Joint detection of psychosis or bipolar disorder risk in clinical practice in the UK:development and validation of a clinical prediction model

<b>Background</b><br/>Efficient detection of individuals at risk of developing psychotic disorders (PSY) or bipolar disorders (BD) is the first critical step to improving mental health outcomes in young people. A novel, transdiagnostic approach to jointly detect individuals at risk for either PSY or BD would maximise the impact of prevention. The aim of the current study is to develop and validate an individualised prediction model to detect the risk of developing PSY or BD in the UK.<br/><br/><b>Methods</b><br/>This RECORD- and TRIPOD+AI-compliant study describes the development and validation of a clinical prediction model to estimate the risk of developing PSY or BD using data from patients with an index diagnosis of a non-organic, non-PSY and non-BD mental disorder recorded in electronic health records from South London and Maudsley (SLaM in the UK) secondary mental healthcare between 1st January 2008 and 10th August 2021. Exclusion criteria included receiving long-acting injectables or clozapine before BD/PSY diagnosis, no recorded contact with SLaM services after index date and an index date falling within the period 01/01/2008-30/06/2008. A LASSO-regularised Cox proportional hazards model was developed to estimate the 6-year risk of developing PSY or BD, incorporating sociodemographic and clinical predictors at index date (n=5), whereas medication (n=4), hospitalisation (n=2) and natural language processing-derived symptom and substance use (n=66) predictors were derived using a 6-month look-back period. Model performance was assessed using internal–external validation, sequentially leaving one borough out for testing and averaging performance across all boroughs. The final model was fit with data across all the boroughs. Performance was assessed via discrimination (C-index), calibration (calibration slope, calibration-in-the-large) and potential clinical utility (decision curve analysis) during internal-external cross-validation. Individuals with lived experience of BD or PSY were not involved in the research or writing process.<br/><br/><b>Findings</b><br/>In total, data from 127,868 patients were included. The sample was composed of 64,980 (50.8%) male, 62,711 (49.0%) female and 89 (0.1%) other gender. For self-assigned ethnicity, it had 71,390 (55.8%) white, 18,025 (14.1%) black, 7,257 (5.7%) other, 6,270 (4.9%) Asian and 5,022 (3.9%) mixed. The mean age was 33.4 years (SD: 18.8, IQR: 17.9-44.9). The cumulative risk incidence of PSY or BD was 0·0827 (95%CI:0·0784-0·0870) within six years (mean follow-up 622 days [SD:687]). The model showed the following performance in internal-external validation: C-index=0·78 (95%CI: 0·78-0·81); calibration slope=1·02, SD:0·14; calibration-in-the-large=0·06, SD:0·02. Decision curve analysis showed that use of the model would detect three additional PSY or BD cases early per 100 patients screened compared to default assessment strategies.<br/><br/><b>Interpretations</b><br/>This study shows that the transdiagnostic clinical prediction model can identify patients at risk of developing PSY or BD and displayed excellent performance. Such a novel approach would enable systematic early detection of young people at risk of PSY or BD, advancing preventive care in real-world clinical practice.