Cross-Institutional Diabetes Prediction: A Comparative Evaluation of Federated Learning Approaches Versus Centralized Machine Learning Models Under Strict Data Privacy Constraints

• Maxwell Benefit

  Author

The proliferation of electronic health records (EHR) and machine learning has created unprecedented opportunities for early diabetes detection, yet stringent data privacy regulations—including PIPEDA, HIPAA, and GDPR—fundamentally obstruct the centralized data aggregation required for traditional predictive modeling. This study addresses the critical research gap in privacy-preserving diabetes prediction through a comparative evaluation of federated learning (FL) architectures against centralized machine learning benchmarks under simulated cross-institutional constraints. Using real-world clinical data from multiple Canadian provinces, we implemented both logistic regression and multilayer perceptron (MLP) models within a FedAvg federated framework, comparing their performance against province-local models and centralized baselines. The federated MLP achieved an AUC of 87.2% and F1-score of 85.4%, demonstrating performance within 3% of the centralized model while completely eliminating patient data sharing across institutions. However, federated logistic regression exhibited a 13% AUC reduction relative to its centralized counterpart (78.1% vs. 91.2%), revealing significant algorithmic sensitivity to the FL paradigm. The findings establish that federated learning, when implemented with appropriately complex architectures like MLP, offers a viable privacy-preserving alternative to centralized diabetes prediction, with critical implications for healthcare systems navigating the tension between predictive accuracy and regulatory compliance.

• PDF

Copyright (c) 2026 Maxwell Benefit (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.