Leveraging OGTT derived metabolic features to detect Binge-eating disorder in individuals with high weight: a "seek out" machine learning approach.

Abstract

Binge eating disorder (BED) carries a 6 times higher risk for obesity and accounts for roughly 30% of type 2 diabetes cases. Timely identification of early glycemic disturbances and comprehensive treatment can impact on the likelihood of associated metabolic complications and the overall outcome. In this study, machine learning techniques were applied to static and dynamic glucose-derived measures to detect BED among 281 individuals with high weight. Data from the classic (2 h) and the extended (5 h) glucose load were computed by multiple algorithms and two models with the most relevant features were trained to detect BED within the sample. The models were then tested on an independent cohort (N = 21). The model based on the 5 h-long glucose load exhibited the best performance (sensitivity = 0.75, specificity = 0.67, F score = 0.71) diagnosing BED in 7 out of 10 cases. Sex, HOMA-IR, HbA1c and plasma glucose in different times, and hypoglycemia events were the most sensitive features for BED diagnosis. This study is the first to use metabolic hallmarks to train ML algorithms for detecting BED in individuals at high risk for metabolic complications. ML techniques applied to objective and reliable glycemic features might prompt the identification of BED among individuals at high risk for metabolic complications, enabling timely and tailored multidisciplinary treatment.