Continuous multi-step ECG and PPG generation through latent disentanglement

Artificial data synthesis is a valuable tool for overcoming the challenges of collecting large datasets, particularly in the medical domain where privacy concerns are significant. However, synthesising complex biomedical signals remains difficult due to their intricate morphologies and subject-specific constraints. While GANs have shown promise, existing methods have largely been limited to short-duration outputs or morphology-centered single-pulse events. In this study, we propose a generative framework combining autoencoders and adversarial learning-based disentanglement to enable multi-step, subject-specific biosignal generation with transferable properties, evaluated under a within-recording, segment-level protocol. We evaluate our model on two modalities: Electrocardiogram (ECG; 17 subjects) and Photoplethysmogram (PPG; 7 subjects). The model generates continuous ECG sequences of approximately 6 s (two forward recurrent steps beyond the seed segment), with a subset of subjects retaining recognisable waveform structure up to approximately 8 s. It shows similar performance for PPG, suggesting adaptability across periodic biosignal modalities. With recurrent subject-conditional generation demonstrated across both ECG and PPG, and exploratory style-transfer capabilities demonstrated for selected ECG subjects, the proposed framework serves as a proof of concept for multi-step sequential biosignal generation.