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

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Endocrinology

Background:

• Accurate blood glucose (BG) forecasting is crucial for type 1 diabetes (T1D) management, particularly for decision support systems and artificial pancreas devices.
• Deep learning (DL) models show promise for BG forecasting, but the impact of various input features on their performance requires further investigation.

Purpose of the Study:

• To assess the impact of different input features, including continuous glucose monitoring (CGM), meal (carbohydrate intake), insulin, and physical activity (PA) data, on the performance of DL models for BG forecasting in T1D.
• To evaluate seven distinct input configurations across five DL models using real-world data from a large T1D cohort.

Main Methods:

• Trained and evaluated five DL models using four weeks of daily-life data from 497 individuals with T1D.
• Compared seven input configurations, ranging from univariate CGM to comprehensive models including CGM, insulin, carbohydrate (CHO) intake, heart rate (HR), and exercise data.
• Assessed model performance using Root Mean Squared Error (RMSE) and Time Gain (TG) at a 30-minute prediction horizon (PH).

Main Results:

• Incorporating additional input features progressively improved DL model performance at the 30-minute PH.
• The CNN-Transformer model demonstrated the greatest improvement; incorporating all features reduced RMSE from 21.02 ± 3.5 mg/dL to 18.63 ± 4.18 mg/dL and increased TG from 10.38 ± 1.33 minutes to 12.12 ± 2.64 minutes.
• Prediction accuracy during exercise significantly improved only when PA data were included (RMSE reduced from 28.72 ± 9.18 mg/dL to 24.7 ± 7.8 mg/dL).

Conclusions:

• The inclusion of diverse data inputs, such as meal, insulin, and PA, enhances the performance of DL models for BG forecasting in T1D.
• While statistically significant improvements were observed, the modest magnitude of these changes suggests limited immediate clinical benefit.
• Further research is needed to optimize DL models and translate performance gains into tangible clinical advantages for T1D management.