Background: Knowing at bedtime whether hypoglycemia is likely to occur while sleeping could empower people with diabetes to take preventive action, and could reduce concerns about either being awoken by a continuous glucose monitor (CGM) alarm, or sleeping through the alarm and not being able to respond. We used CGM and self-care data collected in the One Drop app to predict the occurrence of overnight hyperglycemia.
Method: Over 500,000 person-nights of sleep, blood glucose (BG), self-care and contextual data from over 3000 app users with CGMs were pooled and used to train and test a supervised learning model to predict, as of bedtime, the minimum BG value for the coming night. A related model was trained to predict the probability of any BG value being lower than 70 mg/dL during the night.
Results: Test-set predictions of minimum overnight BG value had a mean absolute relative deviation (MARD) of 18.6%. For test-set predictions of the probability of BG lower than 70 mg/dL, the area under the receiver operating characteristic curve (AUC) was 82.2%. About 30% of predictions could be identified at prediction time as having higher accuracy. For those, test-set minimum BG prediction MARD was 15.4%, and probability of BG lower than 70 mg/dL prediction AUC was 87.0%.
Conclusion: Pooling sleep, BG, behavioral and self-care data from thousands of One Drop app users who also use CGM accurately predicts the probability of overnight hypoglycemia. Having such predictions at bedtime can facilitate preventative action, reduce anxiety and improve both sleep and quality of life.
Y. Wexler: Employee; Self; One Drop. D. Goldner: Employee; Self; One Drop. G. Merchant: Employee; Self; One Drop. A. Hirsch: Employee; Self; Informed Data Systems Inc. B. Huddleston: Employee; Self; One Drop. Stock/Shareholder; Self; One Drop. J. Dachis: Board Member; Self; One Drop. Employee; Self; One Drop. Stock/Shareholder; Self; One Drop.
