Explainable Depression Classification Based on EEG Feature Selection From Audio Stimuli.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Psychiatry

Background:

• Electroencephalogram (EEG)-based depression detection is advancing with AI.
• Existing methods often overlook feature-model associations and individual feature contributions.
• A need exists for explainable AI in EEG-based depression recognition.

Purpose of the Study:

• To develop an innovative EEG-based depression detection method using Ant-Lion Optimization (ALO) and Multi-Agent Reinforcement Learning (MARL).
• To enhance model explainability and identify critical EEG features for depression recognition.
• To analyze dynamic brain state transitions in response to audio stimuli.

Main Methods:

• Integration of ALO and MARL for feature fusion analysis in EEG data.
• Application of Explainable Artificial Intelligence (XAI) for feature interpretability.
• Utilizing Time-Delay Embedded Hidden Markov Model (TDE-HMM) for inferring brain states.
• Hyper-parameter optimization of XGBoost classifier.

Main Results:

• Achieved high accuracy (93.69%), sensitivity (88.60%), specificity (97.08%), and F1-score (91.82%) on an EEG dataset.
• Outperformed state-of-the-art feature selection methods.
• XAI identified minimum Power Spectral Density (PSD), Sample Entropy (SampEn), and Rényi Entropy (Ren) as critical features.

Conclusions:

• The proposed ALO-MARL approach with XAI offers a robust and explainable method for EEG-based depression detection.
• Identified key EEG features provide insights into the neurophysiological underpinnings of depression.
• The findings support the clinical application of AI in depression recognition and understanding brain dynamics.