CMPB · Under review · 2026

Federated Diabetes

Privacy-preserving federated learning for diabetes risk prediction, externally validated on 1.28 million BRFSS patients.

Rajveer Singh Pall · 2026

ProblemDiabetes risk models trained on one hospital's data often fail elsewhere — and privacy rules block pooling patient records to fix that.

ApproachTrain locally at each site, share only model updates, and compare four privacy-preserving ways of combining them.

ResultTested on 1.28 million people it never trained on, the federated model generalizes 40% better than one trained centrally.

SignificanceHospitals can collaborate on better risk models without ever sharing a patient record.

The Discovery in One Figure

Centralized training

Wider generalization gap between training and real-world performance.

→

Federated training

40% narrower gap — without any site ever sharing a patient record.

Generalization gap, centralized vs. federated

The Paper in Five Minutes

Hospitals cannot share patient records, so how do they train one model together? Federated learning sends the model to the data instead of the data to a server. This paper checks what federated papers usually don't: whether the jointly trained model still works on a different population of 1.28 million people, whether it treats the elderly fairly, whether its probabilities are honest, and whether formal privacy protection actually holds. Three of the four go surprisingly well — the federated model beats the centralized one externally. The fourth is an honest negative: differential privacy destroys the model at realistic settings, and the paper says so plainly.

The Research Question

Clinical risk models trained centrally tend to fail on out-of-domain patients, and privacy constraints often prevent pooling raw records across institutions in the first place. The goal: a diabetes risk model trained without sharing data that still holds up on a population it never saw — audited on generalisation, fairness, calibration, and privacy simultaneously.

How It Works

Demographically partitioned NHANES client nodes train locally; four aggregation strategies — FedAvg, FedProx, FedNova, SCAFFOLD — are compared over 50 rounds under non-IID participation, with a DP-SGD privacy sweep (Opacus) tracing the utility–privacy frontier and isotonic calibration repairing probability estimates.

Split realisticallythree demographically unequal simulated hospital nodes

Train federatedFedAvg · FedProx · FedNova · SCAFFOLD vs centralized controls

Validate externally1.28M BRFSS records — a population no node ever saw

Audit fairness + calibrationelderly–young gap · ECE with recalibration ladder

Stress privacyDP-SGD ε-sweep — where utility survives, and where it collapses

The Results

0.757External AUC (BRFSS)

1.28MExternal patients

0.001Calibration error (ECE)

40%Narrower generalization gap

External validation on 1.28M BRFSS respondents reaches AUC 0.757 with ECE 0.001 after calibration — a generalization gap 40% narrower than centralized training, with utility recovered above ε≈10. Subgroup fairness is reported across demographic strata, and the full pipeline is reproducible end-to-end.

External AUC on 1.28M unseen records

FedAvg (federated)0.757

Centralized, same architecture0.749

Centralized XGBoost0.700

The federated model generalises better than both centralized controls.

External validation ROC on 1.28M BRFSS records — External validation on 1.28M BRFSS records: federated and centralized models compared where it counts.

Elderly–young fairness across training strategies — the axis most clinical papers never report.

Privacy–utility trade-off across the epsilon sweep — The privacy–utility trade-off: DP-SGD across the ε sweep, including the collapse the paper reports plainly.

What This Changes

It moves federated-learning evaluation beyond "privacy by architecture" toward audit-first deployment evidence — including the honest negative on differential privacy at strict budgets, stated plainly rather than buried.

↗ View on GitHub ↗ Full research page

Citation

Rajveer Singh Pall. "Privacy-Preserving Federated Learning for Diabetes Risk Prediction Across Demographically Heterogeneous Hospital Nodes." Under journal review, 2026.