Scholarlink

<p>Emotion identification using electroencephalogram (EEG) data is now recognised as a critical strategy for developing intelligent human-computer interaction (HCI) and brain-computer interface (BCI) technologies. This paper explores the effectiveness of three hybrid deep learning (DL) frameworks—CNN1D-LSTM, CNN1D-GRU, and GRU-LSTM—for categorizing emotional states utilizing EEG signals from an online available dataset. Every algorithm was trained and assessed using three optimization strategies: Adam, RMSprop, and Stochastic Gradient Descent (SGD). Experimental findings show that when optimized with the Adam method, the CNN-GRU outperformed with an F1 score of 99.5% and a 5-fold cross-validation average F1 Score of 97.5% (95% CI 95.8–99.1%). Furthermore, statistical analyses, including t-tests,<i> p</i>-values, Cohen’s d, Glass’s Δ, and ANOVA, confirmed the model’s significance and stability. The results demonstrate the durability and reliability of hybrid DL models in capturing the complex temporal and spatial correlations in EEG data. The findings have important implications for developing real-time affective computing platforms and improving the usability of EEG-based emotion recognition in real-world HCI and BCI systems.</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

Enhanced emotion detection in EEG using CNN1D and RNN hybrids with adaptive optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月8日

查看全文
  • Madhusudan G Lanjewar,
  • Manikandan Vinodh Kumar,
  • Kamini G Panchbhai,
  • Rajesh K. Parate

摘要

Emotion identification using electroencephalogram (EEG) data is now recognised as a critical strategy for developing intelligent human-computer interaction (HCI) and brain-computer interface (BCI) technologies. This paper explores the effectiveness of three hybrid deep learning (DL) frameworks—CNN1D-LSTM, CNN1D-GRU, and GRU-LSTM—for categorizing emotional states utilizing EEG signals from an online available dataset. Every algorithm was trained and assessed using three optimization strategies: Adam, RMSprop, and Stochastic Gradient Descent (SGD). Experimental findings show that when optimized with the Adam method, the CNN-GRU outperformed with an F1 score of 99.5% and a 5-fold cross-validation average F1 Score of 97.5% (95% CI 95.8–99.1%). Furthermore, statistical analyses, including t-tests, p-values, Cohen’s d, Glass’s Δ, and ANOVA, confirmed the model’s significance and stability. The results demonstrate the durability and reliability of hybrid DL models in capturing the complex temporal and spatial correlations in EEG data. The findings have important implications for developing real-time affective computing platforms and improving the usability of EEG-based emotion recognition in real-world HCI and BCI systems.