We developed a machine learning model that predicts risk of declining engagement (churn) with an mHealth app (Glooko) that supports diabetes self-management. The model uses self-monitored blood glucose (SMBG), demographic, and app engagement data.
We selected 7134 PWD using meter and/or insulin pump devices who used the Glooko diabetes management app in a remote (outside of the clinic) setting between 1/1/2019-7/1/2019 (40,349 total weekly predictions). We trained a gradient boosting algorithm to predict risk of churn in the next 28 days.
We built features using data from the prior 28 days; features included outcomes (counts of glucose checks, hypo/hyperglycemia occurrences, mean blood glucose), app engagement (counts of app clicks and meter and/or insulin pump uploads, duration of sessions), and demographics (diabetes type, gender, age).
The training cohort had median age=22 years (IQR:11-49), 52% female, 60% type 1 diabetes (T1D). We performed out-of-sample validation using 4647 individuals (8646 total weekly predictions). The validation cohort had median age=20 years (IQR:11-49), 52% female, 60.5% T1D. The algorithm predicted individuals at risk of churning with a precision of 0.67 and a recall of 0.8. This correlates to sensitivity=80%, specificity=58%, and positive predictive value=67%.
The present model indicates that it is feasible to predict declining engagement with an mHealth app to support diabetes self-management. The ability to predict churn may help clinics design effective mHealth-based RPM programs. These predictions can allow clinicians to proactively reach out to PWDs at risk of program dropout. Future research should examine whether dropout from RPM programs predicts worsening diabetes-related outcomes.
S. Babikian: Employee; Self; Glooko, Inc. V. Singh: Employee; Self; Glooko, Inc. M.A. Clements: Consultant; Self; Glooko, Inc. Other Relationship; Self; Glooko, Inc.
