ECG-Based Human Emotion Recognition Using Generative Models

Recent advancements in Human Emotion Recognition (HER) technology can enable active and assisted living systems to respond more intuitively to the emotional needs of users, enhancing their overall quality of life. This chapter presents a comprehensive study using only Electrocardiogram (ECG) data for the recognition of human emotions. We introduce three distinct processing pipelines to assess the effectiveness of generative models in recognizing human emotions. Each pipeline incorporates the Pan-Tompkins algorithm for signal denoising and resampling. The first pipeline uses the output of the Pan-Tompkins algorithm to train a Support Vector Classifier (SVC). The second pipeline integrates an Autoencoder for feature extraction before the SVC training phase. The third pipeline incorporates a Sparse Autoencoder to encode signals and feed the resulting sparse latent space into an SVC. Our experimental results, based on the MAHNOB and ASCERTAIN datasets show promising results of the Autoencoder pipeline for feature extraction, particularly in reducing model bias towards specific emotional classes. Although the Sparse Autoencoder pipeline exhibits a slight decrease in overall accuracy, it makes a valuable contribution in minimizing class bias.