DefinePK

Abstract

Understanding emotions through EEG signals plays a crucial role in fields like healthcare, human-computer interaction, and affective neuroscience. It classified Disgust, Fear, Happy, Neutral, and Sad as five emotional states, combining eye-blinking features with differential entropy attributes extracted from EEG across the SEED-V dataset. The features were standardized via padding, while the EMOTION labels were encoded by one-hot encoder to enhance the capability of model training. The advanced DL architectures CNN, LSTM, and TFT were considered. The linear regression model in this study is that of LSTM; CNN is the secondary modeling method, while TFT has relatively less performance but excels in interpretability. The results suggest that LSTM outperformed both CNN and TFT, reaching the maximum accuracy of 80%. There was a difference of about 5 to 10% from that of CNN at 78% to that of TFT at 60%. Also, the macro average F1-score of LSTM improves upon that of CNN, 0.77, and of TFT, 0.58, indicating better performance for LSTM in the capture of long-term dependencies in sequential EEG data. While CNN is mostly concerned with spatial features, TFT has been challenged in dealing with fine temporal dependence. Therefore, LSTM incurs obstacles of increased computation complexity, longer training time, and potential overfitting when it comes to smaller datasets. LSTM-based models further inherit a limitation concerning their non-interpretability owing to a detailed internal structure. Even with these setbacks, the findings highlight the importance of physiologic (EEG) and behavioral (eye-blinking) feature integration for solid emotion recognition, paving the road for advancements in mental health care, brain-computer interface, and adaptive human-computer interaction.