Researchers create machine learning-based classifier that could aid early diagnosis of psychosis

The onset of psychosis can be predicted before it happens, using a machine learning tool that can classify brain MRI scans between healthy people and those at risk of a psychotic episode .

An international consortium including researchers from the University of Tokyo used the classifier to compare scans of more than 2,000 people from 21 global sites. Around half of the participants had been identified as being at high clinical risk of developing psychosis.

Using the training data, the classifier was 85% accurate in differentiating between people who were not at risk and those who later showed overt psychotic symptoms. Using new data, it was 73% accurate. The work was published in Molecular Psychiatry.

This tool could be useful in future clinical settings because although most people with psychosis make a full recovery, earlier intervention generally leads to better outcomes with less negative impact on people’s lives.

Anyone can experience a psychotic episode, which usually involves delusions, hallucinations, or disorganized thinking. There is no single cause, but it can be triggered by illness or injury, trauma, drug or alcohol use, medication or genetic predisposition.

Although it can be scary or disturbing, psychosis is treatable and most people recover. Because the most common age for a first episode is during adolescence or early adulthood, when the brain and body are going through many changes, it can be difficult to identify young people who need help.

“At most, only 30% of clinically high-risk individuals later show overt psychotic symptoms, while the remaining 70% do not,” explained Associate Professor Shinsuke Koike of the Graduate School of Arts and Sciences at the University of Tokyo.

“Therefore, clinicians need help identifying those who will continue to exhibit psychotic symptoms using not only subclinical signs, such as changes in thinking, behavior and emotions, but also certain biological markers.”

The consortium of researchers worked together to create a machine learning tool that uses brain MRI scans to identify people at risk for psychosis before it begins.

Previous studies using brain MRI have suggested that structural differences occur in the brain after the onset of psychosis. However, this is the first time that differences in the brains of people who are at very high risk but have not yet suffered from psychosis have been identified.

The team from 21 different institutions in 15 different countries brought together a large and diverse group of adolescents and young adults.

According to Koike, MRI research on psychotic disorders can be difficult because variations in brain development and MRI machines make it difficult to obtain highly accurate and comparable results. Additionally, among young people, it can be difficult to differentiate between changes that occur due to typical development and those that are due to mental illness.

“Different MRI models have different parameters that also influence the results,” Koike explained. “Just like with cameras, various instruments and shooting specifications create different images of the same scene, in this case the participant’s brain. However, we were able to correct for these differences and create a classifier well suited to predicting the onset of psychosis.”

Participants were divided into three groups of people at high clinical risk: those who later developed psychosis; those who have not developed psychosis; and people with uncertain follow-up status (1,165 people in total for the three groups) and a fourth group of healthy controls for comparison (1,029 people).

Using the scans, the researchers trained a machine learning algorithm to identify patterns in the participants’ brain anatomy. From these four groups, the researchers used the algorithm to classify participants into two main groups of interest: healthy controls and those at high risk who later developed overt psychotic symptoms.

During training, the tool was 85% accurate in classifying results, while in final testing using new data, it was 73% accurate in predicting which participants were at high risk of developing psychosis . Based on the results, the team considers that providing brain MRI scans to individuals identified as being at clinically high risk could be useful in predicting future onset of psychosis.

“We still need to test whether the classifier will work well for new datasets. Since some of the software we used is best suited for a fixed dataset, we need to create a classifier that can robustly classify MRIs new sites and machines,” a challenge that a national brain science project in Japan, called Brain/MINDS Beyond, is addressing,” Koike said.

“If we succeed, we will be able to create more robust classifiers for new datasets, which can then be applied to real-world, routine clinical settings.”