Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

anyECG-chat: A Generalist ECG-MLLM for Flexible ECG Input and Multi-Task Understanding

Haitao Li, Ziyu Li, Yiheng Mao, Ziyi Liu, Zhoujian Sun, Zhengxing Huang

TL;DR

anyECG-chat introduces a generalist ECG-MLLM capable of handling dynamic-length, reduced-lead, and multi-ECG inputs to perform diverse tasks including report generation, fine-grained localization, and multi-ECG comparison. It relies on the anyECG dataset family and a three-stage curriculum that progressively aligns ECG perception with instruction-following in a LLaMA-based model augmented by LoRA adapters. The architecture couples a pre-trained ECG encoder with a modality connector and strategic input tokens to support dynamic ECG inputs, while the evaluation demonstrates robust cross-task performance, including zero-shot and multi-turn capabilities. This work provides a practical framework for flexible, interactive ECG understanding with potential impact on clinical workflow and home-monitoring contexts.

Abstract

The advent of multimodal large language models (MLLMs) has sparked interest in their application to electrocardiogram (ECG) analysis. However, existing ECG-focused MLLMs primarily focus on report generation tasks, often limited to single 12-lead, short-duration (10s) ECG inputs, thereby underutilizing the potential of MLLMs. To this end, we aim to develop a MLLM for ECG analysis that supports a broader range of tasks and more flexible ECG inputs. However, existing ECG-QA datasets are often monotonous. To address this gap, we first constructed the anyECG dataset, which encompasses a wide variety of tasks, including report generation, abnormal waveform localization, and open-ended question answering. In addition to standard hospital ECGs, we introduced long-duration reduced-lead ECGs for home environments and multiple ECG comparison scenarios commonly encountered in clinical practice. Furthermore, we propose the anyECG-chat model, which supports dynamic-length ECG inputs and multiple ECG inputs. We trained the model using a three-stage curriculum training recipe with the anyECG dataset. A comprehensive evaluation was conducted, demonstrating that anyECG-chat is capable of supporting various practical application scenarios, including not only common report generation tasks but also abnormal waveform localization for long-duration reduced-lead ECGs in home environments and comprehensive comparative analysis of multiple ECGs. Our code and data are available at: https://github.com/CuCl-2/anyECG-chat.

anyECG-chat: A Generalist ECG-MLLM for Flexible ECG Input and Multi-Task Understanding

TL;DR

anyECG-chat introduces a generalist ECG-MLLM capable of handling dynamic-length, reduced-lead, and multi-ECG inputs to perform diverse tasks including report generation, fine-grained localization, and multi-ECG comparison. It relies on the anyECG dataset family and a three-stage curriculum that progressively aligns ECG perception with instruction-following in a LLaMA-based model augmented by LoRA adapters. The architecture couples a pre-trained ECG encoder with a modality connector and strategic input tokens to support dynamic ECG inputs, while the evaluation demonstrates robust cross-task performance, including zero-shot and multi-turn capabilities. This work provides a practical framework for flexible, interactive ECG understanding with potential impact on clinical workflow and home-monitoring contexts.

Abstract

The advent of multimodal large language models (MLLMs) has sparked interest in their application to electrocardiogram (ECG) analysis. However, existing ECG-focused MLLMs primarily focus on report generation tasks, often limited to single 12-lead, short-duration (10s) ECG inputs, thereby underutilizing the potential of MLLMs. To this end, we aim to develop a MLLM for ECG analysis that supports a broader range of tasks and more flexible ECG inputs. However, existing ECG-QA datasets are often monotonous. To address this gap, we first constructed the anyECG dataset, which encompasses a wide variety of tasks, including report generation, abnormal waveform localization, and open-ended question answering. In addition to standard hospital ECGs, we introduced long-duration reduced-lead ECGs for home environments and multiple ECG comparison scenarios commonly encountered in clinical practice. Furthermore, we propose the anyECG-chat model, which supports dynamic-length ECG inputs and multiple ECG inputs. We trained the model using a three-stage curriculum training recipe with the anyECG dataset. A comprehensive evaluation was conducted, demonstrating that anyECG-chat is capable of supporting various practical application scenarios, including not only common report generation tasks but also abnormal waveform localization for long-duration reduced-lead ECGs in home environments and comprehensive comparative analysis of multiple ECGs. Our code and data are available at: https://github.com/CuCl-2/anyECG-chat.

Paper Structure

This paper contains 33 sections, 1 equation, 5 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. AnyECG Dataset
  4. AnyECG-chat Architecture
  5. ECG encoder
  6. Large Language Model.
  7. Modality Connector
  8. Dynamic ECG Input
  9. Training Recipe
  10. Experiments
  11. Report Generation
  12. Localization Task
  13. Results
  14. Zero-shot Single Lead Localization
  15. Multi-ECG Comparison
  16. ...and 18 more sections

Figures (5)

  • Figure 1: The overview of anyECG-chat architecture.
  • Figure 2: Results of Localization and Zero-Shot Single Lead ECG Localization. Since LLaVa-Med, MEIT and PULSE failed to provide second-level localization, scoring 0, they are omitted from the figure.
  • Figure 3: The Score Distribution on MIMIC Multi-ECG QA.
  • Figure 4: Statistics of the MIMIC Multi-ECG QA dataset.
  • Figure 5: Comparison of LLM-based scoring and human scoring across different score levels.