Predictors of social and occupational functioning in help-seeking youths accessing an early intervention in psychosis programme: A real-world cohort study using machine learning algorithms

INTRODUCTION: Social and occupational impairment is common in early psychosis, yet predictors of functional outcome in real-world early intervention in psychosis (EIP) remain not fully understood. Machine-learning algorithms (MLAs) may help identify cases at risk of poor functioning and inform tailored interventions. METHODS: We analysed data from 338 adolescents and young adults accessing an Italian EIP programme between 2012 and 2024. Social and occupational functioning at 12 months was assessed with the Social and Occupational Functioning Assessment Scale (SOFAS) and served as the primary outcome. Prediction models were developed in the complete-case set with available 12-month SOFAS (n = 173), using seven MLAs (Generalized Linear Model [GLM], Bayesian GLM, Random Forest, Support Vector Machine, K-Nearest Neighbours, LASSO, XGBoost). Performance was evaluated with repeated 5 × 5-fold cross-validation in the training set and once in the held-out test set. RESULTS: Of 338 participants, 61.0% were male, mean age was 20.93 years (range 15-35). Overall, 42.6% met criteria for first-episode psychosis, 32.0% for an ultra-high risk mental state, and 25.4% for no at-risk mental state. Baseline SOFAS was the most robust predictor of 12-month functioning. LASSO and XGBoost indicated that higher general psychopathology scores, more severe behavioural and cognitive symptoms, presence of schizotypal personality traits, longer duration of untreated psychosis, and lower education were associated with poorer functional outcomes. CONCLUSIONS: Baseline social and occupational functioning emerged as a pragmatic prognostic marker of 12-month functioning in early psychosis. Findings supported function-focused, personalised early intervention strategies and warrant multicentre validation of ML-based prognostic tools in early psychosis programmes.