A Privacy-Preserving Federated Learning Approach for Diabetes Risk Classification using Glucose-Level based Labeling
Abstract
Today, advancements in healthcare technology have made it possible to identify diseases at early stages and communicate patient information in real-time via connected healthcare systems. Nevertheless, securely storing and sharing sensitive healthcare data remains a major challenge. To address this issue, this paper introduces Differentially Private Stochastic Gradient Descent (DP-SGD), a privacy-preserving federated learning method for simulated diabetes-risk classification. Without exchanging their local patient data, several clients work together to develop a global neural network model. DP-SGD uses Gaussian noise and gradient clipping to give example-level differential privacy during local training. The proposed framework was evaluated using a simulated Federated Health Records Dataset containing 1,000 records. The target label is deterministically derived from a glucose-level threshold, while glucose level itself is excluded from the model input to prevent direct target leakage. The model uses seven non-glucose health features and is trained across five federated clients. With a noise multiplier of 1.0, the final model achieved an accuracy of 0.845, while precision, recall, and F1-score for the positive class were 0. Privacy accounting estimated ϵ =262.673 for δ =0.00625 under conservative RDP accounting. The results demonstrate the feasibility of combining federated learning and DP-SGD for privacy-preserving learning, while also highlighting the predictive limitations of the simulated glucose-derived classification task.