Personalized Blood Glucose Forecasting Models Using Machine Learning

Accurate blood glucose forecasting can enable improved decision support and glycaemic control for diabetes patients. However, complex patient-specific behaviours make reliable personalized predictions challenging. This work develops data-driven machine learning models to provide personalized blood glucose forecasts. Long short-term memory (LSTM) neural networks are adapted to incorporate patient context through modular architecture layers. The models are trained on continuous glucose monitoring (CGM) data from diabetes patients. Customization techniques such as transfer learning and similarity matching are introduced to tailor the forecasting models to individual patient characteristics and subgroups. Experiments demonstrate significantly improved predictive performance over population-based models, with over 85% accuracy on 30-minute ahead forecasts. The customized LSTM framework has the potential to provide more effective blood glucose predictions and strengthen personalized diabetes management.