Functional connectivity signatures of major depressive disorder: machine learning analysis of two multicenter neuroimaging studies

Major depressive disorder (MDD) is a prevalent and increasing global health concern, underscoring the need for improved diagnostic tools and treatments. The potential of artificial intelligence and machine learning (ML) in medicine has spurred interest in developing neuroimaging biomarkers for psychiatric disorders. While initial small-scale studies using resting-state fMRI showed promising results in identifying MDD, larger datasets are needed to validate these findings and account for the heterogeneity inherent in MDD. This study leverages two of the largest resting-state fMRI datasets for MDD to address this limitation, utilizing both traditional ML techniques (support vector machines, SVM) and advanced deep learning (DL) methods (graph convolutional neural networks, GCN) to classify patients with MDD from healthy controls and identify potential neurophysiological signatures of the disorder. Functional connectivity (FC), the statistical dependence of neurophysiological signals between brain regions, is used as a feature for classification. The study aims to overcome limitations of previous work by using large-scale data and exploring the potential of DL to uncover complex patterns associated with MDD.

Previous research using smaller fMRI datasets showed high classification accuracy for MDD using multivariate pattern recognition, but larger datasets have revealed lower accuracy, likely due to increased heterogeneity within the patient group. While univariate analysis of fMRI data has identified some consistent FC differences in MDD, it may miss more complex patterns. Recent advancements in ML and DL, especially GCNs, which are designed to handle graph-structured data like FC matrices, have enabled the exploration of multivariate patterns and offer the potential to visualize important features, addressing the 'black box' criticism of some ML methods. The current study builds upon this prior work, addressing the need for large-scale data analysis to improve the robustness and generalizability of MDD classification models.

This study utilized resting-state fMRI data from two large consortia: REST-meta-MDD (N=2338) and PsyMRI (N=1039). Data were preprocessed using standard techniques, and time courses from cortical and subcortical regions (defined by the Harvard-Oxford atlas) were extracted. Pearson correlation was used to calculate FC matrices, which served as features for classification. Three classifiers were used: linear SVM, non-linear rbf SVM, and GCN. A 5-fold cross-validation scheme was employed for model evaluation. Hyperparameter tuning was performed using best practices and empirical evaluation on a subset of the training data. Post-hoc visualization included GCN-Explainer to identify important connections and an ablation study to assess the contribution of each brain region to classification performance. Univariate t-tests were also conducted to investigate group differences in FC. Classification tasks included MDD vs. healthy controls (HC), medicated vs. non-medicated MDD patients, and male vs. female. The Abide and UK Biobank datasets were used for benchmarking.

The study achieved a mean classification accuracy of 61% for MDD vs. HC, which is significantly better than chance but lower than previously reported in smaller-scale studies. Classification accuracy for medicated and non-medicated subgroups was similar (around 62%). However, sex classification accuracy was substantially higher (73-81%), highlighting the potential influence of clinical heterogeneity. Visualization techniques (GCN-Explainer and ablation study) consistently identified thalamic hyperconnectivity as a prominent feature associated with MDD across both datasets. Univariate analysis revealed predominantly reduced FC in MDD, except for increased connectivity in the thalamus, though effect sizes were small. Attempts to predict symptom severity based on FC were unsuccessful.

The lower classification accuracy observed in this study compared to smaller studies suggests that increased sample size and resulting clinical heterogeneity may reduce model performance, a known challenge in neuroimaging research of psychiatric disorders. The consistent finding of thalamic hyperconnectivity across multiple visualization methods and datasets supports its potential importance as a neurophysiological marker for MDD. The lack of improved classification accuracy when analyzing medicated and non-medicated patients separately indicates that medication use has little influence on the results, at least with these techniques. The substantial difference in accuracy between MDD classification and sex classification further underscores the issue of clinical heterogeneity in MDD. The lack of success in predicting symptom severity using FC suggests that other factors, such as symptom complexity and comorbidity, contribute significantly to the clinical picture of MDD.

This study provides a realistic estimate of MDD classification performance using FC in large, multi-site datasets. While the accuracy was insufficient for clinical use, the results highlight thalamic hyperconnectivity as a potential neurophysiological signature of MDD. Future research should focus on addressing clinical heterogeneity through improved diagnostic approaches and integration of multi-modal data to improve classification accuracy and better understand the complex neural mechanisms underlying MDD.

The study's limitations include the significant clinical and technological heterogeneity inherent in large, multi-site datasets, which likely reduced classification accuracy. The reliance on a single neuroimaging modality may limit the ability to capture the full complexity of the disorder. The relatively low accuracy of the classifiers affects the reliability of visualization techniques, especially for deterministic algorithms like SVM. Analyzing FC as a stationary feature may overlook the importance of dynamic changes in neural activity. The use of resting-state fMRI may not capture the full complexity of brain activity during tasks or under specific stimuli.