Increasing conceptual clarity and confounders identification: a pragmatic way to enhance prognostic precision in ENIGMA clinical high risk for psychosis (CHR-P)

Abstract

Zhu and colleagues [1] utilized structural magnetic resonance imaging data from the ENIGMA Clinical High-Risk for Psychosis (CHR-P) Working Group cohort (based on 21 sites) to assess the ability of machine learning to predict psychosis. The primary outcome, transiton to psychosis, occurred in 144 out of 1165 CHR-P individuals (12.36%) and the study examined whether neuroimaging data processed through machine learning could discriminate between three CHR-P subgroups (transitioned, not transitioned, unknown outcome) and healthy controls.

The classifier achieved an accuracy of 85% on the training dataset and 73% on the independent confirmatory dataset. CHR-P individuals who did not transition to psychosis were more likely to be classified as healthy controls compared to those who developed psychosis (classification rate to healthy controls: CHR-P transitioned, 30%; CHR-P not transitioned, 73%; CHR-P unknown outcome, 80%).