Differential predictive performance of machine learning models for schizophrenia relapse prediction using behavioral versus physiological passive sensors: A systematic review and meta-analysis.

On an international level, 1% of the population develops schizophrenia, and over 80% of those relapse, even with the best treatment. Relapse prognosis is important and needs to be identified early; however, this is not routinely performed in clinics, and traditional clinical monitoring does not reflect real-time changes in symptoms. The incubation of passive sensing technologies in smartphones and wearable devices can enable continuous and objective monitoring of the behavioral and physiological markers of relapse risk. This systematic review and meta-analysis aimed to evaluate the predictive validity of behavioral and physiological sensors and multimodal approaches for predicting relapse in patients with schizophrenia-spectrum disorders. According to the PRISMA 2020 guidelines, the PubMed, Google Scholar, Scopus, IEEE Xplore, ArXiv, and PsycInfo databases were accessed (2015-2025). Backward and forward citation searches of key studies were also used to augment the search strategy. Of the 741 records, 55 studies were included, and 50 were included in the meta-analysis. The pooled AUCs for the behavioral sensors, physiological sensors, and multimodal systems were 0.720, 0.729, and 0.806, respectively, with the multimodal systems demonstrating the best performance. In the current study set, overall, deep learning models performed better than classical machine learning methods (AUC 0.789 vs. 0.719), and larger studies showed greater predictive accuracy than smaller ones. The methodological quality assessed by the risk of bias was generally high. Overall, multimodal passive sensing showed the best predictive accuracy and may be useful for the continuous monitoring of relapse in a clinical environment.