Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ECG-Agent: On-Device Tool-Calling Agent for ECG Multi-Turn Dialogue

Hyunseung Chung, Jungwoo Oh, Daeun Kyung, Jiho Kim, Yeonsu Kwon, Min-Gyu Kim, Edward Choi

TL;DR

ECG-MTD tackles the gap in on-device ECG understanding by enabling multi-turn dialogues with precise measurements. The authors introduce ECG-Agent, a tool-calling LLM, and ECG-MTD, a diverse, realistic dataset for training and evaluating such agents across lead configurations. Results show that tool-based ECG-Agents outperform baseline ECG-LLMs in response quality, while compact on-device variants (1B–3B) achieve performance close to larger models, highlighting practical viability for wearables and smartphones. This work advances private, accurate ECG interpretation through on-device, interactive AI systems.

Abstract

Recent advances in Multimodal Large Language Models have rapidly expanded to electrocardiograms, focusing on classification, report generation, and single-turn QA tasks. However, these models fall short in real-world scenarios, lacking multi-turn conversational ability, on-device efficiency, and precise understanding of ECG measurements such as the PQRST intervals. To address these limitations, we introduce ECG-Agent, the first LLM-based tool-calling agent for multi-turn ECG dialogue. To facilitate its development and evaluation, we also present ECG-Multi-Turn-Dialogue (ECG-MTD) dataset, a collection of realistic user-assistant multi-turn dialogues for diverse ECG lead configurations. We develop ECG-Agents in various sizes, from on-device capable to larger agents. Experimental results show that ECG-Agents outperform baseline ECG-LLMs in response accuracy. Furthermore, on-device agents achieve comparable performance to larger agents in various evaluations that assess response accuracy, tool-calling ability, and hallucinations, demonstrating their viability for real-world applications.

ECG-Agent: On-Device Tool-Calling Agent for ECG Multi-Turn Dialogue

TL;DR

ECG-MTD tackles the gap in on-device ECG understanding by enabling multi-turn dialogues with precise measurements. The authors introduce ECG-Agent, a tool-calling LLM, and ECG-MTD, a diverse, realistic dataset for training and evaluating such agents across lead configurations. Results show that tool-based ECG-Agents outperform baseline ECG-LLMs in response quality, while compact on-device variants (1B–3B) achieve performance close to larger models, highlighting practical viability for wearables and smartphones. This work advances private, accurate ECG interpretation through on-device, interactive AI systems.

Abstract

Recent advances in Multimodal Large Language Models have rapidly expanded to electrocardiograms, focusing on classification, report generation, and single-turn QA tasks. However, these models fall short in real-world scenarios, lacking multi-turn conversational ability, on-device efficiency, and precise understanding of ECG measurements such as the PQRST intervals. To address these limitations, we introduce ECG-Agent, the first LLM-based tool-calling agent for multi-turn ECG dialogue. To facilitate its development and evaluation, we also present ECG-Multi-Turn-Dialogue (ECG-MTD) dataset, a collection of realistic user-assistant multi-turn dialogues for diverse ECG lead configurations. We develop ECG-Agents in various sizes, from on-device capable to larger agents. Experimental results show that ECG-Agents outperform baseline ECG-LLMs in response accuracy. Furthermore, on-device agents achieve comparable performance to larger agents in various evaluations that assess response accuracy, tool-calling ability, and hallucinations, demonstrating their viability for real-world applications.
Paper Structure (11 sections, 2 figures, 5 tables)

This paper contains 11 sections, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Method
  3. ECG-MTD Dataset
  4. ECG-Agent
  5. Experiments
  6. Experimental Settings
  7. Training Setup
  8. Evaluation
  9. Experimental Results
  10. Conclusion
  11. Acknowledgement

Figures (2)

  • Figure 1: An example of a single conversational turn from a multi-turn dialogue in the ECG-MTD test set. The baselines output inaccurate responses to a user inquiry about heart rate, while the ECG-Agent (1B) produces a precise, measurement tool-based response.
  • Figure 2: (Left) Overall pipeline of ECG-MTD dataset construction. Dialogue Scenario defined by seven topic categories acquired from online medical consultation datasets and three English proficiency levels from the official CEFR guidelines. Action sequences constructed from pre-defined user and agent actions. A topic category, a user CEFR level, and an action sequence sampled to prompt Gemini-2.5-Flash to generate ECG multi-turn dialogue. (Right) Training framework of ECG-Agent. User query is processed to the ECG-Agent, where the agent is able to use three different tools to generate a response to the user. This process is repeated in multiple turns during multi-turn dialogue.