186 - HYPO- AND HYPERGLYCEMIA PREDICTION FROM POOLED CONTINUOUS GLUCOSE MONITOR DATA
Abstract
Background and Aims
Proactive self-care prevents hypo- and hyperglycemia, improving time-in-range and A1c. To be proactive, people with diabetes guess if BG is rising or falling and act accordingly. An inaccurate guess can lead to a harmful over or under correction (hypo- and hyperglycemia). To prevent such harm, we used app-entered BG data from CGMs to accurately predict upcoming low/high BG.
Methods
Data included contextual information, self-care, and BG values from app users with CGMs. Data were used to train a supervised learning model. The model generated predictions of each user’s BG 30 and 60 minutes into the future, as well as whether BG would be low (<70 mg/dL) or high (>180 mg/dL) in the next 30 minutes, one hour, and four hours.
Results
The mean absolute relative difference (MARD) for 30-minute predictions was 4.3%, with 99.7% of predictions falling in Zone A of the Clarke Error Grid, and 99.9% in Zone A or B. The MARD for 60-minute predictions was 13.4%, with 79.4% in Zone A, and 98.4% in Zone A or B. Hypoglycemia predictions showed 93.2% recall and 89.4% precision at 30 minutes, 83.2% recall and 74.1% precision at one hour, and an area under the ROC curve (AUC) of 91.9% at 4 hours. Hyperglycemia predictions showed 98.9% recall and 97.6% precision at 30 minutes, and 95.0% recall and 92.6% precision at one hour.
Conclusions
Pooling BG data from thousands of One Drop app and CGM users confers accurate, short-term BG forecasts, which can facilitate proactive, safer self-care.
