Social anxiety prediction based on ERP features: A deep learning approach.

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Social Anxiety Disorder (SAD) diagnosis relies on subjective scales, limiting accuracy.
• Electroencephalography (EEG) offers objective neurophysiological data for anxiety detection.
• Existing EEG methods often neglect task-related Event-Related Potential (ERP) features.

Purpose of the Study:

• To investigate the efficacy of task-related ERP features for SAD detection using deep learning.
• To compare the performance of EEGNet with other machine learning models for SAD recognition.
• To evaluate the stability and generalizability of the findings across datasets.

Main Methods:

• Collected EEG data from 63 participants viewing facial expressions.
• Extracted task-relevant ERP features and employed the EEGNet model for SAD prediction.
• Compared EEGNet against DeepConvNet, ShallowConvNet, BiLSTM, and SVM, validating on a prior dataset.

Main Results:

• EEGNet achieved 99.16% accuracy using Late Positive Potential (LPP) ERP components.
• ERP features demonstrated superior accuracy for SAD recognition compared to time- and frequency-domain features.
• Higher accuracy was observed for neutral and negative facial stimuli; findings were consistent across datasets.

Conclusions:

• Task-related ERP features, particularly LPP, show significant potential for objective SAD recognition.
• The EEGNet model effectively utilizes ERPs for high-accuracy SAD detection.
• Recognizing SAD using neutral or negative facial stimuli is more effective.