Predictive Modeling in Mental Health

Abstract

Mental health prediction models -- forecasting risk of depression onset, suicidal behaviour, psychosis transition, and treatment response -- have the potential to transform clinical practice by enabling early intervention for high-risk individuals before crises emerge. The methodological challenges are substantial: mental health outcomes are shaped by complex longitudinal interactions between biology, psychology, and social environment; diagnostic boundaries are fluid and cultural; and the consequences of both false positive and false negative predictions are clinically significant. Recent advances in structured EHR analytics, digital phenotyping from smartphone and wearable data, and transformer-based analysis of clinical documentation have expanded the predictor space substantially beyond what traditional psychiatric risk scores could incorporate. We evaluated five mental health prediction frameworks applied to 184,642 patients across 18 European mental health services in Paris, Amsterdam, Stockholm, Copenhagen, and Berlin between 2017 and 2024, spanning depression, suicidality, first-episode psychosis, and bipolar disorder. Frameworks ranged from clinical questionnaire scores through multimodal deep learning integrating EHR, digital phenotyping, and clinical documentation. Performance was benchmarked using our Mental Health Prediction Clinical Value Score (MHPCVS), integrating predictive accuracy, calibration, clinical utility, fairness, and ethical acceptability. The multimodal deep learning framework achieved the highest MHPCVS (0.914), improving 12-month depression relapse prediction AUC from 0.68 (PHQ-9 baseline) to 0.89 and suicide attempt prediction from 0.64 to 0.86. Digital phenotyping from smartphone sensor data contributed the largest single predictor gain for mood disorders, while clinical notes NLP contributed most for psychosis transition prediction. Ethical acceptability -- a critical dimension for mental health AI -- required active design and was not automatically achieved through accuracy optimisation alone.