AI Tool May Help Predict Psychosis Before It Occurs - Medscape

The onset of psychosis in individuals at clinical high risk (CHR) may be predicted before it occurs using a machine learning tool that classifies MRI brain scans, an international consortium of researchers reported.

Prior studies using brain MRI have revealed structural differences in the brain after the onset of psychosis. The new work demonstrates structural differences in the brains of those at high risk who have not yet experienced psychosis.

"The results suggest that when considering adolescent brain development, baseline MRI scans for CHR individuals may be helpful to identify their prognosis," researchers wrote.

Only about 30% of CHR individuals go on to develop overt psychotic symptoms. Investigators say with further refinements, their tool could allow clinicians to identify people who may benefit from early intervention by combining biological markers with subclinical signs such as changes in thinking, behavior, and emotions.

"Although the machine learning-aided classification is not ready for use in clinical settings immediately, the findings suggest that when we properly consider machine- and protocol-related bias, as well as nonlinear developmental trajectory during adolescence, we can build a machine learning classifier for the brain images scanned before the onset of psychosis," study author Shinsuke Koike, PhD, with the University of Tokyo, Tokyo, Japan, told Medscape Medical News.

The study was published online on February 9 in Molecular Psychiatry.

Building a Classifier

Earlier studies that sought to use MRI data to predict outcomes in CHR individuals were limited by smaller and less diverse sample sizes, the researchers noted.

The international ENIGMA consortium includes investigators from 21 institutions in 15 countries who used T1-weighted structural brain MRI scans from 1165 adolescents and young adults at CHR and 1029 similarly aged healthy controls to build their classifier.

The CHR group included 144 individuals who went on to develop overt psychosis, 793 who did not develop psychosis, and 228 with uncertain psychosis follow-up status.

In the training set, the classifier was 85% accurate at differentiating between CHR individuals who developed overt psychotic symptoms and healthy controls. In the validation set, it was 73% accurate.

The researchers noted that the brain changes seen in CHR individuals who went on to develop psychosis are consistent with those observed in other studies, with the superior temporal, insular, and superior frontal areas contributing most to differentiating those who developed psychosis from healthy controls.

"These areas could be informative in improving understanding of pathophysiology linked to psychosis onset," researchers wrote.

'Major Step Forward'

Reached for comment, Cameron S. Carter, MD, professor and chair, Department of Psychiatry and Human Behavior School of Medicine, University of California at Irvine, California, said this "very well executed study utilizes a unique multisite dataset and a sophisticated machine learning method to address a very important question: Can we use MRI brain imaging to predict who develops psychosis among those individuals who are at increased clinical risk for the illness?"

The findings that a structural MRI measure can distinguish high-risk converters from healthy controls are "an important one and among other things can provide useful indications as to what changes in the brain are associated with developing psychosis," said Carter, who wasn't involved in the study.

It's important to note that the classification model was not trained to distinguish converters from non-converters "because the authors surmise that these differences will be subtle to detect," Carter told Medscape Medical News. "Future studies, however, will need to take on this challenge, which will most likely require new, larger samples of harmonized clinical and MRI data if the approach is to have clinical impact."

This is important for both groups, he added.

"Positive prediction of conversion will permit more intensive early intervention in these individuals," he said. "Predicting those who will not convert serves these individuals and families by reducing the stress of being at risk for psychosis and will ensure that they receive appropriate care for their nonpsychosis-related symptoms."

Also providing perspective on this research, Donald C. Goff, MD, professor of psychiatry and vice chair for psychiatry research at NYU Grossman School of Medicine in New York City, noted that treatments for schizophrenia are "pretty good, but there continues to be quite a few people who don't respond, and so we really are pushing toward earlier and earlier identification of people at risk, but also trying to understand what's going on in terms of the biology of the brain early on; what causes the onset of the illness."

This new work is a "major step forward as it relates to attempts to identify psychosis, and particularly schizophrenia, at the earliest possible stage and hopefully intervene either to prevent its development or at least halt the progression of the process," Goff told Medscape Medical News.

The accuracy in the replication sample was around 70%, "which is good, but it's not quite good enough for clinical use," he noted. "Ultimately, we'll probably use many different kinds of measures to make it more accurate, but this is an important step toward clinical prediction."

This research was supported by grants from AMED, JST Moonshot R&D, JSPS KAKENHI, Takeda Science Foundation, SENSHIN Medical Research Foundation, and the International Research Center for Neurointelligence at the University of Tokyo. Koike and Goff had no relevant disclosures. Carter is editor in chief of Biological Psychiatry: Cognitive Neuroscience and Neuroimaging.

Adblock test (Why?)