EEG-based emotion analysis using non-linear features and ensemble learning approaches

Recognition of emotions based on electroencephalography (EEG) has become one of the most emerging topics for healthcare, education system, knowledge sharing, gaming, and many other fields in the last few decades. This paper proposes three non-linear features (Higuchi fractal dimension, sample entropy, and permutation entropy) and eight ensemble learning approaches to predict six basic emotions (hope, interest, excitement, shame, fear, and sad). To increase the recognition rate of each classifier, we utilized a randomized grid search technique for tunning the hyperparameter of each algorithm. Moreover, the impact of electrodes on each emotion was observed using the recursive feature elimination (RFE) method. In addition, the synthetic minority oversampling technique (SMOTE) is used to handle the imbalanced sample distribution of each emotion. Then, we have conducted all the experiments on DEAP and AMIGOS datasets and calculated the computation time and accuracy to assess each algorithm's performance. Besides, we observe statistical significance to compare the algorithm's performance. From the experimental result, we achieved the highest average accuracy, 89.38% and 94.62%, using Higuchi fractal dimension on DEAP and AMIGOS datasets, respectively. We also observed that the Higuchi fractal dimension outperforms the sample entropy and permutation entropy in terms of accuracy. Our proposed method increased the average recognition rate by 8.22% and 1.77%, respectively, compared with existing approaches working on the same dataset. Along with accuracy, precision, recall, F-measure, and AUC-ROC have been evaluated using the same experimental setting on DEAP and AMIGOS datasets. The results of each classifier have been shown in the table and graph. Moreover, the proposed method outperforms existing approaches dealing with shallow machine learning techniques.