Personalized machine learning models for noninvasive hypoglycemia detection in people with type 1 diabetes using a smartwatch: Insights into feature importance during waking and sleeping times

Le résumé fourni par la source

Hypoglycemia is a major challenge for people with diabetes. Therefore, glycemic monitoring is an important aspect of diabetes management. However, current methods such as finger pricking and continuous glucose monitoring systems (CGMS) are invasive, and hypoglycemia has still been shown to occur despite advancements in CGMS. Consequently, a growing body of research has been directed toward noninvasive hypoglycemia detection, relying on data from medical devices and wearables that can record physiological changes elicited by hypoglycemia. Consumer-grade wearables such as smartwatches remain an attractive yet underexplored candidate for such applications. Therefore, we explored the potential of a consumer-grade wearable for hypoglycemia prediction and investigated differing feature importance during waking and sleeping times. Smartwatch data from 18 adults with type 1 diabetes was collected, preprocessed, and imputed. Machine learning (ML) models were built using a tree-based ensemble algorithm to detect hypoglycemic events registered by CGMS. Models were built in a personalized manner using the same participant's data for training and testing, with separate modeling for daytime and nighttime. The relative importance of input features on model decisions was analyzed using SHAP (SHapley Additive exPlanations). Seventeen personalized models were built with an average area under the receiver operating characteristic curve (AUROC) score of 0.74 ± 0.08. Average specificity and sensitivity were 0.76 ± 0.18 and 0.71 ± 0.15, respectively. Time-of-day, activity, and cardiac features showed comparable importance in daytime models (29.9%, 28.5%, and 24%, respectively), while in nighttime models, cardiac features demonstrated the highest importance (42.2%) followed by time-of-day features (37.5%) and respiratory features (15.2%). In summary, we demonstrate the potential of consumer-grade wearables in noninvasive hypoglycemia detection. By additionally considering different physiological states (waking and sleeping) during modeling, our results offer further insights into differences in relative feature importance influencing the model's decision, guiding future research in this area.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.