Multivariate Machine Learning Analyses in Identification of Major Depressive Disorder Using Resting-State Functional...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Diagnosing major depressive disorder (MDD) is challenging due to data complexity and individual variability.
• Resting-state functional connectivity (rs-FC) offers potential biomarkers but requires robust analytical methods.
• Machine learning (ML) can address high dimensionality and improve diagnostic accuracy in MDD.

Purpose of the Study:

• To evaluate the clinical utility of rs-FC for MDD diagnosis.
• To identify an optimal rs-FC-based ML model for individualized MDD classification.
• To explore the relationship between rs-FC features and MDD symptom severity.

Main Methods:

• A progressive three-step ML analysis was conducted on a large, multicenter dataset (1021 MDD patients, 1100 controls).
• Six ML algorithms and two dimension reduction techniques were employed to assess classification performance.
• A linear regression model analyzed the correlation between rs-FC features and Hamilton Depression Scale scores.

Main Results:

• The eXtreme Gradient Boosting (XGBoost) model demonstrated superior classification performance (accuracy=0.728, AUC=0.831) for distinguishing MDD patients from controls.
• Key rs-FCs identified by XGBoost were located within/between the default mode, limbic, and visual networks.
• The XGBoost model accurately predicted individual Hamilton Depression Scale scores (adjusted R²=0.180).

Conclusions:

• The XGBoost model utilizing rs-FC provides a robust and interpretable approach for MDD diagnosis.
• rs-FC features identified by XGBoost have clinical relevance for predicting MDD symptom severity.
• This study highlights the potential of ML-driven rs-FC analysis for personalized MDD assessment.