Machine Learning to Identify Predictors of Lapses in a Beginners’ Exercise Program for Adults with Type 1 Diabetes

People with type 1 diabetes (T1D) are advised to perform the amounts of exercise recommended for the general population but encounter additional barriers including blood glucose management and fear of hypoglycemia. We previously reported that 20 overweight adults (29.5±5.1 kg/m²) with T1D (HbA1c 7.2±1.1%) that received personalized biosensor feedback, completed mood diaries, and accessed on-demand instructional videos for 10 weeks increased their exercise to levels that partially met recommendations (median 64 [IQR 20, 129] minutes/week). Herein we use machine learning methods to predict the occurrence of lapses (i.e., exercise not performed). We modeled exercise lapses using 95 possible predictive features derived from the prior 7 days of: 1) continuous glucose monitor summary statistics (mean, variability, time >180, 70-180, <70 mg/dL) for each daytime, nighttime, and 24hr window; 2) negative glycemic experiences around exercise (hypo/hyperglycemia, excessive insulin-on-board, nocturnal hypoglycemia); 3) subjective fear of hypoglycemia entered each morning and evening; 4) subjective sleep quality; and 5) illness. We created three different models using Random Forest (RF), Support Vector Machine (SVM), and Deep Neural Network (DNN) with an 80:20 train-test split of the data (960 training & 240 testing days). We used upsampling to balance the data between a lapse and non-lapse days. First, the SVM achieved 71.2% with the most important features being sickness, mean nighttime glucose levels, 24hr time <70 mg/dL, and variability. Next, the RF achieved a higher accuracy (74.1%) with the most important feature being sleep (significance 0.16) followed by the negative glycemic experiences around exercise (0.09 - 0.14). The most accurate model was the DNN with 5-fold cross-validation (75.2% accuracy, 77.5% precision, 71.4% sensitivity). Machine learning models that leverage biosensor, mood, and sleep data may be used to predict exercise lapses for adults with T1D, which could inform just-in-time motivational prompts in future mobile exercise interventions. Model performance might be improved by separating day from subject-level effects and adding a sleep biosensor to complement sleep ratings.