Predicting future depressive episodes from resting-state fMRI with generative embedding.

Area of Science:

• Neuroimaging
• Computational Psychiatry
• Machine Learning

Background:

• Major depressive disorder (MDD) often follows a relapsing course after a first episode, highlighting the need for early detection and intervention strategies.
• While predictive modeling for depression exists, the use of functional magnetic resonance imaging (fMRI) for predicting future depressive episodes remains underexplored.

Purpose of the Study:

• To investigate the predictive utility of generative models (dynamic causal models, DCMs) and functional connectivity (FC) for forecasting future depression in adults using resting-state fMRI data.
• To identify the most effective combination of modeling approaches and classifiers for predicting depression onset.

Main Methods:

• Utilized a large UK Biobank dataset (N=906) of resting-state fMRI data from never-depressed adults, with 50% experiencing a depressive episode over three years.
• Employed generative embedding via regression DCM (rDCM) combined with support vector machine (SVM) and analyzed functional connectivity (FC) using independent component analysis.
• Implemented nested cross-validation for model training and a held-out test set for performance evaluation, assessing various feature sets and classifiers.

Main Results:

• The combination of rDCM and SVM achieved the highest prediction accuracy (0.62) and AUC (0.64) on the test set, outperforming FC-based predictions (0.59 accuracy, 0.61 AUC).
• SHAP value analysis indicated that predictive brain connections were broadly distributed across networks, not localized to specific regions.
• While promising, the achieved accuracy suggests that fMRI alone may not be sufficient for clinically relevant depression prediction.

Conclusions:

• Generative embedding models, particularly rDCM, show potential for improving the early detection of individuals at risk for depression using fMRI data.
• Predictive models based on functional connectivity also offer utility, though rDCM demonstrated a slight advantage.
• Future advancements in clinical utility may necessitate integrating fMRI data with other biological or clinical modalities for more robust depression risk prediction.