PHYSIGEN: Physiological signal generation for class imbalance mitigation based on generative AI and machine learning

The study of emotion recognition is quite popular in recent years due to the impact of emotions on human behavior and social interactions. Understanding and identifying emotions has become very crucial nowadays because it influences decision-making, communication, and relationships. Emotion recognition can be performed in two different ways—unimodal or multimodal, depending on the number of physiological signals used. In this work, a multimodal approach has been adopted to classify emotions in four quadrants of the valence–arousal plane. This study uniquely compares synthetic minority over-sampling technique (SMOTE) and conditional generative adversarial network (CTGAN) for multimodal physiological emotion recognition and introduces a class-conditional CTGAN strategy that enhances minority-class sample diversity. The physiological signals that have been used are ECG, EEG, and Galvanic Skin Response (GSR), taken from the ASCERTAIN dataset, which is inherently class imbalanced. To address the class imbalance issue, data augmentation techniques like SMOTE and CTGAN are used to balance the dataset. The study evaluates the performance of Decision Tree (DTree), support vector machine (SVM), logistic regression (LR), linear discriminant analysis (LDA), and k-Nearest Neighbors (kNN) in emotion classification. It is observed that CTGAN-based augmentation improved SVM accuracy from 46.8% to 71.74%, while recognition of the minority class HAHV increased from 3.8% (original) to 47.2% (CTGAN). Similar improvements were observed across LR and LDA, demonstrating that generative adversarial network (GAN)-based synthesis significantly enhances minority-class detection.