and reducing calibration error (ECE) by 28.9–41.8% relative. Training cost increased due to the Transformer encoder and DP overhead, but the privacy–utility trade-off favored FedHybNet versus non-DP FL. These results indicate FedHybNet is practical for privacy-aware, interpretable diabetes risk prediction in distributed clinical settings.