Validation of Electrocardiogram Based Photoplethysmogram Generated Using U-Net Based Generative Adversarial Networks.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiovascular Signal Processing

Background:

• Photoplethysmogram (PPG) is crucial for detecting atrial fibrillation (AF).
• A significant limitation is the scarcity of openly available AF PPG datasets.
• This data gap hinders the development and validation of AF detection algorithms.

Purpose of the Study:

• To develop a novel Generative Adversarial Network (GAN) for synthesizing realistic PPG signals from electrocardiogram (ECG) data.
• To evaluate the morphological similarity and physiological relevance of the generated PPG signals compared to reference data.
• To assess the utility of synthetic PPG data in augmenting real AF PPG datasets for improving AF classification model performance.

Main Methods:

• A U-net-based Generative Adversarial Network (GAN) was employed to synthesize PPG signals from paired ECG recordings.
• Morphological similarity was quantified using Percent Root Mean Square Difference (PRD) and Pearson correlation coefficient.
• Heart Rate Variability (HRV) analysis was performed to compare generated PPG with reference ECG.
• Classification models were trained using various combinations of real and generated AF PPG data to evaluate performance.

Main Results:

• Generated PPG signals exhibited high morphological similarity to reference PPG, with a mean PRD of 27% and Pearson correlation of 0.94.
• No significant difference in Heart Rate Variability (HRV) was observed between reference AF ECG and generated PPG (p=0.248).
• AF classification models trained with augmented datasets (real + generated PPG) achieved high test accuracy (0.962) and F1 scores (0.969), comparable to models trained solely on real data (accuracy 0.961).
• Models trained exclusively on generated AF PPG data demonstrated a test accuracy of 0.945, confirming the value of synthetic data.

Conclusions:

• The proposed GAN-based method successfully synthesizes physiologically relevant AF PPG signals from ECG.
• Generated AF PPG data can effectively augment limited real-world datasets.
• The synthetic data is valuable for training robust AF detection classifiers, addressing the challenge of data scarcity.