The signs of mental health conditions, such as schizophrenia and bipolar disorder, tend to start in teenage years or early adulthood. If we could identify the individuals who are at risk for developing these conditions, we could provide treatment and support a lot earlier than we do now. For example, in Scotland it takes on average 10 years to receive a diagnosis of bipolar disorder (Bipolar Scotland, 2024), and who knows how much additional damage is done to a person while they wait for the correct diagnosis and treatment?
The difficulty is: what things predict someone’s risk of developing a mental health condition? This is a complicated question, as so many things have been found to influence this, such as our genes, our environment, the events we experience etc (NIMH, 2025).
In the study “Joint detection of risk for psychotic disorders or bipolar disorders in clinical practice”, the team have attempted to use electronic health care records of individuals who had received mental health care (e.g., with Community Mental Health Teams or inpatient hospital care) to create statistical models that can identify ‘predictors’ for an increased risk of bipolar or psychotic disorders. They then used these predictors to see if they could identify individuals with this increased risk better than the assessment tools currently in use.
There are many factors that influence someone’s risk of mental health conditions. Could they be brought together to help us identify those at risk?
Methods
The study used data from the South London and Maudsley NHS Trust to screen the mental health records of over 1 million residents (all records were anonymous). Any resident who did not have a diagnosis of a psychotic or bipolar disorder, between 2008 and 2021, were included in the study, people with brain injuries/structural abnormalities were excluded. They used a mix of selected features, such as age and medication, features identified by Natural Language Processing (a type of AI used to find similarities in free-text) and a variety of statistics to develop a prediction model. The authors followed relevant guidelines to apply the various statistical tests to create their models. These models were trained using data from 4 of the boroughs in the NHS trust, and then tested on the remaining boroughs to see if it actually worked.
Results
The entire study included over 127,000 people, with a good balance of men and women. In this group, they found that 3,150 people were diagnosed with a psychotic disorder or bipolar disorder in a 6-year period. All of the identified “predictors” were first included in the model, but several of these predictors were dropped as they had little to no statistical influence on psychotic/bipolar disorder diagnosis.
Using a final model of 28 different predictors, the authors found it performed well across all the boroughs, where the model predicted correctly about 80% of the time. The authors also used a decision curve to establish how harmful it would be for the model to give a wrong answer, also known as false positives. From this they concluded that it was more harmful to not use the model’s prediction than it was for someone to have any unnecessary assessment.
The decision curve also demonstrated that using the model would identify more individuals with psychotic/bipolar disorders compared to the usual assessment methods.
The authors also looked at different ethnic backgrounds and found that the model didn’t show any significant differences when applied to specific backgrounds.
For anyone worried about the use of AI (the Natural Language Processing used in this study), the authors also tried excluding any predictors identified through this method and found it was still effective. So being able to use a simpler model may increase its accessibility in the absence of AI expertise.
Although there are overlaps in psychotic and bipolar disorders, there are differences between them. Using a combined model to look at both of these together might mean predictors and people at high risk are missed. Thankfully, the authors also looked at psychotic disorders and bipolar disorder separately. In doing so, there appeared to be no significant difference in the effectiveness of the model, meaning a combined model can be used for identifying those at risk for psychotic disorders and bipolar disorders together.
Conclusions
In conclusion, the authors were able to use this model to identify individuals at an increased risk of bipolar disorder and psychotic disorders. Using mental health care records in this way gives the potential to flag individuals at risk much earlier.
In this study, the authors created a model for predicting risk, flagged people at high risk and showed, with good accuracy, that those people did have a diagnosis of a psychotic or bipolar disorder.
A next step would be to apply this model to see if it can indeed identify completely new cases of these disorders in a clinical setting, but only time will tell.
Using mental health care records in this way gives the potential to flag individuals at risk much earlier.
Strengths and limitations
Strengths
- The study uses a large sample of real-world clinical data, rather than a specifically curated dataset, so it is less likely to suffer from sampling bias.
- The authors use clearly described and well explored methods to look at real-world data.
- The authors were able to demonstrate, not just the correlation of the model with cases, but also clearly show the potential real-world impact of identifying new cases.
Limitations
- The study used data from a specific area of the UK, so this may not transfer completely to other areas of the UK or beyond. However, the authors do report that the boroughs included are diverse in their backgrounds and likely reflective of the UK population.
- These models are only applicable to secondary mental health care, so there may be important aspects from family history, GP care and other settings that are being missed
- The use of Natural Language Processing has its own limitations. For example, language detection is not completely accurate.
This study focused on specific areas of London, how would it work in other areas?
Implications for practice
Despite the limitations mentioned above, this is very exciting work. Being able to use health records to predict who will be most at risk for developing certain conditions could allow for much earlier treatment and better outcomes for those individuals. If we could show that this approach works in other areas of the UK, it could be put into regular clinical practice.
How this would work on a larger scale in the clinical setting is unclear, but it is promising. We know, for example, that secondary mental health services are already under extreme pressure. Would the early identification of those most at risk help or hinder their efficiency? Helpfully, the authors give a breakdown of how they feel integration could be achieved.
It is important to note these models would not replace the assessments we already use, but could be a supportive way to identify those individuals who we should be assessing sooner.
This approach could usefully supplement current clinical practice
Statement of interests
Amy Ferguson declares no conflicts of interest.
Edited by
Simon Bradstreet.
Links
Primary paper
Maite Arribas, Andrea de Micheli, Kamil Krakowski, Daniel Stahl, Christoph Correll, Allan Young, Ole Andreassen, Eduard Vieta, Celso Arango, Philip McGuire, Dominic Oliver, Paolo Fusar-Poli (2026) Joint detection of risk for psychotic disorders or bipolar disorders in clinical practice in the UK: development and validation of a clinical prediction model. The Lancet Psychiatry. Vol 13 (1) 14 – 23
Other references
Bipolar Scotland (2024) Bipolar: The Essential Guide
National Institute of Mental Health (2025) Bipolar Disorder
