A clustering-based federated deep learning approach for enhancing diabetes management with privacy-preserving edge artificial intelligence

The increasing prevalence of diabetes necessitates innovative glucose prediction methods that prioritize patient privacy. While edge artificial intelligence (AI) offers potential, its limitations in resource-constrained devices can be mitigated through federated learning (FL). However, challenges remain in accounting for patient variability and optimizing FL for glucose prediction. This research introduces a novel personalized clustering-based federated deep learning (Clu-FDL) model to address these challenges. We develop tailored models that enhance prediction accuracy by clustering patients based on carbohydrate (CHO) intake patterns. Utilizing Simple Recurrent Neural Network (SimpleRNN) and Gated Recurrent Unit (GRU) methods, the study evaluates the performance of local patients who contribute to training the cluster and global (non-cluster) models. The results show that the Clu-FDL approach achieves high precision (0.93), recall (0.96), and F1 scores (0.95), along with low Root Mean Square Error (RMSE) values (11.08 ± 1.77 mg/dL). Additionally, for new patients with different data durations, analysis based on 0.25–3 days of data indicates that Clu-FDL models exhibit greater stability, with smaller RMSE and higher precision, recall, and F1 scores compared to non-clustering models. The study identifies that SimpleRNN and GRU models are most effective for new patients with 9 and 6 days of data. This privacy-preserving, clustering-based personalized approach empowers patients to manage their diabetes effectively. • Introduce a clustering-based federated deep learning model for glucose prediction. • Achieved high prediction accuracy with personalized patient models. • Demonstrate improved model stability with minimal data from new patients. • Ensure patient privacy using federated learning on edge artificial intelligence devices. • Evaluate simple recurrent neural network (SimpleRNN) and gated recurrent unit (GRU) models; GRU performs better.