Bi-LSTM-Based Multivariate Time Series Model for Predicting Blood Glucose in Type-1 Diabetes

Proactive glucose level prediction offers a vital advantage in the daily management of Type-1 diabetes, where unexpected fluctuations can lead to dangerous hypoglycemic or hyperglycemic events. This work introduces a Bi-LSTM- based deep learning model tailored for multivariate time series forecasting, targeting 30-minute and 60-minute blood glucose prediction horizons. Unlike univariate models, our approach incorporates multiple physiological signals—such as CGM values, insulin dosages (basal and bolus), and carbohydrate consumption—to capture the underlying temporal and causal relationships affecting glucose regulation. The model is trained on the OhioT1DM dataset, which comprises high-resolution (5-minute interval) data from 12 Type-1 diabetic subjects. The bidirectional architecture enables the model to process sequential patterns in both forward and backward directions, improving its sensitivity to evolving trends and sharp variations. Evaluation results highlight the model's ability to deliver accurate short- and mid-term predictions, thus supporting timely therapeutic actions and personalized diabetes care.