Summary: Researchers have developed a machine-learning classifier that uses structural MRI scans to identify people at high risk of developing psychosis. Trained on multisite data, the model reached 85% accuracy on its training set and 73% accuracy on independent test data. This approach could enable earlier, more targeted intervention and improve outcomes for individuals at clinical high risk.
An international research consortium led in part by investigators at the University of Tokyo analyzed T1-weighted structural MRI data from more than 2,000 participants across 21 sites to detect subtle brain differences that precede overt psychotic episodes. About half of the participants were identified clinically as being at high risk for psychosis. By leveraging harmonized measures of cortical thickness, surface area and subcortical volumes, the team trained a machine-learning model to distinguish those who later developed psychosis from healthy controls.
Key facts
- The classifier discriminates between healthy controls and individuals who later developed psychosis using structural MRI features.
- On the training dataset the model achieved 85% accuracy; on a held-out independent dataset it achieved 73% accuracy.
- Early MRI-based identification could support clinical decision-making and earlier interventions, but further validation is needed across different scanners and clinical settings.
Source: University of Tokyo
Predicting psychosis with MRI and machine learning
Psychotic episodes—characterized by hallucinations, delusions, or disorganized thinking—can have many triggers, including illness, injury, substance use, medication effects, trauma or genetic vulnerability. Although many people recover with appropriate treatment, outcomes are generally better when intervention begins early. Detecting which individuals at clinical high risk will go on to develop psychosis remains a major clinical challenge: historically, only around 20–30% of clinical high-risk (CHR) people transition to overt psychosis, while the majority do not.
To address this, the consortium pooled data from 1,165 adolescents and young adults identified as CHR (including 144 who later developed psychosis) and 1,029 healthy controls collected across 21 sites in 15 countries. The team harmonized imaging measures to correct for site and scanner differences and adjusted for non-linear age and sex effects. The resulting dataset allowed training of a classifier focused on structural brain features that best separate CHR individuals who developed psychosis from healthy controls.
Key regional contributors to classification included cortical surface area measures in the right superior frontal gyrus, right superior temporal gyrus and bilateral insular cortices. During model development, data from 20 sites were used for training and internal validation, while remaining samples were reserved for independent external validation. In training, accuracy against healthy controls reached 85%; in the independent confirmatory dataset accuracy was 73%.
When the classifier was applied to CHR participants who did not develop psychosis or whose follow-up status was uncertain, those groups were more often classified as similar to healthy controls (classification-to-HC rates: CHR-PS+, 30%; CHR-PS-, 73%; CHR-UNK, 80%). These differences underscore both the potential and the current limitations of MRI-based prediction: the tool shows promise but will require additional refinement and prospective testing before routine clinical use.
Associate Professor Shinsuke Koike of the University of Tokyo emphasized challenges in multisite MRI research: scanner models and acquisition parameters vary like different cameras producing varied images of the same scene. The team addressed this with harmonization techniques but acknowledges the need for classifiers that generalize robustly to new scanners and clinical populations. A national project in Japan (Brain/MINDS Beyond) aims to support this next step by developing models that handle greater variability in incoming data.
In practical terms, incorporating baseline MRI into assessments for people already identified as clinically at high risk may improve prognostic accuracy and guide earlier, targeted interventions—provided further prospective validation confirms clinical utility. Future studies should test the classifier in routine clinical settings, evaluate integration with clinical and cognitive measures, and refine approaches to ensure reproducible performance across sites and equipment.
Funding: This research received support from AMED (Grant Numbers JP18dm0307001, JP18dm0307004, JP19dm0207069), JST Moonshot R&D (JPMJMS2021), JSPS KAKENHI (JP23H03877, JP21H02851), Takeda Science Foundation, SENSHIN Medical Research Foundation, and the International Research Center for Neurointelligence (WPI-IRCN), University of Tokyo.
About this psychosis research news
Author: Joseph Krisher
Source: University of Tokyo
Contact: Joseph Krisher, University of Tokyo
Image credit: Neuroscience News
Original research: Open access. “Using Brain Structural Neuroimaging Measures to Predict Psychosis Onset for Individuals at Clinical High-Risk” by Shinsuke Koike et al., published in Molecular Psychiatry.
Abstract
Using Brain Structural Neuroimaging Measures to Predict Psychosis Onset for Individuals at Clinical High-Risk
Machine learning applied to structural MRI (sMRI) may assist disease classification, but predicting psychosis onset has been challenging. The study constructed a model distinguishing individuals at clinical high risk who later developed psychosis (CHR-PS+) from healthy controls (HCs), using T1-weighted sMRI measures harmonized across sites. The dataset comprised 1,165 CHR participants (CHR-PS+ n = 144; CHR-PS- n = 793; CHR-UNK n = 228) and 1,029 HCs from 21 sites. After harmonization and correction for age and sex effects, CHR-PS+ (n = 120) and HC (n = 799) data from 20 sites were used to train the classifier. External validation sets were used to evaluate performance. The classifier achieved 85% accuracy on the training data and 73% accuracy on an independent confirmatory dataset. Regional cortical surface area measures—particularly right superior frontal, right superior temporal and bilateral insular cortices—contributed strongly to classification. CHR-PS- and CHR-UNK individuals were more frequently classified as similar to HCs. The results support the potential value of baseline MRI in prognostic assessment of CHR individuals, while highlighting the need for prospective studies to determine clinical utility.