Differential predictive performance of machine learning models for schizophrenia relapse prediction using behavioral versus physiological passive sensors: A systematic review and meta-analysis.
Meesala Krishna Murthy
✦ AI-curated · Sources linked
In 30 seconds
This systematic review and meta-analysis assessed passive-sensor machine-learning models for predicting relapse among people with schizophrenia-spectrum disorders. Searches identified 55 studies, of which 50 contributed to the meta-analysis. Pooled area under the curve (AUC) was 0.720 for behavioral sensors, 0.729 for physiological sensors, and 0.806 for multimodal systems, which had the highest predictive performance. Deep learning models also outperformed classical machine-learning methods in the included studies (AUC 0.789 vs. 0.719). The authors conclude that multimodal sensing may support continuous relapse monitoring.
Key findings
- Pooled AUC was 0.806 for multimodal systems, compared with 0.720 for behavioral and 0.729 for physiological sensors.
- Deep learning models had higher AUC than classical machine-learning methods (0.789 vs. 0.719).
- The review included 55 studies; 50 were included in the meta-analysis.
Why it matters
Relapse risk may change between clinical assessments, while conventional monitoring may not capture real-time symptom changes. The higher pooled predictive performance for multimodal passive sensing suggests that combining behavioral and physiological signals could be useful for continuous monitoring. These results provide an evidence base for further evaluation of sensor-based prediction in clinical settings.
What to keep in mind
Predictive performance varied with study size and model type; the abstract does not report prospective clinical validation or whether sensor-guided monitoring improves patient outcomes.