On-Device LLM Agent Enables Multi‑Turn ECG Dialogue

Global: On-Device LLM Agent Enables Multi‑Turn ECG Dialogue

Researchers led by Hyunseung Chung and colleagues submitted a paper to arXiv on January 28, 2026, describing ECG‑Agent, the first large‑language‑model (LLM) based tool‑calling system designed for multi‑turn electrocardiogram (ECG) conversations. The work aims to address gaps in existing multimodal LLMs, which typically lack conversational depth, on‑device efficiency, and precise handling of ECG measurements such as P‑, Q‑, R‑, S‑, and T‑wave intervals.

Background and Motivation

Recent advances have extended multimodal LLMs to ECG classification, report generation, and single‑turn question answering. However, those models often require cloud resources and cannot sustain interactive, back‑and‑forth dialogues that clinicians might need during patient assessment.

Dataset Introduction

To support development and testing, the authors released the ECG‑Multi‑Turn‑Dialogue (ECG‑MTD) dataset, which comprises realistic user‑assistant exchanges covering a variety of ECG lead configurations. The dataset is intended to benchmark conversational ability and tool‑calling performance in a clinical context.

Model Architecture

ECG‑Agent was implemented in several scale variants, ranging from lightweight models suitable for on‑device deployment to larger models that run on more powerful hardware. Each variant incorporates a tool‑calling interface that enables the agent to retrieve or compute specific ECG measurements during a dialogue.

Performance Evaluation

Experimental results reported in the paper show that ECG‑Agents outperform baseline ECG‑LLMs in response accuracy. Moreover, the on‑device agents achieve performance comparable to their larger counterparts across metrics that assess answer correctness, tool‑calling success, and incidence of hallucinated information.

Potential Impact

The findings suggest that on‑device LLM agents could be integrated into clinical workflows, providing clinicians with immediate, interactive ECG analysis without reliance on external servers, thereby enhancing privacy and reducing latency.

Future Directions

The authors indicate plans to expand the ECG‑MTD dataset, refine the tool‑calling mechanisms, and explore integration with electronic health‑record systems to further validate real‑world applicability.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.