STRUCTURAL NEUROIMAGING UNCOVERS DIFFERENT CARDIOMETABOLIC RISK PROFILES IN DEPRESSION: A DATA-DRIVEN APPROACH ON UK BIOBANK

Abstract Background Major depressive disorder (MDD) is a heterogeneous psychiatric condition associated with a high risk of mortality considering both high rates of suicide [1] and of severe comorbid cardiometabolic dysfunctions [2]. Environmental variables such as adverse childhood experiences (ACEs) can furtherly interact with the disorder’ s prognosis [3-4]. However, not all MDD patients exhibit the same illness trajectories, thus leading to inconsistent findings between clinical outcomes and the underlying neurobiology [5]. Therefore, stratifying MDD based on neurobiological data could prompt the reveal of distinct disorder subtypes and enable personalized treatment approaches. Aims & Objectives We employed an unsupervised machine learning approach to stratify MDD patients into biologically-driven clusters using multimodal structural neuroimaging data, which were then characterized for different cardiometabolic profiles, suicide attempts (SA), and ACEs. Method Cortical thickness, cortical and subcortical grey matter volumes, surface area, and white matter fractional anisotropy (FA) were extracted from 3636 individuals affected by MDD among the UK Biobank cohort. A stability-based relative clustering validation approach [6] was applied to these data to identify neurobiologically homogeneous patients’ subgroups. K-means was implemented for clustering, and support vector machine was trained to predict clusters’ labels through a 2-fold cross-validation iterated 100 times. The best clustering solution among 2 to 5 clusters was identified by minimizing the prediction error. A subsample of 727 individuals (20%) was used to optimize the model’ s hyperparameters through a parallel grid search, while the optimized pipeline was applied on the remaining 2909 (80%). Confounding effects of age, sex, and volumetric scaling factor were controlled in cross-validation. Mann-Whitney U and Chi-squared tests were used to explore differences and associations between the identified clusters and age, sex, structural neuroimaging features, BMI, presence of any cardiovascular disease, diagnosis of diabetes, history of SA, childhood emotional, physical, and sexual abuse, childhood emotional and physical neglect. Results The optimized stability-based clustering pipeline identified 2 clusters (N cluster1=1361, N cluster2=1548) with a normalized stability of 0.084, meaning that the input features could discriminate clusters’ labels with a 91.6% of accuracy. The second cluster was significantly associated with a higher prevalence of cardiovascular diseases (χ2=4.846, p=0.027), comorbid diabetes (χ2=13.245, p<0.001), and childhood physical abuse (χ2=4.101, p=0.043). Remarkably, the same cluster also showed lower cortical thickness (p<0.001), lower cortical and subcortical volumes (p<0.001), and higher surface areas (p<0.001) in most regions of interest, as well as lower FA (p<0.001) in all white matter tracts. No significant associations and differences for SA, age, sex, and BMI were found. Discussion & Conclusion Our data-driven approach revealed that structural neuroimaging can uncover depression subtypes linked to different cardiometabolic risk profiles, which were also associated with a childhood history of physical abuse. These results can hence contribute to the understanding of MDD pathophysiology and possibly help to detect vulnerable individuals which may benefit from tailored treatments in a precision psychiatry framework. References [1]DONG, M., ZENG, L. N., LU, L., LI, X. H., UNGVARI, G. S., NG, C. H., CHOW, I. H. I., ZHANG, L., ZHOU, Y. &XIANG, Y. T. 2019. Prevalence of suicide attempt in individuals with major depressive disorder: a meta- analysis of observational surveys. Psychol Med, 49, 1691-1704. [2]HARE, D. L., TOUKHSATI, S. R., JOHANSSON, P. &JAARSMA, T. 2014. Depression and cardiovascular disease: a clinical review. Eur Heart J, 35, 1365-72. [3]SOUAMA, C., LAMERS, F., MILANESCHI, Y., VINKERS, C. H., DEFINA, S., GARVERT, L., STEIN, F., WOOFENDEN, T., BROSCH, K., DANNLOWSKI, U., GALENKAMP, H., DE GRAAF, R., JADDOE, V. W. V., LOK, A., VAN RIJN, B. B., VÖ LZKE, H., CECIL, C. A. M., FELIX, J. F., GRABE, H. J., KIRCHER, T., LEKADIR, K., HAVE, M. T., WALTON, E. &PENNINX, B. 2023. Depression, cardiometabolic disease, and their co- occurrence after childhood maltreatment: an individual participant data meta-analysis including over 200,000 participants. BMC Med, 21, 93. [4]TUNNARD, C., RANE, L. J., WOODERSON, S. C., MARKOPOULOU, K., POON, L., FEKADU, A., JURUENA, M. &CLEARE, A. J. 2014. The impact of childhood adversity on suicidality and clinical course in treatment-resistant depression. J Affect Disord, 152-154, 122-30 [5]BUCH, A. M. &LISTON, C. 2021. Dissecting diagnostic heterogeneity in depression by integrating neuroimaging and genetics. Neuropsychopharmacology, 46, 156-175. [6]LANDI, I., MANDELLI, V. &LOMBARDO, M. V. 2021. reval: A Python package to determine best clustering solutions with stability-based relative clustering validation. Patterns, 2, 100228.