Machine learning predicts weight change in a real-world digital health application

Abstract Background Digital health applications offer scalable support for weight loss, but individual outcomes and retention vary substantially. It remains unclear how predictive information evolves during treatment, whether app engagement adds predictive information beyond early weight response, and whether the same factors predict weight loss and retention. Methods Using real-world data from the 12-week digital health application Oviva Direkt for Obesity , we developed machine learning models to predict clinically meaningful weight loss, continuous relative weight change, and full app retention. Weight-change and full-retention analyses included 17,953 and 26,714 adults, respectively. Models evaluated baseline characteristics, 4-week weight, and weekly app engagement metrics separately and in combination to quantify their incremental predictive value and assess how performance changed when early-treatment information was added to pretreatment characteristics. Results Baseline characteristics alone showed limited discrimination for clinically meaningful weight loss at weeks 8 and 12 (ROC AUC 0.54–0.58). Adding 4-week weight substantially improved performance (ROC AUC 0.86–0.92), whereas app engagement alone yielded modest improvements (ROC AUC 0.63–0.67). Adding app engagement to models including 4-week weight provided little further improvement (ROC AUC 0.87–0.92). A similar pattern was observed for continuous weight-change prediction. In contrast, full-retention prediction improved from ROC AUC 0.81 with baseline characteristics to 0.90 with app engagement, while 4-week weight provided no additional benefit. Weight-loss prediction was driven primarily by baseline weight, 4-week weight, and self-defined target weight, whereas engagement-related variables like goal setting and personalized logging recommendations were more influential for full retention. Conclusions Predictive performance for weight-loss outcomes improved substantially when information from the first four weeks, particularly 4-week weight, was incorporated. Accordingly, the week 8 and week 12 models should be interpreted as predictions updated during treatment rather than predictions based solely on pretreatment characteristics. Weight-loss response and sustained intervention use were characterized by distinct predictive information and may require different forms of early, targeted support. These findings provide a basis for further research on the early identification of individuals at risk of suboptimal outcomes and the development of more personalized digital obesity care. External validation in independent, relevant cohorts is required before the models can be considered for clinical implementation.