Group Sparse Representation Enhances Brain Network Classification of Major Depressive Disorder in Two Chinese Cohorts

Area of Science:

• Neuroimaging
• Computational Psychiatry
• Network Neuroscience

Background:

• Major depressive disorder (MDD) is characterized by disruptions in functional brain network organization.
• Conventional methods like Pearson correlation (PC) have limitations in capturing these complex network alterations.
• Identifying robust biomarkers for MDD diagnosis remains a critical challenge in clinical neuroscience.

Purpose of the Study:

• To evaluate the diagnostic efficacy of three distinct brain network construction methods: Pearson correlation (PC), sparse representation (SR), and group sparse representation (GSR).
• To compare the classification performance of PC, SR, and GSR networks in differentiating individuals with MDD from healthy controls (HCs).
• To validate the findings using an independent dataset.

Main Methods:

• Functional magnetic resonance imaging (fMRI) data were acquired from 117 Chinese participants (61 MDD, 56 HCs).
• Whole-brain networks were constructed using PC, SR, and GSR from time-series signals of 116 brain regions.
• A linear support vector machine (SVM) classifier with LASSO feature selection and leave-one-out cross-validation (LOOCV) was employed for classification. An independent dataset was used for validation.

Main Results:

• The group sparse representation (GSR) network demonstrated superior classification performance compared to PC and SR.
• GSR achieved an area under the receiver operating characteristic curve (AUC) of 0.85, accuracy of 0.81, and sensitivity of 0.95.
• These superior results were corroborated in an independent dataset, identifying 17 key brain connections and 27 critical brain regions within the GSR network.

Conclusions:

• Group sparse representation (GSR) offers a robust and effective approach for constructing brain networks for MDD diagnosis.
• GSR-based brain networks show significant potential as a diagnostic tool, outperforming traditional methods like Pearson correlation (PC).
• These findings advocate for the integration of advanced network construction techniques in neuroimaging research for psychiatric disorders.