FEDERATED LSTM FRAMEWORK FOR REAL-TIME INSULIN PREDICTION USING EDGE-TO-CLOUD CGM INTEGRATION

Abstract

Background: Effective diabetes management relies heavily on timely and accurate insulin prediction, particularly with the increased use of Continuous Glucose Monitoring (CGM) systems. However, conventional centralized machine learning (ML) approaches often require the transfer of sensitive health data to remote servers, raising concerns about patient privacy, high communication overhead, and limited adaptability to individual glucose patterns. Problem: Existing prediction models struggle to balance personalization, privacy, and computational efficiency. Centralized learning lacks scalability and exposes raw data, while traditional models like decision trees or statistical regressors fail to capture the temporal complexity inherent in CGM data. Methods: This study proposes a federated learning (FL) framework for real-time insulin prediction, utilizing Raspberry Pi devices as edge nodes for local data processing and model training. Each client trains a lightweight Long Short-Term Memory (LSTM) model on individual CGM sequences. The cloud server aggregates model weights using a secure federated averaging algorithm, avoiding raw data transfer and enabling decentralized model optimization. Results: The proposed federated LSTM approach was evaluated using simulated multi-client CGM datasets and compared against baseline models: centralized LSTM and decision tree. Results demonstrate superior performance by the federated model, achieving higher accuracy (91.4%), F1-score (90.0%), and lower mean squared error (3.75), while maintaining communication efficiency and strong personalization across clients. Conclusion: This work demonstrates the feasibility of deploying a scalable, privacy-preserving, and accurate insulin prediction system using federated deep learning (DL) and edge computing. The proposed architecture supports real-time glucose monitoring and personalized insulin management, offering a practical solution for decentralized digital healthcare.