Machine learning classification of patients with major depressive disorder and healthy controls using magnetic resonance imaging data

Abstract

Major depressive disorder (MDD) is a prevalent and complex psychiatric disorder affecting more than 300 million people worldwide. It is characterized by a highly heterogeneous spectrum of symptoms, including low mood and self-esteem, loss of interest, sleep disturbances, and loss/gain of appetite. Psychotherapy and pharmacotherapy are established as the first line of treatment. Nevertheless, up to a third of patients do not respond to such interventions. The development of personalized, more successful treatment strategies requires a comprehensive understanding of the pathophysiology of depression and a set of corresponding biomarkers. There is a growing interest in investigating large-scale structural and functional brain alterations via neuroimaging techniques. Machine learning techniques have gained popularity in neuroimaging due to higher performances in classifying mental disorders and elucidating the brain's structural and functional connectivity patterns. The current findings' reproducibility is restricted by either small sample sizes or inadequate consideration of demographic factors and site-related differences – site effect - in the case of large multi-site samples. Careful heed of demographic factors' role in classification performance and meticulously considering site-related differences between subjects is expected to fill this gap. In this work, I focus on discriminating depressive subjects from healthy controls using shallow machine learning algorithms based on structural pre-segmented brain features. An unprecedented initiative in terms of the number of sites included, a large dataset from ENIGMA MDD Consortium allows for extensive analysis and generalizable results. Furthermore, I investigate if higher classification performances can be achieved by analyzing high-resolution cortical vertex-wise maps and integrating volumetric characteristics, such as cortical thickness, with shape characteristics (sulcal depth and cortical curvature). Moreover, I test if deep non-linear classification algorithms, such as convolutional neural network (CNN), could potentially reveal complex patterns of brain organization, contributing to better detection of depression-related alterations compared to simple classification models. The results show that the investigated machine and deep learning models yielded accuracy close to random chance, regardless of the data modality or resolution. Furthermore, the integration of volumetric and shape characteristics did not yield high results. I detected the presence of the site effect, which was addressed by a ComBat harmonization tool. However, ComBat failed to improve the classification performance of the models. More sophisticated classification models that can incorporate both demographic and clinical information could improve classification performance based on the brain's morphology in future studies. Finally, I investigate the validity of functional subject-specific parcellation maps as a potential predictor of MDD. This proof-of-concept study uses subject-specific resting-state fMRI-based parcellations to reveal the effect of a single session of 10 Hz rTMS on healthy subjects. I applied RSFC-Snowballing and RSFC-Boundary mapping parcellation methods to obtain complementary node and boundary maps of functional brain organization, which were analyzed via Support Vector Machines (SVM) with a novel feature selection method. This approach revealed a slower and more complex response in boundaries compared to nodes located primarily in the posterior cingulate cortex and precuneus. These results highlight the potential benefits of subject-specific parcellations in future psychiatric analyses as they might capture distinct temporal and spatial differences. The development of new personalized, more successful treatment strategies requires a comprehensive understanding of the pathophysiology of depression and a set of corresponding biomarkers. The heterogeneity of large worldwide samples in terms of socio-demographic, clinical, and genetic factors requires more sophisticated analytical approaches and solutions to achieve a more general overview of the pathophysiology of depression. Recently established scientific consortiums enable the investigations of unprecedentedly large datasets, requiring more powerful big-data analytical tools.