Background: Mobile health (mHealth) delivered goal setting interventions that include financial incentives (FI) can increase physical activity (PA). It is not clear whether a supervised machine learning [ML] algorithm can be used to set better goals and boost intervention effects.

Objective: To examine whether an mHealth PA intervention with FI is improved with the incorporation of an optional ML-driven goal setting algorithm.

Methods: A 17-week pre-post quasi-experiment was conducted among users of the Sprout app, an mHealth PA intervention with FI targeting North American employees (March-July 2022). The study consisted of a five-week ‘run-in’ (baseline) period, where users (i.e., registered with the app for at least five weeks) could earn FI for meeting static daily step goal (i.e., $0.15 USD each day). A 12-week ‘intervention’ period followed. During this period users were given the option of either (a) continuing with their static goal (control) or (b) opting into ML-driven adaptive goals (intervention). The supervised ML algorithm set new daily step goals at the beginning of each week based on users' recent daily step count patterns and cluster-based comparisons (i.e., portion of standard deviation of cluster to which user belonged added to goal). Independent-samples t-tests examined between-group differences in mean daily step count change from baseline (‘pre’: baseline vs. ‘post’: Week 6, Week 12; p<0.05) controlling for covariates (e.g., age, gender).

Results: A total of 1,249 participants (control: n=447; intervention: n=802) were included (59.4% between 30 and 50 years of age; 42.5% women; BMI: 27.9±9.8 kg/m²; baseline daily step count: 6,366±3,617; days registered on app: 633±497). There were no between-group differences in mean daily step count change from baseline at Week 6 (mean difference [MD] [95% CI]: 49 [-579, 480]; p=0.855) nor Week 12 (MD [95% CI]: 443 [-110, 998]; p=0.116).

Conclusion: Compared to controls, PA did not improve among those opting into the ML-driven adaptive goal setting condition. Future research should test different ML approaches (e.g., reinforcement learning) more likely to set refined, PA promoting goals.