Human emotion recognition using Machine learning techniques based on the physiological signal

Background: The study of physiological signals-based emotion detection - Electroencephalograms (EEG), Electrocardiograms (ECG), and Galvanic Skin Reactions (GSR) - has gained attention in recent years. Objective: We intend to develop machine learning-based depth physiological signal categorization for affective level and basic emotions. Methods: In this research work Random Forest (RF), K-Nearest Neighbors (KNN), and Logistic Regression (LR) with feature extraction techniques for evaluating the performance of the AMIGOS and ASCERTAIN datasets. We have used the accuracy, precision, recall, F1-measure, Cohen-kappa score, and Matthews's Correlation Coefficient (MCC) as performance evaluation metrics. Results: The proposed method effectively classified consciousness states using time-domain features with classifiers RF, LR, and KNN achieving superior results. On the AMIGOS dataset, RF improved EEG accuracy by 1.57% and F1 score by 0.25, KNN improved ECG accuracy by 1.56% and F1 score by 0.021, and RF improved GSR accuracy by 8.145% and F1 score by 0.32. On the ASCERTAIN dataset, RF improved EEG F1 score by 0.34, LR improved ECG accuracy by 22.42% and F1 score by 0.22, and RF improved GSR accuracy by 37.5% and F1 score by 0.28. Conclusion: The proposed methodology is validated through performance evaluation and comparison with related works. Results show that the RF classifier performs best for EEG and GSR signals. For ECG signals, KNN outperforms the AMIGOS dataset, while LR is superior on the ASCERTAIN dataset. Significance: Consequently, it has been observed that the proposed method might be considered a valuable tool for classifying emotional states while making automated decision-making.