Akademik Veri Yönetim Sistemi
Signal, Image and Video Processing, cilt.17, sa.7, ss.3783-3791, 2023 (SCI-Expanded)
Emotion recognition through physiological signals is of great importance for the examination of psychological states and the development of biofeedback-based applications. Thanks to the emergence of the concept of metaverse and the integration of physiological signal trackers into smart devices, this field of study has become a necessity and a subject of interest to researchers. In this study, an algorithm is proposed for emotion detection according to the two-category (valence–arousal) emotion model. ECG signal recordings from the MAHNOB-HCI database were used in the study. First, the noise on the ECG signals is eliminated in the preprocessing step. R peaks were detected by applying the Pan–Tompkins algorithm to the preprocessed ECG signals. Then, for each recording, the P-QRS-T fragment and the maximum and minimum values of the P, Q, R, S, and T waves were obtained as morphological features and combined with selected heart rate variability features to obtain a feature vector. By applying an automated feature engineering algorithm to this feature vector, new feature vectors with increased weight of distinctive features and increased number of samples are obtained as output. These features are classified with three different learning algorithms: support vector machines, feedforward neural network, and bidirectional long short-term memory. As a result of the study, good results were obtained with the bidirectional long short-term memory algorithm compared to the literature. According to these results, with bidirectional long short-term memory, the accuracy obtained was 78.28% for the valence category and 83.61% for the arousal category.