Machine learning diagnosis of schizophrenia using structural and functional images of the brain

Abstract. This study aims to provide earlier, and faster schizophrenia diagnosis based on neurocognitive, structural, and behavioral measures using machine learning. This is because current ways of diagnosis, while accurate, delay patients ability to seek medical care before observable, lasting symptoms develop and are prone to error and discrimination. To provide diagnosis, we used Linear Support Vector Machine (linear SVM), Random Forest (RF), Multilayer Perceptron (MLP), and k-Nearest Neighbor (kNN), all trained with neurocognitive and behavioral measures combined with either structural or functional MRI data or both of 99 subjects from the OpenNeuro public dataset. 100 iterations of classification were run, and results showed a higher-than-average accuracy for all classifiers using all combinations of parameters, with a highest accuracy of 0.75 using linear SVM trained with behavioral and neurocognitive measures and fMRI data. We found correlations between structural changes in AAL3 brain regions and n-back working memory task performance, noting that the inferior parietal gyrus, right precuneus, supplementary motor area, and the central lateral thalamic nucleus have the highest feature importance. This means that future studies can select these features for further clinical examination or for machine learning diagnosis. We conclude that linear SVM provides the highest average diagnostic accuracy, and that fMRI data often leads to more accurate algorithmic decisions than sMRI data and thus should be weighed more in future studies.