Dissecting multimodal predictors of psychosis risk and remission in youth at ultra‑high risk of developing psychosis: An exploratory study

Early identification of individuals at risk of developing psychosis enables timely intervention and better clinical outcomes. Current approach relies on clinical assessments, such as the Comprehensive Assessment for At‑Risk Mental States (CAARMS), which provides an Ultra‑High‑Risk (UHR) classification but have limited predictive precision. Artificial intelligence (AI) models integrating neuroimaging, clinical, genetic, and biomolecular data show promise for improving prognostic accuracy, yet their clinical applicability remains uncertain. We systematically evaluated the contribution of six data modalities, ranging from socio‑environmental factors to polygenic risk scores, using data from 56 UHR participants in the Longitudinal Youth at‑Risk Study (LYRIKS) cohort. Modalities were iteratively included and excluded to generate candidate models tested on two tasks: (1) predicting transition to psychosis within 12 months and (2) predicting remission from UHR status within the same period. Statistical significance was assessed via permutation testing. Of the 56 UHR participants, 12 developed psychosis and 26 achieved remission from UHR status. The model combining CAARMS, socio-environmental risk factors, and social functioning (CAARMS + RISK + HiSoC) achieved the best overall performance for both predicting conversion (MCC = 0.71, SP = 0.93, SE = 0.78) and remission (MCC = 0.64, SP = 0.82, SE = 0.77), and was the only model to significantly outperform null models in the conversion task. Behavioural and socio‑environmental modalities demonstrate strong predictive value for forecasting outcomes in UHR within multimodal AI frameworks. While molecular and genetic modalities remain promising, further advances are needed to translate their high‑dimensional complexity into clinically useful predictive tools.