Neuroimaging Data Informed Mood and Psychosis Diagnosis Using an Ensemble Deep Multimodal Framework

Investigating neuroimaging data to identify brain-based markers of mental illnesses has gained significant attention. Nevertheless, these endeavors encounter challenges arising from a reliance on symptoms and self-report assessments in making an initial diagnosis. The absence of biological data to delineate nosological categories hinders the provision of additional neurobiological insights into these disorders. This study explores the use of neuroimaging to identify brain-based markers for mental illnesses, addressing the limitations of existing diagnoses. Previous research showed the potential of integrating structural neuroimaging data by treating diagnostic categories as uncertain and adjusting them to align better with biological data. Building on this, our current research incorporates multimodal neuroimaging data, combining fMRI with structural MRI, and introduces methodological advances to enhance diagnosis by creating more homogeneous categories based on MRI-derived neurobiological information. Unlike other studies that reclassify psychiatric groups purely based on biological data, our approach integrates neuroimaging and symptom-based categories using ensemble methods, deep learning, and data fusion. This strategy aims to improve symptom-based categorization by identifying biologically-based categories that help distinguish between correctly classified, challenging, and noisy samples. Our goals include identifying potential biomarkers for existing symptom-based categories, determining biologically homogeneous groups, and mitigating label noise across mood and psychosis categories. We analyzed the relationship between biological findings and existing categories, highlighting discrepancies between brain imaging features and symptom-based categories, and assessing the potential of augmenting label categories for sample heterogeneity. Notably, visualization techniques provided insights into distinct brain patterns in well-classified versus challenging samples. We used a deep convolutional framework and bagging approaches for diagnostic classification, finding that ensemble deep models outperformed individual models, and multimodal frameworks consistently surpassed unimodal approaches. In sum, this work highlights the potential of combining existing symptom-based categorization with multimodal data and advanced data-driven approaches to improve the categorization of mental illness.