P2E-WGAN: ECG waveform synthesis from PPG with conditional wasserstein generative adversarial networks

Electrocardiogram (ECG) is routinely used to identify key cardiac events such as changes in ECG intervals (PR, ST, QT, etc.), as well as capture critical vital signs such as heart rate (HR) and heart rate variability (HRV). The gold standard ECG requires clinical measurement, limiting the ability to capture ECG in everyday settings. Photoplethysmography (PPG) offers an out-of-clinic alternative for non-invasive, low-cost optical capture of cardiac physiological measurement in everyday settings, and is increasingly used for health monitoring in many clinical and commercial wearable devices. Although ECG and PPG are highly correlated, PPG does not provide much information for clinical diagnosis. Recent work has applied machine learning algorithms to generate ECG signals from PPG, but requires expert domain knowledge and heavy feature crafting to achieve good accuracy. We propose P2E-WGAN: a pure end-to-end, generalizable deep learning model using a conditional Wasserstein generative adversarial network to synthesize ECG waveforms from PPG. Our generative model is capable of augmenting the training data to alleviate the data-hungry problem of machine learning methods. Our model trained in the subject independent mode can achieve the average root mean square error of 0.162, Fréchet distance of 0.375, and Pearson's correlation of 0.835 on a normalized real-world dataset, demonstrating the effectiveness of our approach.