Diff-TGEG: a novel diffusion model for text-guided electrocardiogram generation

Cardiovascular diseases (CVDs) represent a significant public health challenge and rank among the principal causes of death globally, making early detection and effective intervention imperative. In this context, the electrocardiogram (ECG) stands as a cornerstone non-invasive methodology, integral to the screening, diagnostic evaluation, and ongoing monitoring of CVDs. However, the scarcity of clinical data constrains the development of automated ECG diagnostic technologies. Therefore, generating ECG that meet clinical standards has become a critical upstream task to drive breakthroughs in this field. This paper proposes a Text-Guided ECG Generation model based on the Diffusion model, named Diff-TGEG. By integrating feature information from clinical text reports as conditional inputs into the diffusion model via an adaptive layer normalization module, the proposed model generates high-fidelity 12-lead ECG that align with the input clinical text. Comparative experiments conducted on the PTB-XL dataset demonstrate that Diff-TGEG consistently outperforms existing methods in terms of ECG generation quality, verifying its effectiveness.