# Transfer Knowledge from Natural Language to Electrocardiography: Can We Detect Cardiovascular Disease Through Language Models?

**Authors:** Jielin Qiu, William Han, Jiacheng Zhu, Mengdi Xu, Michael Rosenberg, Emerson Liu, Douglas Weber, Ding Zhao  
**arXiv:** [2301.09017](https://arxiv.org/abs/2301.09017)  
**Date:** 2023-01-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2301.09017v2/

## Key points

- Recent advancements in Large Language Models (LLMs) have gained attention for their ability to learn powerful embeddings from large-scale datasets
- LLMs have been effective in various downstream applications
- The authors propose a novel approach that uses LLMs to enhance cardiovascular disease diagnosis and automatic Electrocardiography (ECG) report generation
- The approach includes an additional loss function using Optimal Transport (OT) to align the distribution between ECG data and language embeddings
- The learned embeddings are evaluated on automatic ECG diagnosis report generation and zero-shot cardiovascular disease detection tasks
- Results show high-quality cardiac diagnosis reports and competitive performance in zero-shot classification compared to supervised baselines
- This work demonstrates the feasibility of transferring knowledge from LLMs to clinical cardiology
- Utilizing LLMs in clinical Electrocardiography can improve cardiovascular disease diagnosis and report generation

## Summary

Recent advancements in Large Language Models (LLMs) have garnered significant attention due to their ability to learn powerful embeddings from large-scale datasets, which have proven effective in various downstream applications. However, it remains unknown whether the knowledge acquired by LLMs can be transferred to the field of clinical cardiology. To address this gap, the authors propose a novel approach that leverages LLMs to enhance cardiovascular disease diagnosis and automatic Electrocardiography (ECG) report generation. The authors introduce an additional loss function using Optimal Transport (OT) to align the distribution between ECG data and language embeddings. By doing so, they aim to facilitate the transfer of knowledge from LLMs to the cardiac domain. The learned embeddings are evaluated on two important tasks: automatic ECG diagnosis report generation and zero-shot cardiovascular disease detection. The results demonstrate that their approach not only generates high-quality cardiac diagnosis reports but also achieves competitive performance in zero-shot classification compared to supervised baselines. This indicates the feasibility of transferring knowledge from LLMs to clinical cardiology. Overall, this work highlights the potential of utilizing LLMs in clinical Electrocardiography for improving cardiovascular disease diagnosis and report generation. It contributes to bridging the gap between natural language processing and cardiology, opening up new avenues for research and development in this domain.