AI-based prediction of depression symptomatology in first-episode psychosis patients: insights from the EUFEST and RAISE-ETP clinical trials

Background Depressive symptoms are highly prevalent in first-episode psychosis (FEP) and worsen clinical outcomes. It is currently difficult to determine which patients will have persistent depressive symptoms based on a clinical assessment. We aimed to determine whether depressive symptoms and post-psychotic depressive episodes can be predicted from baseline clinical data, quality of life, and blood-based biomarkers, and to assess the geographical generalizability of these models. Methods Two FEP trials were analyzed: European First-Episode Schizophrenia Trial (EUFEST) ( n = 498; 2002–2006) and Recovery After an Initial Schizophrenia Episode Early Treatment Program (RAISE-ETP) ( n = 404; 2010–2012). Participants included those aged 15–40 years, meeting Diagnostic and Statistical Manual of Mental Disorders IV criteria for schizophrenia spectrum disorders. We developed support vector regressors and classifiers to predict changes in depressive symptoms at 6 and 12 months and depressive episodes within the first 6 months. These models were trained in one sample and externally validated in another for geographical generalizability. Results A total of 320 EUFEST and 234 RAISE-ETP participants were included (mean [SD] age: 25.93 [5.60] years, 56.56% male; 23.90 [5.27] years, 73.50% male). Models predicted changes in depressive symptoms at 6 months with balanced accuracy (BAC) of 66.26% (RAISE-ETP) and 75.09% (EUFEST), and at 12 months with BAC of 67.88% (RAISE-ETP) and 77.61% (EUFEST). Depressive episodes were predicted with BAC of 66.67% (RAISE-ETP) and 69.01% (EUFEST), showing fair external predictive performance. Conclusions Predictive models using clinical data, quality of life, and biomarkers accurately forecast depressive events in FEP, demonstrating generalization across populations.