Machine Learning May Improve Diagnostic Accuracy in Schizophrenia, Autism Spectrum Disorder

A recent study aimed to assess the ability of classifiers to identify individuals with schizophrenia, autism spectrum disorder, and typical development based on findings from magnetic resonance imaging.

Authors of a recent study published in Translational Psychiatry reported that machine learning combined with neuroimaging may help psychiatrists diagnose certain neuropsychiatric disorders.

The researchers compared 6 machine learning classifiers to determine their effectiveness in analyzing magnetic resonance imaging (MRI) of the brain in order to diagnose schizophrenia and autism spectrum disorder (ASD). The study included 26 participants at ultra-high risk for psychosis, 17 participants with first-episode psychosis, and 106 typically developing individuals.

Objective biomarkers are needed in neuropsychiatry, the researchers said, due to the unreliable nature of subjective diagnosis. Patient heterogeneity, clinician inconsistency (eg, differing opinions), and nomenclature inadequacy all affect diagnostic accuracy.

Structural MRI scans from all subjects were processed using the FreeSurfer image analysis suite. FreeSurfer data included cortical thickness (150 regions), surface area (150 regions), and subcortical volume (36 regions).

The researchers ran the images through all 6 classifiers (logistic regression, support vector machine, random forest, adaptive boosting, decision tree, and κ-nearest neighbor) 4 times. After processing, the instances were extracted, binarized, and correlated with their subjective clinical scores.

The researchers found the best overall results using the logistic regression classifier with the cortical thickness feature group, which produced an overall accuracy of 69%. The classifiers performed reasonably well in distinguishing between schizophrenia and ASD. Using the whole brain feature image, both support vector machine (a supervised discriminative classification method) and the κ-nearest neighbor learning algorithm had an accuracy of 75%, while logistic regression had an accuracy of 70%. Results were similar among all classifiers using subcortical images.

When distinguishing between ASD and typically developing individuals, support vector machine prevailed with a 75.8% accuracy while measuring whole brain features. To distinguish between schizophrenia and typically developing participants, logistic regression had 70.5% accuracy.

Although study limitations include a small sample size and the fact that not all participants were taking medication, the authors published “new knowledge” about machine learning classifiers for individuals with these disorders, as well as information on biomarkers that may help predict future disease.


Yassin W, Nakatani H, Zhu Y, et al. Machine-learning classification using neuroimaging data in schizophrenia, autism, ultra-high risk and first-episode psychosis. Transl Psychiatry. 2020;10(1):278. doi:10.1038/s41398-020-00965-5