Scholarlink

Background <p>Emotion decoding from Electroencephalography (EEG) remains limited by cross-subject variability and low structural generalisation. This study aimed to develop a calibration-free (inference without per-subject retraining), entropy-guided graph neural network (EGGN) to enhance accuracy, Architectural interpretability, and fairness in EEG-based emotion recognition.</p> Methods <p>We used the Shanghai Jiao Tong University Emotion EEG Dataset (SEED) (<i>n</i> = 15, 3 sessions, 62 channels) and the Database for Emotion Analysis using Physiological Signals (DEAP) (<i>n </i>= 32, 40 trials, 32 channels) datasets, segmented into 3-s epochs with a sampling rate of 128&#xa0;Hz. Models included Graph Convolutional Network (GCN), Transformer, and Entropy-Guided Graph Neural Network (EGGN), trained using 10-fold cross-validation. Graph edges were derived via entropy gradients; dual-attention pooling and fairness metrics were applied. Saliency and graph fingerprint analyses were performed on SEED; calibration-free and meta-learning evaluations were conducted on DEAP.</p> Results <p>EGGN achieved the highest accuracy: 84.2% (valence) and 83.1% (arousal) on SEED, outperforming Transformer (80.5%, 79.8%) and GCN (78.1%, 76.4%). Entropy-based edge modelling improved accuracy by 2.8–3.9%; dual-attention pooling further increased accuracy (+ 3.6–4.1%) and Neuropsychological Alignment Score (NAS) defined by cosine similarity between predicted saliency maps and established emotion-related brain regions (+ 0.14–0.17). Fairness metrics showed minimal gender bias (Disparate Impact Ratio (DIR) = 0.97; Equal Opportunity Difference (EOD) = –1.1%). Graph similarity peaked at 0.74 for high arousal (SEED–DEAP); Uniform Manifold Approximation and Projection (UMAP) silhouette score was 0.64. Calibration-free inference yielded a minor performance drop (EGGN Δ = –1.4% to –0.8%).</p> Conclusion <p>EGGN enables robust, interpretable, and fair emotion decoding from electroencephalography across subjects and datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Analysis of the generalization ability of graph neural networks in cross-subject EEG emotion recognition

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Lingyue Wang,
  • Lei Guo,
  • Xinsheng Yang,
  • Ying Li

摘要

Background

Emotion decoding from Electroencephalography (EEG) remains limited by cross-subject variability and low structural generalisation. This study aimed to develop a calibration-free (inference without per-subject retraining), entropy-guided graph neural network (EGGN) to enhance accuracy, Architectural interpretability, and fairness in EEG-based emotion recognition.

Methods

We used the Shanghai Jiao Tong University Emotion EEG Dataset (SEED) (n = 15, 3 sessions, 62 channels) and the Database for Emotion Analysis using Physiological Signals (DEAP) (n = 32, 40 trials, 32 channels) datasets, segmented into 3-s epochs with a sampling rate of 128 Hz. Models included Graph Convolutional Network (GCN), Transformer, and Entropy-Guided Graph Neural Network (EGGN), trained using 10-fold cross-validation. Graph edges were derived via entropy gradients; dual-attention pooling and fairness metrics were applied. Saliency and graph fingerprint analyses were performed on SEED; calibration-free and meta-learning evaluations were conducted on DEAP.

Results

EGGN achieved the highest accuracy: 84.2% (valence) and 83.1% (arousal) on SEED, outperforming Transformer (80.5%, 79.8%) and GCN (78.1%, 76.4%). Entropy-based edge modelling improved accuracy by 2.8–3.9%; dual-attention pooling further increased accuracy (+ 3.6–4.1%) and Neuropsychological Alignment Score (NAS) defined by cosine similarity between predicted saliency maps and established emotion-related brain regions (+ 0.14–0.17). Fairness metrics showed minimal gender bias (Disparate Impact Ratio (DIR) = 0.97; Equal Opportunity Difference (EOD) = –1.1%). Graph similarity peaked at 0.74 for high arousal (SEED–DEAP); Uniform Manifold Approximation and Projection (UMAP) silhouette score was 0.64. Calibration-free inference yielded a minor performance drop (EGGN Δ = –1.4% to –0.8%).

Conclusion

EGGN enables robust, interpretable, and fair emotion decoding from electroencephalography across subjects and datasets.