NEUROANATOMICAL BIOTYPING BY SEMI-SUPERVISED MACHINE LEARNING FOCUSING ON HETEROGENEITY IN TREATMENT-RESISTANT DEPRESSION

Abstract Background Treatment-resistant depression (TRD) is a highly biologically heterogeneous disorder, which may be due to underlying neuroanatomical heterogeneity. However, neuroimaging studies to date have not yielded consistent results, possibly due in part to a lack of analytical approaches that account for clinicodemographic confounds. The heterogeneity through discriminative analysis (HYDRA) method (Varol et al. 2017) takes into account clinical confounding factors and enables biotyping based on neuroanatomical heterogeneity. Previous studies have reported neuroanatomical biotypes in psychiatric disorders such as schizophrenia (Chand et al. 2020; Lalousis et al. 2022). However, no study has explored neuroanatomical heterogeneity in TRD using HYDRA. Aims & Objectives This study aimed to classify the biotypes of TRD by examining the neuroanatomical signatures of the subtypes and exploring their clinicodemographic characteristics. Methods This study was approved by the ethical committee at Keio University School of Medicine and the Keio Certified Review Board. Fifty-five healthy controls (37.1±15.2 years old, 55 females [48%]) and 117 patients with TRD (44.9±12.8 years old, 50 females [43%]) provided written informed consent. They underwent structural MRI for T1- and T2-weighted images. Structural preprocessing was performed using FreeSurfer 6 for parcellation with 160 ROIs (68 bilateral cortical volumes, 68 white matter (WM) volumes, and 24 other subcortical volumes) for HYDRA clustering. Fifty healthy controls (HC) and 103 TRD patients were successfully preprocessed, and three covariates (age, sex, and total intracranial volume) were included in HYDRA for their data, identifying biotypes based on volumetric ROI metrics. Cortical volume differences in biotypes between TRD and HCs were examined using Permutation Analysis of Linear Models (FMRIB Software Library, version 6.0.5) with vertex-wise analysis (corrected p-value <0.05). WM and other subcortical volumes were compared to HCs in the ROI under false discovery rate (FDR) correction. Results HYDRA identified two optimal clusters and named them TRD1 and TRD2, respectively. Volumetric vertex- wise analysis revealed that TRD1 exhibited smaller volume than HCs in the left precentral, left medial orbitofrontal, right lateral orbitofrontal, right postcentral, and right middle temporal cortex (p <0.05). Conversely, TRD2 showed a larger volume than HCs in the left precuneus, left postcentral, right pericalcarine, and right supramarginal cortex (p <0.05). Comparing TRD1 to HC regarding WM and other subcortical volumes, only the left accumbens showed a significant volume decrease in TRD1. The comparisons of clinical data (age, sex, socioeconomic status, psychosis onset, duration of illness, and Montgomery Å sberg Depression Rating Scale scores) between TRD1 and TRD2 did not show significant differences. Discussion & Conclusion The present study revealed two distinct TRD biotypes: one with increased brain volume and the other with decreased brain volume. HYDRA's ability to biotype based on key depression-related regions, such as the precuneus cortex, orbitofrontal cortex, and nucleus accumbens, aligns with prior studies and offers insight into biologically clustered illness. References 1.Varol, E. et al. (2017) ‘HYDRA: Revealing heterogeneity of imaging and genetic patterns through a multiple max-margin discriminative analysis framework’, NeuroImage, 145(Pt B), pp. 346–364. 2.Chand, G.B. et al. (2020) ‘Two distinct neuroanatomical subtypes of schizophrenia revealed using machine learning’, Brain: a journal of neurology, 143(3), pp. 1027–1038. 3.Lalousis, P.A. et al. (2022) ‘Neurobiologically Based Stratification of Recent-Onset Depression and Psychosis: Identification of Two Distinct Transdiagnostic Phenotypes’, Biological psychiatry, 92(7), pp. 552–562.