Decoding Major Depressive Disorder Neuroconnectivity: Heterogeneity and Demographic Effects
Poster No:
1819
Submission Type:
Abstract Submission
Authors:
Farzad Farahani1, Dongnhu Truong1, Rouhollah Abdollahi1, Gamal Abdel-Azim1, Christopher Whelan1, Shuwei Li1, Julio Molineros1
Institutions:
1Johnson & Johnson, Spring House, PA
First Author:
Co-Author(s):
Introduction:
Major Depressive Disorder (MDD) affects 3.8% of the global population (GBD Results, 2022), with disrupted functional connectivity playing a key role in its neurobiology. This study utilized ICA-based connectome derived from ~40K patient resting-state functional magnetic resonance imaging (rs-fMRI) data from the UK Biobank (UKB; Miller et al. 2016), a biomedical data repository with 500K voluntary participants aged 40-69 at recruitment. The objectives were to (1) identify connectivity disruptions in MDD, and (2) develop a predictive model for MDD classification against healthy controls (HC), excluding participants with any other mental/neurologic disorders. Additionally, we explored heterogeneity within the HC to further evaluate and enhance model accuracy and generalizability.
Methods:
UKB Data (ICA Connectome): Our study included 13,791 subjects (12,656 HC vs. 1,135 MDD) in the UKB with rs-fMRI data and confirmed MDD diagnosis determined by International Classification of Diseases codes. The dataset featured 21 independent components derived through group-ICA decomposition, with 210 edges (correlations) as input features from the lower triangular part of the ICA connectome. To address the imbalanced class distribution, SMOTE (Synthetic Minority Over-sampling Technique) resampling was used to augment the minority class (MDD).
Clustering HC: K-means clustering was used to classify controls into two clusters (C1 and C2) because of the observed heterogeneity of the healthy sample. Subsequent analyses explored both the entire dataset and each cluster independently, offering insights into distinctive patterns within the control group that contributed to a more nuanced understanding of factors influencing downstream MDD classification based on the ICA connectome.
Classification Models: Various classifiers (including logistic regression, support vector machines, and XGBoost) were used, each followed by 5-fold cross-validation. Stacking classifiers leveraged individual model strengths for enhanced classification performance. Performance was compared across scenarios using metrics such as accuracy, precision, recall, and F1-score.
Clustering HC: K-means clustering was used to classify controls into two clusters (C1 and C2) because of the observed heterogeneity of the healthy sample. Subsequent analyses explored both the entire dataset and each cluster independently, offering insights into distinctive patterns within the control group that contributed to a more nuanced understanding of factors influencing downstream MDD classification based on the ICA connectome.
Classification Models: Various classifiers (including logistic regression, support vector machines, and XGBoost) were used, each followed by 5-fold cross-validation. Stacking classifiers leveraged individual model strengths for enhanced classification performance. Performance was compared across scenarios using metrics such as accuracy, precision, recall, and F1-score.
Results:
Deciphering Heterogeneity in MDD
Average correlation matrices from rs-fMRI data reveal neuroconnectivity patterns in HC and MDD groups (Fig 1A). There were significant group differences highlighted in the binary matrix through network-based statistics (Fig 1B) (Zalesky et al., 2010). Initial ensemble classification models showed low relative accuracy (Fig 1G; purple bars). The t-SNE projection further emphasized the intricacies of group separation (Fig 1C). Notably, delving into the heterogeneity within the HC group uncovered distinctive clusters (C1 and C2) (Fig 1D). While there are significant differences in age distribution between the full HC group and MDD (Fig 1E), C2 showed a similar age distribution to MDD patients (Fig 1F), which improved prediction performance of MDD vs C2 (Fig 1G).
Neural Signatures: SHAP Analysis Unravels Key Edges/Components in MDD Classification
SHAP matrix displays edges that significantly contributed to the model's performance upon training on the HC (C2) and MDD groups, as shown in Fig 2A (Lundberg & Lee, 2017). In Fig 2B, the top 20 influential edges are color-coded (red and blue) in the summary plot, indicating their positive and negative effects on the model's performance, respectively. Components 10, 8, 12, and 18 (members of the set of significant edges) had the strongest effect in classification accuracy (as shown in the chord diagram Fig 2C).
Average correlation matrices from rs-fMRI data reveal neuroconnectivity patterns in HC and MDD groups (Fig 1A). There were significant group differences highlighted in the binary matrix through network-based statistics (Fig 1B) (Zalesky et al., 2010). Initial ensemble classification models showed low relative accuracy (Fig 1G; purple bars). The t-SNE projection further emphasized the intricacies of group separation (Fig 1C). Notably, delving into the heterogeneity within the HC group uncovered distinctive clusters (C1 and C2) (Fig 1D). While there are significant differences in age distribution between the full HC group and MDD (Fig 1E), C2 showed a similar age distribution to MDD patients (Fig 1F), which improved prediction performance of MDD vs C2 (Fig 1G).
Neural Signatures: SHAP Analysis Unravels Key Edges/Components in MDD Classification
SHAP matrix displays edges that significantly contributed to the model's performance upon training on the HC (C2) and MDD groups, as shown in Fig 2A (Lundberg & Lee, 2017). In Fig 2B, the top 20 influential edges are color-coded (red and blue) in the summary plot, indicating their positive and negative effects on the model's performance, respectively. Components 10, 8, 12, and 18 (members of the set of significant edges) had the strongest effect in classification accuracy (as shown in the chord diagram Fig 2C).
Conclusions:
In conclusion, addressing network heterogeneity within the control group significantly improved prediction performance of MDD. This highlighted the gross network differences between MDD and age-matched HC (C2) in the ICA-based connectome dataset. We identified four major components (8, 10, 12, and 18) from significant edges that significantly impacted MDD classification accuracy based on SHAP analysis of the connectivity patterns.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 2
Modeling and Analysis Methods:
Classification and Predictive Modeling
fMRI Connectivity and Network Modeling 1
Keywords:
Computational Neuroscience
FUNCTIONAL MRI
Machine Learning
Psychiatric Disorders
Other - Major Depressive Disorder
1|2Indicates the priority used for review
Provide references using author date format
Zalesky, A., Fornito, A., Bullmore, E. T. "Network-based Statistic: Identifying Differences in Brain Networks." Neuroimage 53.4 (2010): 1197-1207.
Lundberg, S. M., Lee, S.-I. "A Unified Approach to Interpreting Model Predictions." Adv. Neural Inf. Process. Syst. 30 (2017).