Scholarlink

<p>In emotional neuroscience and healthcare analytics, objectively estimating pain intensity from EEG signals remains a major challenge. This study proposes a compact yet powerful hybrid EEG-based regression framework that integrates both handcrafted descriptors and deep temporal-spatial representations. After preprocessing, artifact correction, and segmentation, raw EEG recordings from 64 channels (250 Hz) are processed through two complementary branches. The deep-learning branch employs an enhanced EEGNetTemporal model to learn spatial-temporal patterns directly from multichannel signals. In parallel, the classical branch extracts statistical and spectral features including PSD bandpowers, Hjorth parameters, and spectral entropy followed by Random Forest based recursive feature elimination (RF-RFE) to retain the most discriminative attributes. Regression-based predictions from both branches are fused using an XGBoost stacking meta-learner, improving robustness and interpretability. Experimental results show that the proposed hybrid model consistently surpasses individual deep or handcrafted pipelines across key evaluation metrics such as <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>, MAE, and RMSE. The proposed framework achieved a subject-independent mean <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> score of 0.619 with a mean absolute error of 1.045 under GroupKFold cross-validation. Ablation analysis confirms that combining deep features with RF-selected classical descriptors substantially enhances pain-intensity regression-based prediction performance.</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

EEG-driven pain intensity estimation using feature-level fusion and deep ensemble models

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Madhuchhanda Basak,
  • Diptadip Maiti,
  • Debasis Chaudhury

摘要

In emotional neuroscience and healthcare analytics, objectively estimating pain intensity from EEG signals remains a major challenge. This study proposes a compact yet powerful hybrid EEG-based regression framework that integrates both handcrafted descriptors and deep temporal-spatial representations. After preprocessing, artifact correction, and segmentation, raw EEG recordings from 64 channels (250 Hz) are processed through two complementary branches. The deep-learning branch employs an enhanced EEGNetTemporal model to learn spatial-temporal patterns directly from multichannel signals. In parallel, the classical branch extracts statistical and spectral features including PSD bandpowers, Hjorth parameters, and spectral entropy followed by Random Forest based recursive feature elimination (RF-RFE) to retain the most discriminative attributes. Regression-based predictions from both branches are fused using an XGBoost stacking meta-learner, improving robustness and interpretability. Experimental results show that the proposed hybrid model consistently surpasses individual deep or handcrafted pipelines across key evaluation metrics such as \(R^2\) , MAE, and RMSE. The proposed framework achieved a subject-independent mean \(R^2\) score of 0.619 with a mean absolute error of 1.045 under GroupKFold cross-validation. Ablation analysis confirms that combining deep features with RF-selected classical descriptors substantially enhances pain-intensity regression-based prediction performance.