Machine learning glucose forecasting models for septic patients

Sepsis-induced glucose fluctuations present major challenges in critical care, underscoring the importance of accurate glucose monitoring and forecasting to improve patient outcomes. This study introduces a suite of forecasting models trained using continuous glucose monitoring data from a diabetic patient with sepsis (19,621 data points). The models include four transformer-based ones (iTransformer, Crossformer, PatchTST, FEDformer), a dynamic linear model (DLinear), and an ensemble zero-shot inference method leveraging ChatGPT-4. Model performance was evaluated for 15-, 30-, and 60-minute prediction horizons with an optimized 30-minute lookback window. PatchTST achieved the lowest mean maximum percentage error (MMPE) for short-term forecasts (3.0% at 15 minutes), while DLinear excelled at longer horizons (7.46% and 14.41% MMPE at 30 and 60 minutes, respectively). The ensemble ChatGPT-4 approach also showed competitive results. Overall, this work offers a toolbox of advanced forecasting models for ICU glucose prediction and management. The comprehensive comparison among the models highlights the promise of machine learning models-particularly DLinear and PatchTST-in supporting glucose monitoring and ultimately digital twin implementations, paving the way toward personalized and adaptive glycemic control in septic patients.