Predicting Depression Risk in Early Intervention in Psychosis Services: A Machine Learning Study Utilising Electronic...

Depression is prevalent in early psychosis and associated with poorer functional outcomes and suicide risk. Early identification of depression risk is critical to inform targeted interventions. This study aimed to develop models to predict depression risk using electronic health records (EHRs) from patients accepted into South London and Maudsley (SLaM) Early Intervention in Psychosis Services.

EHRs from 3,559 patients were analysed using a nested cross-validated machine learning framework. Structured sociodemographic, medication, and 66 natural language processing (NLP)-derived clinical note variables were used to train support vector regressors (SVRs) and classifiers (SVCs) to predict future changes in depressive symptoms, antidepressant prescriptions (AD0-6m) and treatment escalation for depression (TE0-6m). Model performance was generalised in out-of-cross-validation samples and across SLaM boroughs. Clustering analyses were also conducted to identify depressive symptom trajectories.

SVRs accurately predicted changes in depressive symptom scores over the first year following referral, with R² ranging 58.5%-72.1%. SVCs predicted AD0-6m with a balanced accuracy (BAC) of 65.6% and TE0-6m with a BAC of 68.0%. Patients with deteriorating trajectories were most likely to receive antidepressants, had higher antipsychotic treatment attempts and a psychotic-affective mixed presentation of symptoms. Model performance remained robust across clusters, and poor-prognosis clusters were captured in the +TE0-6m outcome.

This study provides an initial demonstration that predictive modelling of clinically documented depressive outcomes in EIP services is feasible using routinely collected EHR data. The findings address a major unmet clinical need and mark a critical step toward integrating depression risk prediction into psychiatric practice.