Evaluation of electrocardiogram: numerical vs. image data for emotion recognition system

Area of Science:

• Physiological signal processing
• Affective computing
• Machine learning for emotion recognition

Background:

• Electrocardiogram (ECG) is vital for cardiovascular health monitoring.
• Previous research explored 1-D ECG for emotion detection, but 2-D ECG image format is less studied.
• A consensus on the impact of ECG input format (1-D vs. 2-D) on emotion recognition system (ERS) accuracy is lacking.

Purpose of the Study:

• To investigate the effect of ECG input format on ERS performance.
• To compare the accuracy of ERS using 1-D numerical ECG data versus 2-D ECG images.

Main Methods:

• Utilized the DREAMER dataset comprising 23 ECG recordings during emotional elicitation.
• Extracted features from 1-D ECG using AUBT and TEAP; extracted features from 2-D ECG images using ORB, SIFT, KAZE, AKAZE, BRISK, and HOG.
• Applied Linear Discriminant Analysis (LDA) for dimensionality reduction and Support Vector Machine (SVM) for classifying valence and arousal.

Main Results:

• 1-D ECG-based ERS achieved 65.06% accuracy and 75.63% F1-score for valence; 57.83% accuracy and 44.44% F1-score for arousal.
• 2-D ECG-based ERS yielded a maximum of 62.35% accuracy and 49.57% F1-score for valence; 59.64% accuracy and 59.71% F1-score for arousal.
• Both 1-D and 2-D ECG input formats demonstrated comparable performance in emotion classification.

Conclusions:

• The study indicates that both 1-D and 2-D ECG data formats are viable for emotion recognition systems.
• The findings highlight the potential of utilizing both numerical and image-based ECG data for enhanced ERS.
• Further research can explore hybrid approaches or optimize feature extraction for each modality.