Back to all papers

Machine Learning-Based Discrimination of Treatment-Resistant Schizophrenia Using Structural Brain Imaging: A Multi-Site Proof-of-Concept Study.

August 22, 2026pubmed logopapers

Authors

Torres-Carmona E,Graff-Guerrero M,Nakajima S,Iwata Y,Ueno F,Koizumi T,Honda S,Tsugawa S,Homma S,Ogyu K,Tarumi R,Song J,Deluca V,Remington G,Gerretsen P,Graff-Guerrero A

Affiliations (5)

  • Centre for Addiction and Mental Health (CAMH), Toronto, Canada; University of Toronto (UofT), Toronto, Canada.
  • Secretaria de Ciencia, Humanidades, Tecnologia e Innovación (SECIHTI), Mexico; INFOTEC Centro de Investigación e Innovación en Tecnologias de la Información y Comunicación.
  • Keio University, Tokyo, Japan.
  • Centre for Addiction and Mental Health (CAMH), Toronto, Canada.
  • Centre for Addiction and Mental Health (CAMH), Toronto, Canada; University of Toronto (UofT), Toronto, Canada. Electronic address: [email protected].

Abstract

Treatment-resistant schizophrenia (TRS) affects 20-30% of individuals with schizophrenia, with persistent symptoms, functional impairment, and reduced quality of life. Clinical identification remains dependent on sequential antipsychotic trials despite reported structural brain differences between TRS and treatment-responsive schizophrenia (TxR). This study evaluated whether structural MRI features could discriminate clinically defined TRS from TxR using machine learning. A total of 225 participants (122 TRS, 103 TxR) from multi-site studies in Canada and Japan were included. Cortical thickness, brain volume, surface area, and intrinsic curvature were derived from T1-weighted MRI using FreeSurfer, with feature engineering generating volumetric-cortical thickness interaction terms. A voting ensemble was evaluated under a primary leakage-controlled NeuroComBat harmonization and a secondary exploratory full-dataset harmonization. Performance was assessed using ROC-AUC, F1 score, precision, and recall. In the leakage-controlled analysis, the voting ensemble achieved a held-out ROC-AUC of 0.57 and macro F1 of 0.56; training cross-validation yielded ROC-AUC of 0.715. The full-dataset harmonization analysis yielded ROC-AUC of 0.60 and macro F1 of 0.62, interpreted cautiously due to leakage. Temporal-occipital cortical thickness and choroid plexus volume interaction terms contributed most to model performance. Structural MRI features may support cross-sectional discrimination of TRS from TxR. The divergence between harmonization strategies highlights the importance of leakage-aware preprocessing in neuroimaging. Validation in independent cohorts is required to establish whether these features contribute to earlier identification of treatment resistance.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAI Slice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.