Towards Generating Realistic Wrist Pulse Signals Using Enhanced One Dimensional Wasserstein GAN

Area of Science:

• Computational medicine
• Artificial intelligence in healthcare
• Biomedical signal processing

Background:

• Deep learning methods are increasingly applied to computational pulse diagnosis.
• A significant challenge is the limited availability of wrist pulse datasets due to privacy and cost concerns.
• Data scarcity hinders the development and performance of computational pulse diagnosis algorithms.

Purpose of the Study:

• To address the data scarcity issue in computational pulse diagnosis by generating synthetic wrist pulse signals.
• To introduce a novel one-dimension generative adversarial network (GAN) for creating realistic wrist pulse data.
• To evaluate the effectiveness of the proposed GAN model against existing methods.

Main Methods:

• Utilized Wasserstein GAN with gradient penalty (WGAN-GP) to generate wrist pulse signals, mitigating mode collapse.
• Compared the performance of WGAN-GP with vanilla GAN, deep convolutional GAN (DCGAN), and Wasserstein GAN (WGAN).
• Trained the models using a dataset of real wrist pulse signals and evaluated using qualitative and quantitative metrics.

Main Results:

• WGAN-GP demonstrated superior performance compared to other GAN models.
• Quantitative metrics including maximum mean deviation (MMD), sliced Wasserstein distance (SWD), and percent root mean square difference (PRD) showed significant improvements.
• Achieved MMD as low as 0.2325, SWD as low as 0.0112, and PRD as low as 5.8748, indicating high-quality synthetic data generation.

Conclusions:

• The proposed WGAN-GP model successfully generates wrist pulse data from small ground truth datasets.
• This approach offers a viable solution to the data scarcity problem in computational pulse diagnosis research.
• The generated synthetic data is expected to enhance the performance of future pulse diagnosis algorithms.