Generating Realistic Multi-Class Biosignals with BioSGAN: A Transformer-Based Label-Guided Generative Adversarial Network

Time series data are commonly used in biomedical applications, but such datasets are often small, expensive to collect, and may involve privacy issues that restrict large-scale deep learning models. Data augmentation techniques for time series data are limited by the need to maintain signal properties, but Generative Adversarial Networks (GANs) offer a promising approach for expanding datasets. This paper presents BioSGAN, a transformer-based label-guided GAN model capable of generating multi-class, class-specific synthetic time-series sequences of arbitrary length. Our proposed model architecture and design strategies produce synthetic sequences that are almost indistinguishable from real signals, enabling data augmentation. To evaluate the quality of the generated data, we propose a wavelet coherence metric that compares the similarity of real and synthetic signals. Our results show that BioSGAN outperforms existing state-of-the-art time-series GAN models, as demonstrated by qualitative visualizations using PCA and t-SNE, as well as quantitative comparisons of the discriminative and predictive power of the synthetic data. BioSGAN source code: https://github.com/imics-lab/BioSGAN