Machine Learning Framework for Depression Subtype Grouping: Integrating High-Resolution Imaging and Clinical Symptom Analysis via Correlation and Clustering

Gaurav Verma^1*Yael Jacob^1,2Laurel Morris^1,3Xin Xing⁴Ai-Ling Lin⁵James Murrough^1,2Priti Balchandani¹Bradley N Delman¹

• ¹Icahn School of Medicine at Mount Sinai, New York, United States
• ²Icahn School of Medicine at Mount Sinai Friedman Brain Institute, New York, United States
• ³University of Oxford, Oxford, United Kingdom
• ⁴University of Nebraska Omaha, Omaha, United States
• ⁵University of Missouri, Columbia, United States

Introduction Major depressive disorder (MDD) affects approximately one in six individuals over their lifetime, with many patients experiencing treatment-resistant depression, characterized by inadequate or insufficient response from at least one antidepressant treatment. Current classification strategies for depression rely primarily on clinical assessment of symptom severity, which are prone to reader bias and test-retest variability. Moreover, these symptom-based subtypes have shown limited utility in predicting treatment response. This study introduces a data-driven, non-biased classification framework that integrates clinical features with high-resolution magnetic resonance imaging (MRI)-derived features. Using canonical correlation analysis (CCA) and hierarchical clustering, the approach identifies distinct subtypes of MDD, offering a more objective and potentially predictive alternative to traditional methods. Materials & Methods Sixty-four participants with MDD currently experiencing a major depressive episode and not currently undergoing treatment completed a battery of 11 clinical symptom severity assessments and scanned with 7T T1-weighted MRI with parameters: TE/TR = 3.62/6000 ms, 320x240x240 array size with 224x168 mm2 field-of-view (FOV) for voxel dimensions of 0.7 mm3 isotropic. The images were automatically segmented using the FreeSurfer 6.0 package and 87 resulting imaging features, along with 11 clinical measures were processed through CCA to derive highly-correlated clinical-imaging phenotypes. An analysis using the Sillhouette and other methods determined an optimal number of clusters for this dataset. Participants with MDD were plotted on axes consisting of highly correlated clinical-imaging phenotypes derived from CCA and subsequently grouped through hierarchical clustering. Results CCA identified three highly correlated (r > 0.9) clinical-imaging variable pairs. The first, an anhedonia-related phenotype, showed high loadings from anhedonia severity and brainstem features. The second phenotype was associated with childhood trauma and anhedonia, with the right frontal pole as the primary imaging feature. The third phenotype linked general distress and perceived stress with the right superior temporal lobe. Hierarchical clustering along these canonical axes revealed two distinct clusters: one characterized by high childhood trauma scores and the other showing scores comparable to healthy controls. Conclusion Taken together, this study presents a novel ML framework for classifying depression using CCA and clustering.