Federated deep learning for early detection and progression modelling of diabetic nephropathy

Abstract

One of the main causes of end-stage renal disease is diabetic nephropathy (DN). Patient privacy-sensitive data is a restriction in centralized machine-learning approaches for the healthcare domain. This research introduces a privacy-preserving federated deep-learning (Fed-DN) method for early diagnosis of DN and modelling disease progression and its impact. Fed-DN combines safe aggregation with differential privacy (DP) using Opacus and multi-modal data sources (electronic health records, fundus imaging, and histopathology). Both simulation tests with multi-modal data sets distributed across hospitals show an efficient privacy-utility trade-off, with Fed-DN achieving near-centralized performance (AUC ≈ 0.90) (Accuracy = 87%) while maintaining a privacy budget of ε ≈ 3.1. This is a good example of the trade-off achieved between model accuracy and privacy protection. The technique does not reveal patient information, which enables collaborative, scalable healthcare analytics. Furthermore, it enhances trust among patients and institutions.