Development and validation of a predictive model for high-intensity mental health service use using electronic health record data

AIMS AND METHOD: This study aimed to develop and evaluate a predictive model using electronic health record (EHR) data from a large south London mental health service, in order to identify patients 3 months following first referral who are at risk of subsequent high-intensity service use over the subsequent 12 months. Early identification of such patients may support proactive and personalised care planning, reducing the need for high-cost episodes of care. Predictive models were developed using information from 18 869 patients newly referred between 2007 and 2011. High-intensity use was defined as the top 10% of estimated mental healthcare expenditure. The model was developed using demographic, clinical and service use variables, and was validated on data from the periods 2012-2017 and 2018-2023. RESULTS: A logistic regression model achieved an area under the receiver operating characteristic (AUROC) of 0.79 in development (sensitivity 0.82, specificity 0.54), with robust performance in validation sets (AUROC 0.81, 0.83, respectively). Key predictors included first 3 months service use, schizophrenia or eating disorder diagnoses and living alone. Natural language processing-derived features did not improve performance. CLINICAL IMPLICATIONS: Routine EHR data performed well in predicting the risk of high-cost care, potentially enabling targeted interventions and more efficient resource allocation.