Scholarlink

Background <p>Cardiovascular disease is a leading cause of death, accounting for nearly one-third of all fatalities worldwide. The electrocardiogram (ECG), an electrophysiological signal representing cardiac activity, is a widely used non-invasive method for diagnosing cardiovascular diseases. However, ECG has limitations for continuous monitoring due to sensor size requirements, including electrodes. The photoplethysmogram (PPG), an optical method measuring blood volume variations, offers a more convenient alternative but provides less detailed information about heart activity than ECG. We introduce a peak-oriented diffusion model for PPG-to-ECG reconstruction, utilizing a U-Net framework composed of an encoder and decoder.</p> Objective <p>We aim for the model to learn the peak differences between ECG and PPG by predicting R peaks from the systolic peaks of PPG, thereby preserving the main components of ECG signals. We evaluate the reconstructed ECG through clinical interpretation and assess it with noise input to verify model robustness under a daily environment.</p> Methods <p>An R peak predictor, connected to the U-Net bottleneck, estimates R peaks incorporated into the decoder. The R peak predictor enables the model to reconstruct the ECG from the PPG while learning peak differences characterized as pulse transit time. We trained and evaluated the model using the MIMIC-III, WESAD, and PPG-DaLiA datasets. We assessed the model performance using root-mean-squared error and Fréchet distance, along with the ECG interpretation. In addition, atrial fibrillation detection and heart rate estimation are performed to verify clinical use of the reconstructed ECG.</p> Results <p>There were 59,980,410 segmented samples from 5,888 patients, split into train and test sets based on each patient, with an 8:2 ratio. The proposed model achieved a root-mean-square-error of 0.220, a Fréchet distance of 6.456, an F1-score of 0.925 for atrial fibrillation detection, and a mean difference of −0.077 [−1.877, 1.723] beats/min for heart rate estimation.</p> Conclusions <p>Our results showed that the peak-oriented diffusion model translates PPG into high-fidelity ECG for clinical applications. We believe our findings expand the usefulness of the AI-based model to mobile healthcare, which involves continuous ECG monitoring in daily life.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A peak-oriented diffusion model for high-fidelity electrocardiogram reconstruction from photoplethysmogram: development and usability study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月6日

查看全文
  • Hyun-Myung Cho,
  • Sungmin Han,
  • Joon-Kyung Seong,
  • Inchan Youn

摘要

Background

Cardiovascular disease is a leading cause of death, accounting for nearly one-third of all fatalities worldwide. The electrocardiogram (ECG), an electrophysiological signal representing cardiac activity, is a widely used non-invasive method for diagnosing cardiovascular diseases. However, ECG has limitations for continuous monitoring due to sensor size requirements, including electrodes. The photoplethysmogram (PPG), an optical method measuring blood volume variations, offers a more convenient alternative but provides less detailed information about heart activity than ECG. We introduce a peak-oriented diffusion model for PPG-to-ECG reconstruction, utilizing a U-Net framework composed of an encoder and decoder.

Objective

We aim for the model to learn the peak differences between ECG and PPG by predicting R peaks from the systolic peaks of PPG, thereby preserving the main components of ECG signals. We evaluate the reconstructed ECG through clinical interpretation and assess it with noise input to verify model robustness under a daily environment.

Methods

An R peak predictor, connected to the U-Net bottleneck, estimates R peaks incorporated into the decoder. The R peak predictor enables the model to reconstruct the ECG from the PPG while learning peak differences characterized as pulse transit time. We trained and evaluated the model using the MIMIC-III, WESAD, and PPG-DaLiA datasets. We assessed the model performance using root-mean-squared error and Fréchet distance, along with the ECG interpretation. In addition, atrial fibrillation detection and heart rate estimation are performed to verify clinical use of the reconstructed ECG.

Results

There were 59,980,410 segmented samples from 5,888 patients, split into train and test sets based on each patient, with an 8:2 ratio. The proposed model achieved a root-mean-square-error of 0.220, a Fréchet distance of 6.456, an F1-score of 0.925 for atrial fibrillation detection, and a mean difference of −0.077 [−1.877, 1.723] beats/min for heart rate estimation.

Conclusions

Our results showed that the peak-oriented diffusion model translates PPG into high-fidelity ECG for clinical applications. We believe our findings expand the usefulness of the AI-based model to mobile healthcare, which involves continuous ECG monitoring in daily life.