AI-Based Prediction Tool Flags Psychotic and Bipolar Disorder Risk
Key Clinical Summary
- A transdiagnostic prediction model estimated 6-year risk of psychotic or bipolar disorders using UK electronic health records.
- The model demonstrated strong discrimination (C-index 0.80) and good calibration across data from multiple London boroughs.
- Decision curve analysis suggested 3 additional early detections per 100 patients screened compared with standard assessments.
A large UK study published in The Lancet Psychiatry reports that an artificial intelligence (AI)–enabled clinical prediction model may accurately identify individuals at risk of developing psychotic disorders or bipolar disorders up to 6 years in advance. Using electronic health records from more than 127,000 patients receiving secondary mental health care in South London, the model demonstrated strong discrimination and calibration, outperforming standard assessment strategies in early case detection.
Study Findings
The study analyzed data from 127,868 patients receiving secondary mental health care at South London and Maudsley (SLaM) NHS Foundation Trust between 2008 and 2021. All patients had an index diagnosis of a non-organic, non-psychotic, and non-bipolar mental disorder at baseline.
Researchers developed a least absolute shrinkage and selection operator-regularized (LASSO) Cox proportional hazards model to estimate 6-year risk for psychotic or bipolar disorders. Predictors included sociodemographic and clinical variables at index date, medication exposure, hospitalization history, and 66 natural language processing–derived indicators of symptoms and substance use, assessed over a 6-month look-back period.
During follow-up, the cumulative 6-year incidence of psychotic or bipolar disorders was 8.27% (95% CI, 7.84–8.70). Internal–external validation across 5 London boroughs showed consistent performance, with a pooled C-index of 0.80 (95% CI, 0.78–0.81). Calibration metrics were robust, including a calibration slope of 1.02 and minimal calibration-in-the-large error.
Decision curve analysis suggested that implementing the model in clinical practice could identify 3 additional cases per 100 patients screened compared with default clinical assessment strategies. The final model was trained on the full dataset after validation.
Clinical Implications
The findings highlight the potential of transdiagnostic, data-driven tools to support earlier detection of severe mental illness in real-world settings. Unlike disorder-specific risk tools, this approach jointly assesses risk for psychotic and bipolar disorders, reflecting overlapping early clinical presentations.
The model leverages routinely collected electronic health records, increasing feasibility for integration into mental health services. Improved early detection could allow targeted monitoring, timely referrals, and preventive interventions, particularly for younger patients who often present with nonspecific symptoms.
However, the authors note external validation beyond the SLaM sample is still needed. Implementation would also require governance around algorithmic decision support and clinician oversight.
Expert Commentary
“Given the crucial need to enhance early detection strategies beyond psychotic disorders, this study provides substantial advancement with strong clinical implications for preventive care,” wrote Maite Arribas, PhD, Early Psychosis: Interventions and Clinical-Detection (EPIC) Lab, Department of Psychosis Studies, Institute of Psychiatry, Psychology, and Neuroscience, King's College London, London, UK, and coauthors.
Although developed using UK National Health Service data, this large validation study highlights how electronic health record–based prediction models could support earlier identification of psychotic and bipolar disorders in routine care. With further validation in US health systems, similar tools may help clinicians stratify risk, guide monitoring, and advance preventive mental health strategies across diverse care settings.
