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 the 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 (1B, 3B) to larger agents (8B, 32B). 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.

For pre-processing PhysioNet2021 and pretraining W2V+CMSC+RLM, follow the instructions in:

• Fairseq-signals repository (uni-modal tasks section)
• W2V+CMSC+RLM guide

Step 2a: Preprocess the PTB-XL dataset:

python src/preprocess/preprocess_ptbxl.py \
    /path/to/ptbxl/records500/ \
    --dest /path/to/output

Step 2b: Generate manifest files for fine-tuning:

python src/preprocess/manifest_ptbxl_10s.py /path/to/output

Step 2c: Fine-tune on the Cardiac Arrhythmia Classification task:

fairseq-hydra-train \
    model.model_path=/path/to/pretrained_model/checkpoints/checkpoint_last.pt \
    +task.data=/path/to/output/from/2b/ptbxl_10s_manifest \
    --config-dir examples/w2v_cmsc/config/finetuning/ecg_transformer \
    --config-name diagnosis

For explanation tool outputs, follow the instructions in: SpectralX repository

After fine-tuning, use the model to produce tool output CSV files.

Classification (requires model path):

python extract_tool_outputs.py \
    --tool classification \
    --ecg_dir /path/to/ecg/files \
    --model_path /path/to/checkpoint_best.pt \
    --output_dir ./results

Measurements (no model needed):

python extract_tool_outputs.py \
    --tool measurements \
    --ecg_dir /path/to/ecg/files \
    --output_dir ./results

python finetune_ecg_dialogue_unsloth.py

python inference_ecg_dialogue.py

If you find this work useful, please cite our paper:

@misc{chung2026ecgagentondevicetoolcallingagent,
      title={ECG-Agent: On-Device Tool-Calling Agent for ECG Multi-Turn Dialogue},
      author={Hyunseung Chung and Jungwoo Oh and Daeun Kyung and Jiho Kim and Yeonsu Kwon and Min-Gyu Kim and Edward Choi},
      year={2026},
      eprint={2601.20323},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.20323},
}