f-GAN: A frequency-domain-constrained generative adversarial network for PPG to ECG synthesis

TL;DR

The paper addresses synthesizing ECG signals from wearable PPG data using GANs to enable continuous cardiovascular monitoring when ECGs are impractical. It compares a baseline GAN with a frequency-domain constraint, showing that the frequency-regularized model achieves more stable training, fewer heart-rate estimation failures, and improved HR accuracy on held-out subjects. The approach demonstrates that subject-agnostic PPG-to-ECG synthesis can yield ECG morphologies with actionable HR information, outperforming direct PPG-based HR estimation in various daily-activity scenarios. This work advances wearable biosignal integration by providing a GAN-based pathway to translate accessible PPG data into ECG-derived insights with practical clinical relevance.

Abstract

Electrocardiograms (ECGs) and photoplethysmograms (PPGs) are generally used to monitor an individual's cardiovascular health. In clinical settings, ECGs and fingertip PPGs are the main signals used for assessing cardiovascular health, but the equipment necessary for their collection precludes their use in daily monitoring. Although PPGs obtained from wrist-worn devices are susceptible to noise due to motion, they have been widely used to continuously monitor cardiovascular health because of their convenience. Therefore, we would like to combine the ease with which PPGs can be collected with the information that ECGs provide about cardiovascular health by developing models to synthesize ECG signals from paired PPG signals. We tackled this problem using generative adversarial networks (GANs) and found that models trained using the original GAN formulations can be successfully used to synthesize ECG signals from which heart rate can be extracted using standard signal processing pipelines. Incorporating a frequency-domain constraint to model training improved the stability of model performance and also the performance on heart rate estimation.

Figures (5)

• Figure 1: Schematics of the generator and the discriminator architectures. "AG" denotes the attention gate when combining features from the encoder with the features from the decoder. Left: architecture of the generator, which takes as input the PPG signal. Right: architecture of the discriminator, which takes as input the real or the synthetic ECG signal.
• Figure 2: Example of a synthesized ECG signal from a PPG signal using a GAN trained with the frequency-domain constraint. The PPG signal (top) was used as the input to the generator, which output a synthetic ECG signal (middle). The synthetic ECG signal looks qualitatively similar to the real (paired) ECG signal. Signal amplitudes shown here were obtained after signal preprocessing steps were performed.
• Figure 3: Distribution across random seeds of model performance on heart rate estimation during all activities for the two objective functions. Across the entire validation set, the model incorporating the frequency-domain constraint performs better (on average across seeds) than the model without the constraint ($t(60) = 3.04, p = 0.005$) and its performance is less variable across random seeds.
• Figure 4: Distribution across random seeds of model performance on heart rate estimation during "active" activities for the two objective functions. When evaluated on signal segments during staircase traversal, table soccer, cycling, driving, walking and working, the model incorporating the frequency-domain constraint has better performance (on average across seeds) than the model without the constraint ($t(60) = 2.98, p = 0.005$). Incorporating the constraint also reduces performance variance.
• Figure 5: Distribution across random seeds of total number of samples where heart rate estimation failed during all activities for the two objective functions. Across the entire validation set (a total of $11276.0$ samples), the model incorporating the frequency-domain constraint results in less samples in which heart rate estimation fails (on average across seeds) than the model without the constraint ($t(60) = 2.97, p = 0.006$), while using existing signal processing pipelines.