Leveraging T1DEXI Study to Develop Deep Learning Models for Predicting Blood Glucose Levels

Background and Aims: Deep learning (DL) models are rapidly gaining popularity and are increasingly being integrated into advanced decision support systems and artificial pancreas technologies to enhance Type 1 Diabetes (T1D) management. In this context, large datasets play a crucial role for developing accurate blood glucose (BG) forecasting algorithms. Leveraging the extensive T1DEXI dataset, this study aims to develop and evaluate advanced DL algorithms for BG prediction. Methods: The dataset includes 497 adults from a 4-week real-world trial of at-home exercise sessions. Participants provided data on exercise, food intake, administered insulin, heart rate, and continuous glucose monitoring (CGM) data. We investigated Long Short-Term Memory (LSTM), Convolutional Neural Network-LSTM (CNN-LSTM), and CNN-Transformer models, using an AutoRegressive with eXogenous inputs (ARX) model as baseline comparator. Models employed as input historical CGM measurements, meal logs, and insulin data to forecast BG levels at 30, 45, 60, and 90 minute prediction horizons (PHs). In addition to RMSE, time gain (TG) is used to quantify the time advantage provided by the algorithms, enabling patients to take preemptive actions against impeding adverse events. Results: presented in Table 1 indicate that the CNN-LSTM and LSTM models achieved the best performance, exhibiting the lowest RMSE and highest TG across all PHs. However, all DL models showed comparable performance, consistently outperforming the ARX model (p-value<0.03).