DISTRIBUTED AND PERSONALIZED INSULIN PREDICTION VIA FEDERATED LEARNING WITH DECISION TREE OPTIMIZATION ON CGM STREAMS

Abstract

Background:Accurate insulin prediction is crucial for effective diabetes management. Traditional centralized models pose challenges in terms of data privacy, model generalization, and adaptability to patient-specific needs. With the growing use of Continuous Glucose Monitoring (CGM) systems, there is a need for intelligent, privacy-preserving frameworks that can learn from distributed data without compromising confidentiality.Problem:Existing models fail to effectively integrate patient-specific physiological data while ensuring data privacy and minimizing communication overhead. There is a need for a secure, decentralized system capable of learning from multiple clients and offering accurate insulin dosage recommendations.Methods:This study proposes a Federated Learning-based insulin prediction model that integrates optimized decision trees. The system follows a structured workflow: raw CGM data are collected and preprocessed (including filtering and segmentation), followed by feature extraction (glucose level, insulin, activity, and carbohydrate intake). Features are distributed across multiple patients. Federated averaging is applied at the central server after each patient performs local model updates. Optimization techniques further enhance model accuracy. Results:The proposed Federated + Optimized Decision Tree model outperforms both local and centralized models, achieving higher accuracy and better ROC performance. The model demonstrates convergence across federated rounds and maintains low prediction error across various patients. Time-series analysis reflects clinically meaningful insulin-glucose interactions. Correlation analysis confirms the model's ability to learn physiological relationships effectively.Conclusion:The integrated federated learning approach, combined with decision tree optimization, ensures both model accuracy and data privacy. Its patient-specific adaptability and real-time prediction capabilities make it a strong candidate for future deployment in smart diabetes management systems.