Prediction of psychotic disorder in individuals with clinical high-risk state by multimodal machine-learning: A preliminary study

Objective markers which can reliably predict psychosis transition among individuals with at-risk mental state (ARMS) are warranted. In this study, sixty-five ARMS subjects [of whom 17 (26.2%) later developed psychosis] were recruited, and we performed supervised linear support vector machine (SVM) with a variety of combinations of.modalities (clinical features, cognition, structural magnetic resonance imaging, eventrelated.potentials, and polyunsaturated fatty acids) to predict future psychosis onset. While single-modality SVMs showed a poor to fair accuracy, multi-modal SVMs revealed better predictions, up to 0.88 of the balanced accuracy, suggesting the advantage of multi-modal machine-learning methods for forecasting psychosis onset in ARMS.