AutoBiGluNet: transformer-based time series modeling for blood glucose prediction in Type 1 diabetes patients

The key to optimizing insulin administration and simplifying the management of Type 1 diabetes (T1D) lies in accurately predicting future blood glucose (BG) levels. Consistently predicting BG levels is a challenging goal due to interindividual biological variability, data quality issues, and the inherent variability of glucose metabolism.

The study aims to predict BG levels across different time horizons by analyzing multimodal data from the BrisT1D and OhioT1DM datasets, comprising CGM measurements, insulin pump data, smartwatch activity data, and dietary carbohydrate data. The purpose of this research was to develop a robust time series model that could handle noise and heterogeneous medical data and that could contribute to clinical decision making for patients with T1D.

A variety of time series transformer models were applied, and the best model was AutoBiGluNet, which is a hybrid deep learning model that uses Autoformer and BiLSTM networks to capture global patterns and temporal dependencies. Data were preprocessed by replacing missing values for time series features through linear interpolation and using zero imputation for other numeric values.

AutoBiGluNet produced the best performance on BrisT1D, achieving an RMSE of 0.0674 ± 0.0006, MAE of 0.0411 ± 0.0004, and R² of 0.9523 ± 0.0003 across five independent runs. On the external OhioT1DM dataset, the model also showed good generalizability, achieving RMSE of 0.88, MAE of 0.52, and R² of 0.93 at the 30-minute prediction horizon.

The model demonstrated strong predictive performance and favorable clinical error-grid results, suggesting potential for future decision-support applications. However, prospective clinical validation is required before considering integration into closed-loop insulin delivery systems.