Subject-independent emotion recognition with EEG bispectral quadratic phase coupling features and explainable machine...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Subject-independent emotion recognition from electroencephalography (EEG) faces challenges due to complex neural dynamics and individual differences.
• Existing methods often rely on high-dimensional models, limiting interpretability and efficiency.

Purpose of the Study:

• To characterize and select compact bispectral quadratic phase coupling (QPC) descriptors for robust EEG-based emotion recognition.
• To evaluate the performance of these descriptors using an Explainable Boosting Machine (EBM) for improved transparency.

Main Methods:

• Extraction of nine bispectral QPC descriptors from frontal EEG rhythms in the DEAP dataset.
• Feature selection using a genetic algorithm with nested leave-one-subject-out (LOSO) cross-validation.
• Classification using an Explainable Boosting Machine (EBM) for interpretable results.

Main Results:

• Bispectral QPC features significantly outperformed power spectral density and differential entropy baselines for arousal and valence classification (p < 0.05).
• The pipeline achieved 70.39% accuracy for arousal and 69.98% for valence on the DEAP dataset, comparable to deep learning models.
• The EBM provided interpretable feature interactions consistent with neuroscience literature, while offering faster inference and a smaller model footprint.

Conclusions:

• Bispectral QPC descriptors offer a compact, transparent, and efficient approach for subject-independent EEG emotion classification.
• The interpretable nature of the EBM framework enhances understanding of the underlying neural correlates of emotion.
• While effective within datasets, cross-dataset transfer (e.g., DEAP to DREAMER) remains a challenge.