Dynamic Multimodal Measurement of Depression Severity Using Deep Autoencoding.

Area of Science:

• Psychiatry
• Computational Neuroscience
• Behavioral Science

Background:

• Depression is a prevalent global psychiatric disorder affecting over 350 million people.
• Current depression assessment relies on subjective clinical interviews and self-report scales.
• Objective, systematic behavioral indicators are lacking in current depression assessment methods.

Purpose of the Study:

• To develop and validate automated classifiers for detecting depression severity.
• To investigate the efficacy of facial movement, head movement, and vocal prosody as behavioral indicators of depression.
• To determine the optimal combination of behavioral modalities for accurate depression severity assessment.

Main Methods:

• Trained logistic regression classifiers on behavioral data (facial movement, head movement, vocalization) from individuals with major depressive disorder (MDD).
• Utilized leave-one-participant-out cross-validation for robust accuracy assessment.
• Compared the performance of individual and combined behavioral modalities in detecting three levels of depression severity (moderately to severely depressed, mildly depressed, remitted).

Main Results:

• Facial movement dynamics demonstrated higher accuracy in detecting depression severity compared to head movement dynamics and vocal prosody.
• Head movement dynamics also showed significant accuracy, outperforming vocal prosody.
• Combining all three modalities (facial, head, vocal) offered only a marginal improvement over using facial and head movements alone.

Conclusions:

• Automatic detection of depression severity from observable behavioral indicators is feasible.
• Facial movement dynamics are a powerful, objective indicator for assessing depression severity.
• Multimodal behavioral analysis, particularly incorporating facial and head movements, provides the most effective approach for objective depression assessment.