Summary
A medRxiv preprint tested a structural MRI model for psychosis risk across four independent clinical high-risk cohorts. Its discrimination was weak, predicted risks transferred poorly, and exploratory analyses found variation by negative-symptom severity.
A structural MRI model designed to predict transition to psychosis showed weak performance when tested across four independent clinical high-risk cohorts, according to a medRxiv preprint posted on October 19, 2026. The analysis included 685 participants, of whom 106 transitioned. Overall discrimination was low, and the model’s predicted risks transferred poorly between cohorts.
The researchers re-estimated a previously published brain-imaging risk model and applied it unchanged to the four cohorts. The result highlights a challenge for clinical prediction tools: an average accuracy score can obscure whether a model works consistently across different groups of people and clinical settings.
Weak discrimination and poor transfer
Across the cohorts, the model’s area under the receiver operating characteristic curve (AUC) was 0.56. AUC measures how well a model ranks people who experience an outcome above those who do not; 0.5 represents chance-level ranking. In the two cohorts not previously used to evaluate the model, AUCs were 0.53 and 0.56—close to chance.
The model’s calibration slope was 0.17. Calibration concerns whether predicted risks correspond to the outcomes observed in a new group. The authors report that the predicted risks did not transfer, indicating that the scores were poorly aligned with outcomes across these cohorts.
These are different aspects of prediction: a model can rank people imperfectly and also give risk estimates that do not carry over reliably. Both matter if a score is intended to inform decisions about people considered at clinical high risk for psychosis.
Performance varied across participant characteristics
The researchers also used exploratory “performance landscapes” to examine how discrimination varied across participant characteristics. They found variation with negative-symptom severity, including in external cohorts. Apparent variation related to age and brain age, however, depended on individual cohorts. No global test remained significant after correction for multiple testing, so these landscape findings warrant cautious interpretation.
The study is a preprint and reports a model-validation analysis, not a test of a clinical intervention. Its results suggest that one diagnostic label can encompass groups for whom the same prediction model performs differently. For this MRI model, the weak overall discrimination and poor risk transfer indicate that its scores were not reliable across the clinical cohorts examined.