Diffusion-Based PPG-to-ECG Translation with Dynamic ROI Selection

Electrocardiography (ECG) is one of the simplest and fastest methods for examining the cardiovascular system. Despite its widespread use in assessing cardiac health, collecting ECG data can be a challenge, often requiring the attachment of multiple sensors to the body or the use of specialized equipment. To overcome these limitations, many research has been conducted to find alternative approaches. One promising avenue is the use of photoplethysmography (PPG) data, which has a high correlation with ECG data and can be easily collected in real-time through wearable devices. Consequently, there has been growing interest in using PPG data to generate ECG data for real-time cardiac analysis. In this study, we propose a diffusion-based PPG-to-ECG translation with dynamic region of interest (ROI) selection, using the diffusion model success in various domains, including image, video, audio, and time-series data. Our proposed methodology addresses a key limitation of traditional diffusion models, which struggle to accurately generate the QRS complex which is an essential component of ECG data. We enhance the model performance by selectively adding noise to the ROI, focusing on the most critical segments. Instead of relying on hyperparameters for ROI selection, we develop a dynamic ROI selection algorithm. Furthermore, instead of using a standard U-Net architecture commonly used in image processing, we propose using a diffusion model suitable for time-series data, DiffWave, enabling more precise and detailed extraction of signal characteristics. To validate the performance of our model, we use a benchmark comprising five bio-signal datasets. Experimental results show that, compared with the baseline RDDM, our method reduced the average RMSE from 0.533 to 0.212 and FD from 129.88 to 14.28 across five benchmark datasets.