Hypoglycemia Prediction from Pre-Exercise Continuous Glucose Monitoring Time Series: Deep-Learning Versus Feature-Based Machine-Learning Approaches

Exercise-induced hypoglycemia poses significant risks to individuals with type 1 diabetes (T1D), discouraging their involvement in physical activity and preventing them from exploiting its known benefits. Although continuous glucose monitoring (CGM) systems may alert the patient of impending high or low glucose levels, they still lack efficient and specific algorithms enabling a safe hypoglycemia prediction prior exercise. Deep learning (DL) and machine learning (ML) approaches now represent potential solutions for the development of clinical decision support systems in the field of diabetes. Thus, this study examined CGM time series preceding the exercise and evaluated the use of DL and feature-based ML approaches to predict hypoglycemia occurring from the start of exercise until the following morning. We analyzed 47 CGM recordings pertaining to T1D young patients that performed a controlled exercise on a treadmill. Recordings were labelled as HYPO or NO-HYPO, based on the occurrence or absence of hypoglycemic episodes, respectively. A Bidirectional-Long Short-Term Memory (Bi-LSTM) model was trained using raw CGM time series and validated through leave-one-out cross validation. Data augmentation techniques were applied to enhance the small and unbalanced dataset. Results were compared to those previously obtained with ML approaches based on clinically relevant CGM features extracted from the same data. The comparison showed that while the Bi-LSTM DL model exhibited very high specificity (SPDL=84.6%vs SPML=76.1%), feature-based ML was superior in terms of sensitivity (SEML=87.2% vs SEDL=67.7%), especially important in clinical decision making. In conclusion, DL and ML-based approaches both revealed their potential for exercise-induced hypoglycemia prediction.