E-DANN: An Enhanced Domain Adaptation Network for Audio-EEG Feature Decoupling in Explainable Depression Recognition.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Depression poses a significant global health challenge.
• AI techniques are increasingly used for objective depression detection.
• Existing AI models often lack explainability and feature importance evaluation.

Purpose of the Study:

• To propose a novel framework, Enhanced Domain Adversarial Neural Network (E-DANN), for explainable depression detection.
• To combine audio-specific physical properties and electroencephalogram (EEG) responses for multimodal feature extraction.
• To improve the clinical trustworthiness of AI-assisted diagnostic systems.

Main Methods:

• Extracting joint features from audio properties and EEG signals.
• Employing E-DANN's feature decoupling framework using adversarial training.
• Utilizing decoupled private features for binary depression classification.
• Applying Explainable Artificial Intelligence (XAI) for feature importance visualization.

Main Results:

• The E-DANN framework achieved high accuracy in classifying individuals with and without depression (accuracy: 92.83%).
• Demonstrated strong performance with specificity (93.56%), sensitivity (91.61%), and F1 score (91.81%).
• XAI approach successfully visualized feature importance and complex interactions.

Conclusions:

• The proposed E-DANN framework is effective for accurate and explainable depression detection.
• This study provides a foundation for developing trustworthy AI diagnostic tools for mental health.
• Integrating multimodal data and explainability enhances AI's clinical utility in diagnosing depression.