A Machine-Learning Model Accurately Predicts Projected Blood Glucose

Background: Clinical guidelines do not specify the frequency of self-monitoring of blood glucose (SMBG) for the ∼85% of people with type 2 diabetes (T2D) not on intensive insulin therapy. For these individuals, insurance covers a limited amount of testing supplies, and blood glucose (BG) checks are infrequent. With limited BG data, people may not know how foods, activity, and other factors affect their BG levels. Other, non-SMBG-dependent, approaches are needed to illustrate how self-care choices affect BG. Objectives: 1.) Use a large dataset to train a machine learning model 2.) Use the model and minimal data inputs to predict BG values at variable times 3.) Test the accuracy of the prediction. Methods: The One Drop | Mobile app collected 1,923,416 BG measurements from 14,706 people with noninsulin treated T2D. Contextual information (CI), when available, included demographics, health metrics (e.g., weight, A1c), and self-care. Inputs to each BG prediction included a prior BG and available CI. The model did not distinguish whether BGs with similar CI were from the same or different users. Forecast horizons were set by the time since prior BG and varied from 10 minutes to several days. Machine learning algorithms to predict BG values were trained and vetted on BGs entered prior to Sept. 2017 (83% of all BGs). BGs (17%) entered from Sept.-Nov. 2017 were held out and predicted. Results: Users were 59% male, with 80% from North America, 9% from Europe, and 11% from elsewhere; 50% were diagnosed with T2D in the past 3 years. The median and mean absolute error of holdout predictions were 14.2 and 21.3 mg/dL respectively, with 91% of predictions within +/-50 mg/dL. Discussion: A machine learning model with minimal inputs accurately predicts future BG values in noninsulin treated people with T2D. Whether people modify their behavior (e.g., eating less carbohydrates) after knowing a planned behavior (e.g., a 60 carbohydrate gram meal) produces an undesirable BG (>140 mg/dL postprandial) is an empirical question worth investigating.