Liao Yilong, Gao Yuan, Wang Fang, Zhang Li, Xu Zhenrong, Wu Yifan
School of Biomedical Engineering, South-Central Minzu University, Wuhan, 430074, China.
South-Central Minzu University, No.182, Minzu Avenue, Hongshan District, Wuhan City, Hubei Province, China.
Sci Rep. 2025 Jun 6;15(1):19869. doi: 10.1038/s41598-025-96616-0.
Emotion recognition is a key research area in artificial intelligence, playing a critical role in enhancing human-computer interaction and optimizing user experience design. This study explores the application and effectiveness of ensemble learning methods for emotion recognition based on multiple physiological parameters. A dataset was systematically constructed by preprocessing data from electroencephalogram (EEG), galvanic skin response (GSR), skin temperature (ST), and heart rate (HR) collected from 38 subjects while watching short videos. We proposed a hybrid model framework combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, trained and optimized using a random seed initialization strategy and a cosine annealing warm restart strategy. To further enhance performance, various strategies were designed and evaluated. The results showed that applying advanced preprocessing techniques significantly improved data quality, while the hybrid model effectively leveraged the advantages of both CNN and LSTM. Incorporating the cosine annealing warm restart strategy further boosted model performance. Using a soft voting ensemble method, the proposed approach achieved a 96.21% accuracy rate in classifying seven emotions-calm, happy, disgust, surprise, anger, sad, and fear, indicating its ability to accurately capture emotional responses to short videos. This study presents an innovative approach to emotion recognition using multiple physiological parameters, demonstrating the potential of ensemble learning for complex tasks. It offers valuable insights for the development of effective applications.
情感识别是人工智能中的一个关键研究领域，在增强人机交互和优化用户体验设计方面发挥着至关重要的作用。本研究探讨了基于多个生理参数的集成学习方法在情感识别中的应用及有效性。通过对38名受试者观看短视频时收集的脑电图（EEG）、皮肤电反应（GSR）、皮肤温度（ST）和心率（HR）数据进行预处理，系统构建了一个数据集。我们提出了一种结合卷积神经网络（CNN）和长短期记忆（LSTM）网络的混合模型框架，并使用随机种子初始化策略和余弦退火热重启策略进行训练和优化。为进一步提高性能，设计并评估了各种策略。结果表明，应用先进的预处理技术显著提高了数据质量，而混合模型有效地利用了CNN和LSTM的优势。采用余弦退火热重启策略进一步提升了模型性能。使用软投票集成方法，该方法在对平静、快乐、厌恶、惊讶、愤怒、悲伤和恐惧七种情绪进行分类时，准确率达到了96.21%，表明其能够准确捕捉对短视频的情感反应。本研究提出了一种利用多个生理参数进行情感识别的创新方法，展示了集成学习在复杂任务中的潜力。它为有效应用的开发提供了有价值的见解。
