Pane Evi Septiana, Wibawa Adhi Dharma, Purnomo Mauridhi Hery
Department of Electrical Engineering, Institut Teknologi Sepuluh Nopember, Building B and AJ, Raya ITS, Surabaya, 60111, Indonesia.
Industrial Training and Education of Surabaya, Ministry of Industry, Surabaya, Indonesia.
Cogn Process. 2019 Nov;20(4):405-417. doi: 10.1007/s10339-019-00924-z. Epub 2019 Jul 24.
For emotion recognition using EEG signals, the challenge is improving accuracy. This study proposes strategies that concentrate on incorporating emotion lateralization and ensemble learning approach to enhance the accuracy of EEG-based emotion recognition. In this paper, we obtained EEG signals from an EEG-based public emotion dataset with four classes (i.e. happy, sad, angry and relaxed). The EEG signal is acquired from pair asymmetry channels from left and right hemispheres. EEG features were extracted using a hybrid features extraction from three domains, namely time, frequency and wavelet. To demonstrate the lateralization, we performed a set of four experimental scenarios, i.e. without lateralization, right-/left-dominance lateralization, valence lateralization and others lateralization. For emotion classification, we use random forest (RF), which is known as the best classifier in ensemble learning. Tuning parameters in the RF model were done by grid search optimization. As a comparison of RF, we employed two prevalent algorithms in EEG, namely SVM and LDA. Emotion classification accuracy increased significantly from without lateralization to the valence lateralization using three pairs of asymmetry channel, i.e. T7-T8, C3-C4 and O1-O2. For the classification, the RF method provides the highest accuracy of 75.6% compared to SVM of 69.8% and LDA of 60.4%. In addition, the features of energy-entropy from wavelet are important for EEG emotion recognition. This study yields a significant performance improvement of EEG-based emotion recognition by the valence emotion lateralization. It indicates that happy and relaxed emotions are dominant in the left hemisphere, while angry and sad emotions are better recognized from the right hemisphere.
对于使用脑电图（EEG）信号进行情感识别而言，挑战在于提高准确率。本研究提出了一些策略，这些策略专注于融合情感偏侧化和集成学习方法，以提高基于EEG的情感识别准确率。在本文中，我们从一个基于EEG的公共情感数据集中获取了EEG信号，该数据集包含四个类别（即快乐、悲伤、愤怒和放松）。EEG信号是从左右半球的成对不对称通道采集的。使用从时间、频率和小波三个域进行混合特征提取的方法来提取EEG特征。为了证明偏侧化，我们进行了一组四个实验场景，即无偏侧化、右/左优势偏侧化、效价偏侧化和其他偏侧化。对于情感分类，我们使用随机森林（RF），它是集成学习中公认的最佳分类器。RF模型中的调优参数通过网格搜索优化来完成。作为RF的比较，我们采用了EEG中两种流行的算法，即支持向量机（SVM）和线性判别分析（LDA）。使用三对不对称通道（即T7 - T8、C3 - C4和O1 - O2）时，从无偏侧化到效价偏侧化，情感分类准确率显著提高。对于分类，RF方法提供了最高的准确率，为75.6%，相比之下SVM为69.8%，LDA为60.4%。此外，小波能量熵特征对于EEG情感识别很重要。本研究通过效价情感偏侧化在基于EEG的情感识别方面取得了显著的性能提升。这表明快乐和放松的情绪在左半球占主导，而愤怒和悲伤的情绪从右半球能得到更好的识别。
