What is the P Risk algorithm?
We have developed a risk algorithm, called P Risk, to help GPs detect the early warning signs of psychosis. The P Risk algorithm uses predictors which are stored in the electronic medical records of patients. A previous nested case (n=11,690) control (n=81,793) study (1) in primary care EHRs found that 12 predefined nonpsychotic warning symptoms and signs (depression, obsessive compulsive disorder, attention deficit hyperactivity disorder, mania, blunted affect, problems with cannabis, problems with cigarette smoking, sleeping problems, suicidal behaviour, bizarre behaviour, social isolation, role functioning problems), sociodemographic factors and consultation frequency were associated with later psychosis.
Using these findings we have developed and internally validated a prediction model (P Risk), using a sample of 300,000 people from 216 GP practices with linked primary care and secondary care EHRs and at least 5 years of follow up data, who had consulted their GP for any nonpsychotic mental health problem (paper submitted for publication May 2022). From this sample 830 diagnoses of psychosis were detected. P Risk discrimination (i.e. the ability to discriminate between those who later develop psychosis and those do not) performance is good (Harrell’s C statistic of 0.77). Once the final model had been identified, we calculated the predicted risk over 6 years for patients whose follow up time exceeded 6 years, or who had experienced an event within 6 years. We calculated sensitivity (i.e. ability to detect true positives), specificity (i.e. ability to detect false positives), and likelihood ratios for thresholds of risk of 0.5%, 1%, 1.5% and 2%. The sensitivity (68.9%) and specificity (70.9%) were similar and high for a 1% risk threshold, the likelihood ratio was 2.37.
External validation of P Risk in a separate EHR database showed slightly improved accuracy. P Risk overcomes the problems that GPs have in identifying patients with the early warning signs of psychosis described above by using existing primary care data to predict future risk. It uses EHR data on GP consultations for 14 proven nonpsychotic predictors for psychosis. It can therefore support individual GP diagnostic skills and is not dependent on care continuity. P Risk could be automated for use on GP IT systems in a similar way to other tools already in use, such as Q Risk for detecting cardiac risk and FRAX for osteoporosis risk.
How will P Risk work in practice?
EMIS will install the P Risk algorithm on participating GP practice computers. P Risk will calculate a risk score when a GP enters a Snomed code for a nonpsychotic mental health problem and a pop up will appear on the GP’s computer. The pop up will show a risk stratification of “low”, “medium” or “high” risk (cut points in the risk probability of P Risk at <=1% low, 2-20% medium and >20% high).
After installation GPs will receive brief and simple training from a CRN clinical research nurse about the P Risk algorithm which will take no longer than 10 minutes and will include:
- the rationale for risk stratification
- the pop up
- suggested extra questions to ask of patients whose risk strata is medium or high to aid the referral decision
- Use of the P Risk Snomed code to indicate that the P Risk algorithm has been triggered.
- the P Risk data about consultation history that will accompany a referral to secondary care for a psychosis assessment.
The GP will decide whether to refer to secondary care for a psychosis assessment depending on the P Risk score and additional questions asked.
(1) Sullivan SA, Hamilton W, Tilling K, Redaniel T, Moran P, Lewis G. Early signs and symptoms of psychosis within primary-care: a nested case-control study using electronic primary-care records. JAMA Network Open 2018.