DiffuSETS: 12-lead ECG Generation Conditioned on Clinical Text Reports and Patient-Specific Information

Jiabo Chen, Yongfan Lai, Deyun Zhang, Yue Wang, Shijia Geng, Hongyan Li, Shenda Hong

Keywords: Cardiology, Electrocardiogram, Signal processing, ECG generation, Diffusion models

Abstract: Heart disease poses a serious threat to human health. As a non-invasive diagnostic tool, the electrocardiogram (ECG) is one of the most commonly used methods for cardiac screening. Obtaining a large number of real ECG samples often entails high costs, and releasing hospital data also necessitates consideration of patient privacy. Due to the shortage of medical resources, precisely annotated ECG data are scarce. In the critical task of generating ECGs, work on generating ECGs from text is extremely rare. Given the differing data modalities, incorporating patient-specific information into the generation process is also challenging. To address these challenges, we propose DiffuSETS, the first method to use a diffusion model architecture for text-to-ECG. Our method can accept various modalities of clinical text reports and patient-specific information as inputs and generates ECGs with high semantic alignment and fidelity. In response to the lack of benchmarking methods in the ECG generation field, we also propose a comprehensive evaluation method to test the effectiveness of ECG generation. Our model achieve excellent results in tests, further proving its superiority in the task of text-to-ECG. Our code and trained models will be released after the acceptance of our paper.

Submission Number: 26