Abstract
AIMS: Our study aimed to develop a machine learning ensemble to distinguish "at-risk mental states for psychosis" (ARMS) subjects from control individuals from the general population based on facial data extracted from video-recordings. METHODS: 58 non-help-seeking medication-naïve ARMS and 70 healthy subjects were screened from a general population sample. At-risk status was assessed with the Structured Interview for Prodromal Syndromes (SIPS), and "Subject's Overview" section was filmed (5-10 min). Several features were extracted, e.g., eye and mouth aspect ratio, Euler angles, coordinates from 51 facial landmarks. This elicited 649 facial features, which were further selected using Gradient Boosting Machines (AdaBoost combined with Random Forests). Data was split in 70/30 for training, and Monte Carlo cross validation was used. RESULTS: Final model reached 83 % of mean F1-score, and balanced accuracy of 85 %. Mean area under the curve for the receiver operator curve classifier was 93 %. Convergent validity testing showed that two features included in the model were significantly correlated with Avolition (SIPS N2 item) and expression of emotion (SIPS N3 item). CONCLUSION: Our model capitalized on short video-recordings from individuals recruited from the general population, effectively distinguishing between ARMS and controls. Results are encouraging for large-screening purposes in low-resource settings.