Individual patterns of activity predict the response to physical exercise as an intervention in mild to moderate depression.

Physical exercise (PE) as antidepressive intervention is a promising alternative, as shown by multiple meta-analyses. However, there is no consensus regarding optimal intensity and duration of exercise, and there are no objective criteria available for personalized indication of treatment. The aims of this study were (1) to evaluate whether individual activity patterns before intervention can predict the response to treatment; and (2) to evaluate whether the patient outcome can be improved by using prior information on treatment efficacy at individual level. The study included subjects with mild to moderate depression randomized to three PE regimens as antidepressive intervention. Features extracted from actigraphy recordings were used for training linear regression ensembles to predict the response to treatment. The Bayesian analysis of coefficients yielded distinct signatures in enriched feature subsets for each PE. Next, we used a counterfactual approach by virtually assigning each patient to the PE regimen predicted to yield best outcome. This procedure significantly increased the remission rates as compared to random assignment to treatment. Our data suggest that the analysis of individual patterns of activity can identify a PE regimen to yield the best results, and that assignment to PE regimen using this information would significantly increase remission rate.