J Korean Soc Biol Ther Psychiatry.  2023 Jun;29(2):35-42. 10.22802/jksbtp.2023.29.2.35.

Development of a Machine Learning Model for Diagnosing Schizophrenia and Bipolar Disorder Based on Diffusion Tensor Imaging: A Preliminary Study

• Lee DK
• Lee H
• Choi HJ
• Kim CE
• Kim SW
• Ryu S

• KMID: 2544129
• DOI: http://doi.org/10.22802/jksbtp.2023.29.2.35

Abstract

Objectives
This study aimed to develop a machine learning model for diagnosing schizophrenia (SZ) and bipolar disorder (BD) based on diffusion tensor imaging (DTI) data.
Methods
We used 3T-magnetic resonance imaging to examine SZ, BD, healthy control (HC) subjects (aged 20-50 years, n=65 in each group). Applying Support Vector Machine (SVM) to fractional anisotropy (FA) values, we built classification models of SZ and HC, BD and HC, and SZ and BD. Features of white matter (WM) tracts were selected through recursive feature elimination, and 5-fold cross validation was performed.
Results
The SVM models classified SZ and BD from HC with a mean accuracy of 83.5% and 75.4%, respectively. The SZ-BD classification model archived 75.0% accuracy. These classification models used FA values in 15-18 WM tracts as features, including the retrolenticular part of the internal capsule, superior corona radiata, cingulum, and superior fronto-occipital fasciculus.
Conclusions
This study presented a preliminary machine learning model to diagnose SZ and BD based on DTI data. Our findings also suggest that there might be a specific pattern of abnormalities in WM integrity that can differentiate the two psychotic disorders.

Keyword

Schizophrenia; Bipolar disorder; Diffusion tensor imaging; Machine learning; Support vector machine