Highlights
- A novel transdiagnostic model achieved a C-index of 0.80 in predicting the 6-year risk of psychotic or bipolar disorders using real-world clinical data.
- The study utilized data from 127,868 patients in the UK, incorporating 77 predictors including sociodemographics, medications, and NLP-derived symptoms.
- Decision curve analysis indicates that implementing the model could detect three additional cases of psychosis or bipolar disorder per 100 patients screened compared to standard care.
- The model showed excellent calibration, suggesting its reliability for individualized risk assessment in secondary mental health settings.
Background: The Imperative for Early Detection in Mental Health
Early intervention is the cornerstone of modern psychiatry, particularly concerning severe mental illnesses like psychotic disorders and bipolar disorders. These conditions often emerge in late adolescence or early adulthood, leading to significant long-term disability, reduced quality of life, and high economic costs if left untreated. Traditionally, clinical pathways for detecting risk have been siloed, focusing on either psychosis or mood disorders independently. However, recent evidence suggests a significant overlap in early-stage symptoms and shared genetic and environmental risk factors between these categories.
There is a critical unmet need for a transdiagnostic approach that can jointly identify individuals at risk for both psychotic and bipolar disorders. Such a tool would allow clinicians to move beyond reactive care toward a proactive, preventive model. The development of clinical prediction models (CPMs) using electronic health records (EHR) offers a promising avenue to achieve this, providing clinicians with objective, data-driven estimates of risk tailored to the individual patient.
Study Design: Leveraging Real-World Evidence and AI
This study, compliant with RECORD and TRIPOD+AI statements, aimed to develop and validate a CPM to estimate the 6-year risk of developing psychotic or bipolar disorders. The researchers utilized data from the South London and Maudsley (SLaM) NHS Foundation Trust, one of the largest secondary mental health care providers in Europe. The dataset included patients of all ages who had an index diagnosis of a non-organic, non-psychotic, and non-bipolar mental disorder recorded between January 1, 2008, and August 10, 2021.
Patient Population and Exclusion Criteria
The final cohort consisted of 127,868 patients. To ensure the model predicted new-onset disorders rather than tracking existing ones, the study excluded patients who received long-acting injectable antipsychotics or clozapine before the target diagnosis. Additionally, a washout period was established to ensure data quality, and patients with no follow-up contact were excluded.
Predictors and Modeling
The researchers employed a least absolute shrinkage and selection operator-regularised (LASSO) Cox proportional hazards model. This approach is particularly effective at handling high-dimensional data while preventing overfitting. A total of 77 predictors were incorporated, derived from a 6-month look-back period from the index date:
- Sociodemographic and clinical predictors (5 variables, including age and ethnicity).
- Medication history (4 variables).
- Hospitalization history (2 variables).
- Natural Language Processing (NLP)-derived signs, symptoms, and substance use (66 variables).
The inclusion of NLP-derived data is a significant methodological strength, as it allows the model to capture nuanced clinical information often buried in unstructured clinical notes, such as specific hallucinations, mood fluctuations, or patterns of substance misuse.
Validation Strategy
To ensure the model’s robustness and generalizability, the team used internal-external validation. This involved leaving out one of the five boroughs served by SLaM for testing while training the model on the remaining four. This process was repeated for each borough, and the performance results were averaged.
Key Findings: Model Performance and Accuracy
The study cohort was diverse, with a mean age of 33.4 years. The gender distribution was nearly even (50.8% male, 49.0% female), and the ethnicity data reflected the urban population of South London (55.8% White, 14.1% Black, and 4.9% Asian). The cumulative risk of developing either a psychotic or bipolar disorder within six years was found to be 0.0827 (8.27%).
Discriminative Power and Calibration
The model demonstrated excellent performance across several key metrics:
- C-index: 0.80 (95% CI 0.78-0.81). This indicates a high ability to distinguish between patients who will develop the disorders and those who will not.
- Calibration Slope: 1.02 (SD 0.14). A slope close to 1 suggests that the predicted probabilities align well with the observed outcomes.
- Calibration-in-the-large: 0.06 (SD 0.02). This metric confirms that the overall risk predicted by the model is close to the actual risk in the population.
These results suggest that the model is not only accurate in its rankings but also provides reliable absolute risk percentages that clinicians can use to inform treatment decisions.
Clinical Utility: Beyond Statistical Significance
While statistical metrics like the C-index are vital, the ultimate value of a prediction model lies in its clinical utility. The researchers performed a Decision Curve Analysis (DCA) to evaluate whether using the model would lead to better clinical decisions than default strategies (such as assuming everyone is at risk or no one is at risk).
The DCA revealed that the model provides a significant net benefit across a wide range of threshold probabilities. Specifically, the researchers estimated that the model could detect three additional cases of psychotic or bipolar disorders early for every 100 patients screened. In a large healthcare system, this translates to hundreds of individuals receiving earlier intervention, potentially altering their long-term disease trajectory.
Expert Commentary: Strengths, Limitations, and Future Directions
This study represents a significant advancement in the application of AI and machine learning to mental health. The use of a transdiagnostic framework recognizes the clinical reality that early-stage psychiatric symptoms are often non-specific. By moving away from rigid diagnostic categories during the risk-assessment phase, the model reflects a more biological and clinical truth.
Strengths
The primary strength of the study is its massive scale and its use of real-world EHR data. Unlike tightly controlled clinical trials, this model was built on the messy, complex data found in daily practice, making it more likely to perform well in actual clinical settings. The integration of NLP to extract clinical signs from free-text notes is also a major technological leap, capturing the “clinical intuition” often documented by nurses and doctors but ignored by traditional database queries.
Limitations
Despite the impressive results, several limitations must be considered. First, the model was developed and validated within a single (albeit large) NHS Trust. External validation in different geographic regions or different healthcare systems (e.g., private insurance-based systems) is necessary to ensure generalizability. Second, as noted by the authors, individuals with lived experience were not involved in the design or writing process. Future iterations of such models should incorporate patient perspectives to ensure the outcomes measured and the methods of communication are patient-centered. Finally, while the model identifies risk, it does not specify which intervention is best for which patient—a necessary next step for precision psychiatry.
Conclusion
The development of this transdiagnostic clinical prediction model marks a pivotal shift toward systematic early detection in psychiatry. With a C-index of 0.80 and clear evidence of clinical utility, the tool offers a robust method for identifying high-risk individuals in secondary care. By integrating such models into electronic health record systems, healthcare providers can transition toward a more preventive approach, potentially reducing the burden of psychotic and bipolar disorders on young people and the healthcare system at large.
Funding and References
This research was funded by the UK Medical Research Council (MR/N013700/1) and the National Institute for Health Research (NIHR) Biomedical Research Centres at South London and Maudsley NHS Foundation Trust and Oxford Health NHS Foundation Trust.
References
Arribas M, de Micheli A, Krakowski K, Stahl D, Correll CU, Young AH, Andreassen OA, Vieta E, Arango C, McGuire P, Oliver D, Fusar-Poli P. Joint detection of risk for psychotic disorders or bipolar disorders in clinical practice in the UK: development and validation of a clinical prediction model. Lancet Psychiatry. 2026 Jan;13(1):14-23. doi: 10.1016/S2215-0366(25)00307-4. Epub 2025 Nov 26. PMID: 41317739.
