Application of machine learning classification for structural brain MRI in mood disorders: Critical review from a clinical perspective

- Machine learning approach for the neuroimaging data is useful for diagnosing individual patients.
- Many of current machine learning studies did not consider clinical correlates.
- Future machine learning studies should integrate variable clinical and neurobiological factors.

Mood disorders are a highly prevalent group of mental disorders causing substantial socioeconomic burden. There are various methodological approaches for identifying the underlying mechanisms of the etiology, symptomatology, and therapeutics of mood disorders; however, neuroimaging studies have provided the most direct evidence for mood disorder neural substrates by visualizing the brains of living individuals. The prefrontal cortex, hippocampus, amygdala, thalamus, ventral striatum, and corpus callosum are associated with depression and bipolar disorder. Identifying the distinct and common contributions of these anatomical regions to depression and bipolar disorder have broadened and deepened our understanding of mood disorders. However, the extent to which neuroimaging research findings contribute to clinical practice in the real-world setting is unclear. As traditional or non-machine learning MRI studies have analyzed group-level differences, it is not possible to directly translate findings from research to clinical practice; the knowledge gained pertains to the disorder, but not to individuals. On the other hand, a machine learning approach makes it possible to provide individual-level classifications. For the past two decades, many studies have reported on the classification accuracy of machine learning-based neuroimaging studies from the perspective of diagnosis and treatment response. However, for the application of a machine learning-based brain MRI approach in real world clinical settings, several major issues should be considered. Secondary changes due to illness duration and medication, clinical subtypes and heterogeneity, comorbidities, and cost-effectiveness restrict the generalization of the current machine learning findings. Sophisticated classification of clinical and diagnostic subtypes is needed. Additionally, as the approach is inevitably limited by sample size, multi-site participation and data-sharing are needed in the future.