Predicting transition to psychiatric disorders in clinical high-risk for psychosis subjects using machine learning and facial features.

BACKGROUND: Clinical-high-risk for psychosis (CHR) status is increasingly viewed as a transdiagnostic risk state, but scalable markers of adverse outcomes remain limited. We tested whether facial dynamics extracted from baseline videos could predict transition to any psychiatric disorder in CHR individuals. METHODS: In the SSAPP cohort, CHR participants were assessed at baseline with SIPS and SCID-5 and followed for a mean of 22.8 months. Baseline videos from 50 participants were analyzed. Facial landmarks and Action Units were extracted with OpenFace 2.0 from Subject Overview (SO) and Memory Recall (MR) recordings. PCA-derived features and AU summaries were entered into class-weighted logistic-regression models using leave-one-out cross-validation. RESULTS: Of the modelled sample, 29 participants transitioned to a psychiatric disorder, including 9 psychotic and 20 non-psychotic outcomes; 21 did not transition. PCA features from SO videos showed the best performance (F1-score = 72%, sensitivity = 66%, specificity = 76%, balanced accuracy = 71%, MCC = 0.41), outperforming MR-derived and AU-based models. Sensitivity was higher for psychotic transitioners (78%) than for non-psychotic transitioners (60%). Predictive signal was concentrated in higher-order PCA components reflecting subtle facial-motor variation. CONCLUSIONS: Facial dynamics extracted from clinical videos may provide scalable markers of adverse outcomes in CHR populations. Future studies should aim at larger samples and external validation.