Depression Recognition Using Machine Learning Algorithms With Eye Tracking, Visual Evoked Potentials, and Auditory...

Area of Science:

• Neuroscience
• Machine Learning
• Psychiatry

Background:

• Current depression assessment relies on subjective psychological scales.
• Limited research exists on using multiple neurophysiological measurements for depression classification.
• Machine learning (ML) offers potential for objective depression diagnosis.

Purpose of the Study:

• To employ ML algorithms integrating eye tracking, visual evoked potentials (VEPs), and auditory P300 for depression classification in Chinese medical students.
• To investigate the relationship between neurophysiological features and depression severity (PHQ-9 scores).
• To identify the most significant neurophysiological predictors of depression.

Main Methods:

• Recruited 66 students with depression and 72 controls.
• Collected eye tracking, VEPs, and auditory P300 data.
• Utilized multivariate logistic regression, six ML classifiers (including Random Forest), and SHapley Additive exPlanations (SHAP) for analysis and feature importance.
• Assessed model performance using ROC curves, AUC, precision, accuracy, recall, and F1 score with five-fold cross-validation.

Main Results:

• Depression group showed lower response search scores and prolonged P100 latencies in VEPs.
• No significant differences in auditory P300 features were found between groups.
• Random Forest classifier achieved superior performance; combined eye tracking and VEP features outperformed single modalities.
• SHAP analysis identified right eye P100 latency as the most significant predictive feature.

Conclusions:

• Depression in Chinese medical students is associated with impaired attention and visual processing, indicated by reduced search scores and prolonged P100 latencies.
• Combining eye tracking and VEP data enhances depression classification accuracy.
• Right eye P100 latency emerges as a critical neurophysiological marker for depression detection.