Nearly 90 million adults in the United States have prediabetes with an elevated risk of developing diabetes in the future. Current models to predict changes in glycemic control perform poorly and do not consider daily health behaviors such as exercise. Automated monitoring of these behaviors could inform enhanced risk prediction models that detect patients likely to experience a decline in glycemic control. If such patients could be identified, effective approaches could be better targeted to improve outcomes more quickly.
We conducted a randomized trial with adults with prediabetes to explore if data from wearables could help inform risk prediction models. Participants were randomly assigned to use a waist-worn or wrist-worn wearable to track activity patterns over six months. We tested three models to predict continuous changes in glycemic control during this time.
In all three models, prediction improved when machine learning was used vs. traditional regression, when baseline information with wearable data was used vs. baseline information alone, and when wrist-worn wearables were used vs. waist-worn wearables. We found the best prediction when using ensemble machine learning methods with data from wrist-worn wearables.
Center for Health Incentives and Behavioral Economics, Leonard Davis Institute of Health Economics