COMBINING A PHYSIOLOGICALLY-BASED MODEL AND AN UNSUPERVISED LEARNING METHOD TO DETECT HIGH-FAT/HIGH-PROTEIN MEALS IN INDIVIDUALS WITH TYPE 1 DIABETES UNDER FREE-LIVING CONDITIONS

Background and Aims Postprandial glycemic excursions in Type 1 Diabetes (T1D) are strongly influenced by macronutrient content (MC); however, insulin dosing strategies are based on carbohydrates only. In this work, we developed an unsupervised learning method to detect high-fat/high-protein (HFHP) meals exploiting features derived from CGM and insulin pump (IP) data. Methods We used 353 4-hour postprandial CGM and IP profiles of 120 individuals with T1D (age: 15.5±11.5 years; weight: 51.3±28.0 kg). MC of each meal was recorded. We fitted a Gaussian mixture model on features extracted from CGM profiles (data-driven features, DDF) and derived from the identification of a physiologically-based model (Model-based features, MBF) combining CGM and IP data. Features were selected using 5-fold cross-validation. Clusters were labelled based on MC. Results The most predictive features were 2 MBF (the area under the curve of glucose rate of appearance in the first 2h after the meal, AUCRa2h, and the half-life of gastric retention, HLGR) and one DDF (the glucose rate of change at 2h, GRC2h). The silhouette index identified the existence of 2 clusters (Fig.1A) with a cross-validation accuracy of 99.7%. Labelling detected statistical differences between the cluster with low (L) and that with high fat/ protein content (H) (Fig.1B). Interestingly, the removal of MBF resulted in unsatisfactory clustering and labeling results. Conclusions We developed a physiology-informed model to detect HFHP meals using CGM and IP data. Next steps include adapting our method to real-time scenarios, enabling timing detection of meals requiring additional insulin dosing.