Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation
Authors: Debaditya Shome, Pritam Sarkar, Ali Etemad
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Quantitative experiments demonstrate that RDDM can generate high-fidelity ECG from PPG in as few as 10 diffusion steps, making it highly effective and computationally efficient. Additionally, to rigorously validate the usefulness of the generated ECG signals, we introduce Cardio Bench, a comprehensive evaluation benchmark for a variety of cardiac-related tasks including heart rate and blood pressure estimation, stress classification, and the detection of atrial fibrillation and diabetes. Our thorough experiments show that RDDM achieves state-of-the-art performance on Cardio Bench. |
| Researcher Affiliation | Academia | Debaditya Shome , Pritam Sarkar , Ali Etemad Queen s University, Canada {debaditya.shome, pritam.sarkar, ali.etemad}@queensu.ca |
| Pseudocode | Yes | Algorithm 1: RDDM Training and Algorithm 2: RDDM Sampling |
| Open Source Code | Yes | We make the code public to the research community1. 1https://github.com/Debaditya QU/RDDM |
| Open Datasets | Yes | WESAD (Schmidt et al. 2018) comprises approximately 24 hours of synchronized Lead-II ECG data (sampled at 700Hz) and PPG data (sampled at 64 Hz) from 15 subjects with annotations of stress and affect states. MIMIC AFib (Bashar et al. 2019) is a subset of the MIMIC-III waveform database (Johnson et al. 2016) with binary labels for AFib. PPG-BP (Liang et al. 2018) consists of 657 pre-windowed PPG signals from 219 subjects... Cuffless BP (Kachuee et al. 2015) is a dataset derived from the MIMIC-II waveform database (Saeed et al. 2011). CAPNO (Karlen 2021) comprises approximately 5.6 hours of paired Lead-II ECG and PPG signals sampled at 300 Hz. BIDMC (Pimentel et al. 2016) is a dataset collected from 53 ICU patients... DALIA (Reiss et al. 2019) is a dataset with approximately 35 hours of synchronized PPG and Lead-II ECG signals, recorded from 15 subjects performing daily-life activities like walking and driving. |
| Dataset Splits | No | In particular, we use the data from 80% of the subjects for training and the remaining 20% of subjects are used for cross-subject evaluation. |
| Hardware Specification | Yes | We train RDDM on 4 NVIDIA A100 GPUs with a batch size of 512. We use a single Nvidia 2080Ti GPU for this experiment. |
| Software Dependencies | No | The paper mentions software like Adam W, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | We choose Adam W (Loshchilov and Hutter 2018) as the optimizer with a Cosine learning rate scheduler, having a base learning rate of 10 4 for a total of 500 epochs. We use a fixed linear variance scheduler with β (0.0001, 0.2). We empirically set the sampling steps to 10 and the ROI window size to Γ = 32 (see related study in Sec. 5.5). The loss coefficients in Equation 9 are set to λ1 = 100 and λ2 = 1. |