Abstract
BACKGROUND: Diabetes Mellitus is a common, chronic metabolic disorder affecting the cardiovascular system, autonomic nervous system, and sleep quality. Diabetes affects diverse physiological data including heart rate variability, distal body temperature, and sleep duration. We hypothesized that biologically informed features from wearable device data, combined with appropriate application of longitudinal data, can capture physiological covariates of diabetes and support the noninvasive detection of diabetes. METHODS: We obtained 4 months and 7 days of wearables data (Oura Ring) from 389 individuals self-reporting diabetes and 10,820 people self-reporting no diabetes diagnosis from the TemPredict database. We selected 36 features of sleep, circadian disruption, and distal body temperature from literature and evaluated whether time windows of these features could be classified to be from individuals self-reporting diabetes (N = 236) or self-reporting no diabetes diagnosis (N = 282). RESULTS: Here we show longer time windows of input perform better, with the best algorithm (21-nights) achieving 0.88 Area under ROC (AUROC) and 0.80 Area under Precision Recall (AUPRC) (0.30 improvement over random). Feature analyses reveal the importance of further derived distal body temperature features (increase AUROC by 0.0724), especially to differentiate other chronic conditions from diabetes. The model achieves 0.80 AUROC and 0.28 improvement over random in AUPRC in an imbalanced cohort drawn from 6,658 individuals, emulating a general population. CONCLUSIONS: These results indicate the value of biologically informed features and longitudinal data for identifying people with diabetes and further, suggest that these methods could make such separations possible for other chronic conditions that affect sleep and inflammation.