Evaluating the Effect of Input Features on Deep Learning Models for Blood Glucose Forecasting

In type 1 diabetes (T1D), predicting future blood glucose (BG) concentration tens of minutes in advance is a key element in decision support systems and artificial pancreas devices. Recently, given the availability of large datasets, deep learning (DL) models for BG forecasting have been investigated, showing promising results. However, the impact of typical input features-such as meal, insulin, and physical activity (PA)- on their performance remains still underexplored. The aim of this study is to assess how these inputs may affect DL models performance.We trained and evaluated five DL models on four weeks of daily-life data from 497 individuals with T1D. Seven input configurations, ranging from univariate continuous glucose monitoring (CGM) models to comprehensive approaches incorporating CGM, insulin, carbohydrate (CHO) intake, heart rate (HR), and exercise data, were evaluated. Results indicate that incorporating additional features progressively enhances performance at the 30-minute prediction horizon (PH), with all models showing similar Root Mean Squared Error (RMSE) and Time Gain (TG). For the CNN-Transformer, the model showing the greatest improvement, the univariate approach achieved an RMSE of 21.02 ± 3.5 mg/dL and a TG of 10.38 ± 1.33 minutes. Incorporating all factors reduced RMSE to 18.63 ± 4.18 mg/dL and increased TG to 12.12 ± 2.64 minutes. Notably, prediction accuracy during exercise improved only when PA data were included, reducing RMSE from 28.72 ± 9.18 mg/dL to 24.7 ± 7.8 mg/dL. While these improvements are statistically significant, their potential clinical benefit remains limited due to the modest magnitude of change.