Predicting individual clinical trajectories of depression with generative embedding

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Major depressive disorder (MDD) exhibits diverse clinical trajectories and treatment responses.
• Predicting individual patient outcomes early is crucial for personalized psychiatric interventions.
• Current methods lack reliable predictors for patient-specific clinical courses.

Purpose of the Study:

• To evaluate a machine learning strategy, generative embedding (GE), for predicting clinical trajectories in MDD patients.
• To assess the utility of functional magnetic resonance imaging (fMRI) data in conjunction with GE for outcome prediction.
• To identify specific neural network properties associated with distinct MDD clinical courses.

Main Methods:

• Utilized fMRI data from 85 MDD patients from the NEtherlands Study of Depression and Anxiety (NESDA) cohort.
• Employed generative embedding (GE), combining generative models with support vector machines (SVMs).
• Analyzed emotional face perception tasks and effective connectivity during a two-year follow-up.

Main Results:

• GE accurately predicted chronic depression versus fast remission (79% balanced accuracy).
• Prediction of gradual improvement versus fast remission was above chance (61% balanced accuracy).
• GE outperformed traditional methods relying on functional connectivity or local activation estimates.

Conclusions:

• Generative embedding shows potential for interpretable clinical predictions in psychiatry.
• Abnormal dynamic changes in emotional face processing networks may indicate a higher risk of unfavorable MDD course.
• Findings suggest GE can identify network mechanisms underlying heterogeneous MDD trajectories.