Electrocardiogram Synthesis

Electrocardiogram (ECG) synthesis

We proposed an AI-based method to synthesize ECGs from various cardiac parameters that represent the internal state of the human body and affect to ECG generation.

Application

The pair data of cardiac parameters and ECGs obtained by the proposed method are useful in many situations.

• Information retrieval by comparing a personal ECG and pre-generated ECGs as shown in the figure below.
• Generation of a large amount of data required for training AI (e.g., Cardiac parameter estimation from ECGs).

Cardiac parameter-aware ECG synthesis

Most existing methods focus only on the ECG waveform shapes. For example, some use ODEs* to represent the ECG waveform shape. Other DNN-based methods synthesize ECGs from shape features (e.g., timing and voltage of R positions). By contrast, the proposed method can synthesize ECGs from cardiac parameters representing various factors that contribute to the actual ECG generative process.

*ODE … ordinary differential equation

Fast ECG synthesis

The existing ECG synthesis method uses the finite element method to rigorously simulate the ECG generation process, which consumes supercomputer-level computational costs. By contrast, the proposed method can synthesize ECGs in a few seconds by using a conditional VAE to simulate the ECG synthesis process of the existing method.

We verified that the proposed VAD could reproduce the ground-truth ECGs. As shown in the figure, the waveform contours of the reproduced ECGs were close to those of the ground-truth ECGs. The Q, R, S, and T positions, which are clinically important ECG features, were also nearly accurate, except for the estimation errors caused by the external tool.

We will verify the proposed method in a clinical environment. In addition, we aim to establish technology to simulate the future states of the whole human body, thereby realizing the Bio Digital Twin to improve patient care and well-being.

1. Ryo Nishikimi, Masahiro Nakano, Kunio Kashino, Shingo Tsukada, “Variational Autoencoder-based Neural Electrocardiogram Synthesis Trained by FEM-Based Heart Simulator, ” Cardiovascular Digital Health Journal, Vol 5, Issue 1, pp. 19-28, February 2024. [Link]

Ryo Nishikimi
Biomedical Informatics Research Group, Media Information Laboratory, NTT Communication Science Laboratories