Artificial Intelligence for Prediction of Exercise-Associated Hypoglycemia in Physically Active People With Type 1 Diabetes

This is the authors' abstract. We don't add key points for this paper.

Research objectives. Physical activity continues to be a major barrier to achieving ideal blood glucose control in individuals with type 1 diabetes due to the ongoing threat of hypoglycemia. Traditional alerts from Continuous Glucose Monitoring (CGM) systems are mostly reactive as they fail to consider the specific context of physical activity. The purpose of this project is to provide a critical overview of evidence regarding AI and ML techniques for identifying when exercise will cause low blood glucose in individuals who are physically active and have type 1 diabetes. Methods. The primary databases searched for this narrative review with a structured literature search were PubMed/MEDLINE and PubMed Central. Key findings. Exercise-specific models based on repeated-measures random forest and logistic regression achieved AUROC values around 0.83, while broader explainable XGBoost models using CGM data showed excellent short-term predictive performance. The use of simple translation of clinical knowledge into tool-based formats such as GlucoseGO has been shown to retain high levels of performance when only using limited number of easily obtained variables. Conclusions. AI will allow for a meaningful integration with traditional monitoring in order to shift clinical practice from detecting problems after they occur toward predicting the level of risk each patient has when participating in an activity. The direction in which this will be clinically beneficial will be through developing cost-effective, interpretable, validated tools for use within the current CGM and automated insulin delivery systems.