Publications / Federated Diabetes

Under journal review2026

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

Rajveer Singh Pall

Federated diabetes screening audited on four deployability axes at once — generalisation, fairness, calibration, and privacy — not just accuracy.

Code ↗

The discovery in one figure

HOSPITALS CAN'T SHARE DATA — SO THE MODEL TRAVELS INSTEADNode A · Young Urbann = 4,500Node B · Elderly Ruraln = 3,406Node C · Mixed Metron = 4,000weights,not dataFederatedFedAvgExternal · 1.28M0.757beats centralized 0.700AUDITED ON FOUR DEPLOYABILITY AXES✓ Generalisation40% smaller gap✓ Fairnesselderly gap −21.7%✓ CalibrationECE 0.276 → 0.001✗ DP privacycollapses at ε ≤ 5
Privacy-preserving federation generalises better than the centralized model it replaces.

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

Federated learning is sold on privacy. But does a federated clinical model also generalise externally, treat age groups fairly, and stay calibrated — simultaneously?

How it works

Three demographically heterogeneous simulated hospital nodes over NHANES; FedAvg / FedProx / FedNova / SCAFFOLD vs centralized controls; external validation on 1.28M BRFSS records; isotonic/Platt/temperature recalibration; DP-SGD ε-sweep with Opacus. A 4-page distillation is prepared for the ML4H 2026 Findings track.

  1. 01
    Split realisticallythree demographically unequal simulated hospital nodes
  2. 02
    Train federatedFedAvg · FedProx · FedNova · SCAFFOLD vs centralized controls
  3. 03
    Validate externally1.28M BRFSS records — a population no node ever saw
  4. 04
    Audit fairness + calibrationelderly–young gap · ECE with recalibration ladder
  5. 05
    Stress privacyDP-SGD ε-sweep — where utility survives, and where it collapses

Experimental results

FedAvg reaches external AUC 0.757 [0.756–0.758], beats centralized XGBoost (0.700), shows a 40% smaller internal–external gap than the matched centralized model (0.031 vs 0.052 — partly explained by data composition, stated openly), narrows the elderly–young fairness gap by 21.7%, and recalibrates from ECE 0.276 to 0.001. DP-SGD collapses utility at ε ≤ 5: current per-node sizes give no free privacy guarantee.

External AUC on 1.28M unseen records
FedAvg (federated)
0.757
Centralized, same architecture
0.749
Centralized XGBoost
0.700

The federated model generalises better than both centralized controls.

  • 0.757FedAvg external AUC [0.756–0.758]
  • −21.7%elderly–young fairness gap reduction
  • 0.276 → 0.001ECE after isotonic recalibration
  • ε ≤ 5DP-SGD utility collapse point

Figures from the paper

External validation on 1.28M BRFSS records: federated and centralized models compared where it counts.
Fig. 1External 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.
Fig. 2Elderly–young fairness across training strategies — the axis most clinical papers never report.
The privacy–utility trade-off: DP-SGD across the ε sweep, including the collapse the paper reports plainly.
Fig. 3The privacy–utility trade-off: DP-SGD across the ε sweep, including the collapse the paper reports plainly.

Figures generated by the paper's own pipeline — reproducible from the repository.

What this changes

Moves FL evaluation beyond 'privacy by architecture' toward audit-first deployment evidence — including the honest negative on differential privacy.

Resources

Code ↗
  • Healthcare AI
  • Privacy · Federated
  • Fairness
  • Deployment Shift

Citation

Manuscript under review — citation will be posted on acceptance. Reach me at rajveerpall04@gmail.com.

Where this leads next