Testing a Computer-Generated Model to Predict High-Risk Youths' Transition to Psychosis

Testing a Computer-Generated Model to Predict High-Risk Youths' Transition to Psychosis

Posted: May 27, 2021
Testing a Computer-Generated Model to Predict High-Risk Youths' Transition to Psychosis

Story highlights

Computer-based machine-learning methods were used to develop tools to predict which high-risk youths would "covert" to full-blown psychosis by the end of adolescence as well as which high-risk youths would not develop psychosis by age 18.

 

An international research team has demonstrated a new computer-based model designed to predict whether individual young people at "clinical high risk" (CHR) for developing psychosis will in fact develop that illness. The results appear in the journal JAMA Psychiatry.

"CHR" status pertains to young people who are at elevated genetic risk, usually due to family history, or who have had preliminary (and often mild) symptoms of psychosis that sometimes precede a "first episode of psychosis," or FEP.

Only a fraction of youths at clinical high risk actually "convert" to psychosis—an event, when it happens, that usually occurs between the late teens and early 30s. An FEP, in turn, often marks the onset of schizophrenia, the most prevalent psychotic disorder, or other psychiatric disorders such as bipolar disorder which can involve psychotic symptoms.

The 3-year transition rate of conversion to first-episode psychosis among high-risk youths has been estimated in past research at between 16% and 35%. That's a large range, and it is not specific, meaning that "observational" methods alone still cannot provide insight about the chances of conversion in specific individuals. Being able to make such predictions is thought to be of great potential benefit for patients and their families, since detecting and treating as early as possible can improve patient outcomes over the long run.

Researchers led by 2007 and 2003 BBRF Independent Investigator David R. Cotter, Ph.D. of the Royal College of Surgeons, Ireland, and 2010 BBRF Distinguished Investigator Philip McGuire, Ph.D., of King's College London, set out to discover a potential set of protein biomarkers that would enable prediction of a CHR individual's likelihood of conversion to psychosis.

Past studies have made correlations between levels of various proteins and susceptibility to psychosis and well as to other illnesses including schizophrenia and depression. While these studies have been encouraging, it has not yet been possible to identify specific sets of proteins whose levels in the blood—higher or lower by varying amounts compared with levels of the same proteins in unaffected individuals—can reliably aid psychosis prediction.

The team led by Drs. Cotter and McGuire took advantage of two study cohorts. Their main sample of 344 individuals, all assessed as "CHR" when they entered the study, was drawn from the European Network of National Schizophrenia Networks Studying Gene-Environment Interactions (EU-GEI). These participants donated blood when recruited into the study, and were then followed-up over several years to see who did and did not convert to FEP.

A second cohort, not focusing on high-risk youths, was part of a UK-based "birth cohort" longitudinally studying the health of persons in the general population born in 1991 and 1992 in Avon, UK. A subset whose data were analyzed for the psychosis prediction study had not reported "psychotic experiences" (PEs) at age 12. PEs are low-level psychotic symptoms that are associated with future risk of psychotic and non-psychotic mental disorders.

Blood samples in this second cohort, as in the first, were essential, providing the basis for comparing blood protein levels in participants who did and did not convert to psychosis by age 18.

A science called proteomics is used to assess the representation of different proteins in the blood. In a critical step, the researchers fed these results into computers, which used machine-learning to develop predictive models. These models correlated data for various proteins with clinical information about each of the participants in the two study cohorts. The models were then back-tested against the clinical records indicating which of the participants did and did not "convert" to psychosis.

A subset of the EU-GEI sample of CHR youths consisting of 133 participants, average age about 22 and about half male, included 49 (37%) who developed psychosis and 84 (63%) who did not. Based on protein analysis of the blood of these 133 participants when they first entered the study, a model was developed "which demonstrated excellent performance for prediction" of conversion to psychosis, the team reported. One model correctly identified those who would go on to develop a psychotic disorder in 98% of high-risk cases, and it correctly identified those who would not convert in 81% of cases.

Another model, based on the 10 most predictive proteins, was also quite accurate. It correctly identified those who would go on to develop a psychotic disorder in 95% of high-risk cases, and it correctly identified those who would not in 88% of cases.

In the second cohort, a model based on protein levels at age 12 correctly identified those who would go on to develop PEs at age 18 in 73% of cases, and it correctly identified those who would not in 71% of cases.

The identity of the specific proteins with predictive power provides potentially powerful insights into psychotic disorder, noted David Mongan, first author of the paper. The most striking implication was the involvement of a number of these proteins in a mechanism called the complement and coagulation cascade. It is involved in the body's immune system, a significant fact tending to echo past findings associating various immune system dysfunctions with both psychosis and with schizophrenia risk. This and other associations will be followed in future research, as will the models used to identify predictive proteins, which must be tested in other, independent cohorts of young people both at high risk and not at high risk for psychosis.