PhyDA: New plug-and-play framework boosts EEG emotion recognition accuracy by 3%

A team of researchers (Hongyu Zhu et al.) has published PhyDA, a plug-and-play framework designed to tackle two persistent challenges in EEG-based emotion recognition: cross-subject variability and label noise. Traditional label-denoising methods lack physiological grounding, while physiology-informed approaches rely on hand-crafted hyperparameters. PhyDA bridges these paradigms with two modules: the Physiological Noise Quantifier (PhyNQ) uses a spectral slope to generate a subject-specific noise score, offering neurophysiologically interpretable quality assessment that adapts per individual. The Data-Adaptive Label Refiner (DALR) directly adopts this score as a contamination ratio to drive a label refinement pipeline, requiring no additional neural network training. This makes PhyDA a lightweight, tuning-free solution that directly mitigates inter-subject label noise.

The framework was rigorously evaluated on three public datasets—DEAP, SEED, and SEED-IV—across seven backbone architectures under strict leave-one-subject-out cross-validation. Results show PhyDA consistently outperforms both general and EEG-tailored label-denoising baselines, achieving average accuracy gains of 2.76%, 2.66%, and 3.32%, respectively. Visualizations further confirm its neurophysiological interpretability and practical robustness. The source code is publicly available on arXiv. While the gains may seem modest, the significance lies in PhyDA’s ability to improve performance without retraining or hyperparameter tuning, making it an attractive plug-in for existing EEG emotion recognition pipelines.