Interbeat Interval Detection from Synthetic Photoplethysmography Signals
No Thumbnail Available
Van Zaen, Jérôme
Introduction: Photoplethysmography (PPG) is a low-cost, non-invasive and convenient technology that can be embedded in wearable devices. It is a promising alternative to electrocardiography (ECG) for monitoring of physiological parameters related to cardiovascular diseases. PPG can typically be used to extract interbeat intervals (IBIs) and retrieve related information such as heart rate variability. However, as ECG is the gold standard for IBI extraction, ECG datasets are much more abundant than PPG datasets. Consequently, a thorough evaluation of PPG-based methods to estimate IBIs is more difficult compared to ECG-based methods. To facilitate the evaluation of PPG-based methods, we propose a generative adversarial network (GAN) to artificially generate realistic PPG signals from ECG data. Method: We developed a GAN architecture to generate a PPG signal from a sequence of spikes located at R-peak positions. This cross-modality signal-to-signal translation approach is based on the CycleGAN proposed by Zhu et al. (2017). The model takes as input a 10-second spike sequence derived from ECG R- peaks and generates the corresponding PPG signal. The main challenge is not only to synthetize a realistic looking PPG signal but to preserve its consistency with the ECG signal. For evaluation, a PPG-based IBI detection algorithm was separately applied to the real and artificially-generated PPG signals and the extracted IBIs are compared to IBIs derived from the corresponding ECG. Result and conclusion: The first results were obtained. Visual inspections show a high similarity between the artificially-generated and real PPG signals and the model seems to have integrated the information from the spike sequence derived from the ECG R-peaks. When compared to ECG-derived IBIs, the error of the PPG-based IBIs detection algorithm on the real dataset is of 0.06 ± 60.25 ms and 20.80 ± 221.99 ms on the artificially-generated dataset.
CinC 2021, Brno (CZ)