Privacy-Preserving Federated Learning for Diabetes Risk Prediction Across Demographically Heterogeneous Hospital Nodes

Abstract

Background and Objective. Diabetes risk models achieve strong discrimination on centralised cohorts, but centralisation is constrained by privacy regulation, and models trained at a single site risk demographic bias and miscalibration. We evaluate whether federated learning (FL) can deliver diabetes risk prediction that is simultaneously generalisable, equitable, calibrated, and privacy-preserving across demographically heterogeneous hospital nodes, without centralising patient data. Methods. NHANES 2013-2020 (n = 15,650) was partitioned into three nodes representing young-urban, elderly-rural, and mixed-metropolitan populations. Four FL aggregation strategies - FedAvg, FedProx (mu = 0.1), FedNova, and SCAFFOLD - were trained over 50 communication rounds and validated externally on BRFSS 2020-2022 (n = 1,282,897). A shared global scaler prevented cross-node data leakage. To separate the effect of federation from architecture, a centralised DiabetesNet with identical architecture was trained on the full NHANES pool and, in a sensitivity analysis, on the matched federated subset (n = 11,906). Fairness was measured as the AUC gap between young (18-39) and elderly (>=60) respondents; post-hoc calibration (Platt, isotonic, temperature scaling) and differential privacy (DP-SGD) were also evaluated. Results. FedAvg achieved the highest external AUC, 0.757 [95% CI 0.756-0.758], exceeding the centralised XGBoost baseline (0.700 [0.698-0.701]) by 0.057 and the architecture-matched centralised DiabetesNet (0.749 [0.748-0.750]). Its generalisation gap (internal minus external AUC) was 0.031, versus 0.052 for the full-pool centralised DiabetesNet - a 40% reduction; a sample-matched sensitivity analysis (gap = 0.024) indicated that data composition, not federation alone, accounts for part of this advantage. FedAvg reduced the elderly-young fairness gap from 0.069 (centralised XGBoost) to 0.054, a 21.7% within-study improvement, and isotonic recalibration reduced expected calibration error (ECE) from 0.276 to 0.001 without loss of discrimination. Under DP-SGD, utility collapsed to AUC around 0.50 at epsilon <= 5 and recovered to 0.769 at epsilon = 10; however, Opacus RDP accounting returned epsilon_actual = infinity at that budget, so epsilon = 10 identifies a utility floor rather than a formal privacy guarantee at the per-node sizes used (n around 3,000-4,500). Conclusions. Across four axes of clinical deployability, federated diabetes screening matched or exceeded centralised baselines while keeping patient data local. Relative to a sample-matched centralised model, federation's advantage rests less on the generalisation gap itself than on privacy preservation, improved elderly fairness, and feasibility under regulatory constraints. The framework is a validated proof-of-concept; prospective evaluation in genuinely independent multi-site cohorts remains the essential next step. Preprint - not peer reviewed. Submitted for publication.