Model

ECGJev

question-conditioned ECG tokens + 0.8B option-scoring decoder for 12-lead ECG QA/classification

Year
2026
Form factor
Clinical 12 lead
Li, Zhuoyuan et al., "ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation", arXiv:2611.02041, CC BY-SA 4.0
PreprintCatalog paper date: 2026Added to atlas: Nov 5, 2026Source

Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.

ECGJev reframes 12-lead ECG interpretation as multiple-choice option scoring. Frozen waveform and spectrogram encoders feed 16 question-conditioned ECG tokens into a 0.8B decision decoder (LoRA), which scores every answer option in one forward pass instead of generating free text.

Signals
ECG
Form factor
clinical-12-lead
Year
2026
Architecture

Frozen waveform + spectrogram ECG encoders → 16 question-conditioned ECG tokens → StartLux-Decision-0.8B option-scoring decoder (LoRA); one forward pass scores all answer options

Pretraining data

Uses frozen pretrained ECG encoders; fine-tuned with LoRA on benchmark train splits

License

arXiv perpetual non-exclusive license; Apache-2.0 (adapted StartLux-Decision NOTICE); CC BY-NC 4.0 (decoder weights)

At a glance
Type
FM adaptation
Modalities
ECG
Form factor
Clinical 12 lead
Parameters
Not reported
Architecture
Frozen waveform + spectrogram ECG encoders → 16 question-conditioned ECG tokens → StartLux-Decision-0.8B option-scoring decoder (LoRA); one forward pass scores all answer options
Venue
arXiv 2611.02041v1 (2026-11-01); submitted to ISBI 2027
Year
2026
DOI
Not reported
Weights
Not released
Licence
arXiv perpetual non-exclusive license; Apache-2.0 (adapted StartLux-Decision NOTICE); CC BY-NC 4.0 (decoder weights)
Caveats
Not reported
Added
Nov 5, 2026
Updated
Nov 5, 2026

Abstract

Clinical interpretation of a 12-lead electrocardiogram (ECG) often requires answering targeted diagnostic questions with clear, quantifiable decisions. We introduce ECGJev, the first ECG framework built on a decision decoder, a pretrained language model trained to return a probability for every listed option. A question-conditioned bridge compresses frozen waveform and spectrogram features into 16 ECG tokens, and a two-stage procedure adapts the bridge and low-rank adapters on the decoder's native option readout using diagnostic labels alone, scoring all options for a question in one forward pass. Evaluated across four ECG benchmarks, ECGJev achieves the highest macro-F1 on CPSC2018 (80.46%), PTB-XL (71.68%) and Chapman (96.86%), and generalizes competitively to the small-sample PTB cohort. Concept probes and perturbation analyses further show that its predictions rely directly on the ECG input, remain stable under option-order and wording variations, and transfer zero-shot to unseen waveform concepts. Our code is available at https://github.com/zhuoyuanli-lab/ECGJev.

Verbatim fromarXiv· arXiv metadata (abstracts are released under CC0)

Key results

How it works

How ECGJev works: Input signals → Model → Downstream tasksInput signals: ECG (Electrocardiogram). Datasets: CPSC2018 (China Physiological Signal Challenge 2018): 12-lead ECG arrhythmia set, redistributed in PhysioNet Challenge 2020, PTB-XL, Chapman-Shaoxing 12-lead ECG database, figshare release (Zheng 2020). Model: ECGJev (Language-aligned). Downstream tasks: ECG rhythm classification.Input signalsECG: ElectrocardiogramECGElectrocardiogramDatasets: CPSC2018 (China Physiological Signal Challenge 2018): 12-lead ECG arrhythmia set, redistributed in PhysioNet Challenge 2020, PTB-XL, Chapman-Shaoxing 12-lead ECG database, figshare release (Zheng 2020)DatasetsCPSC2018 (China Physiological Signal…ModelECGJevLanguage-alignedDownstream tasksECG rhythm classificationECG rhythm classification

Datasets linked in the hub; pretraining vs evaluation roles follow the paper.

Pretraining data
Uses frozen pretrained ECG encoders; fine-tuned with LoRA on benchmark train splits
Architecture
Frozen waveform + spectrogram ECG encoders → 16 question-conditioned ECG tokens → StartLux-Decision-0.8B option-scoring decoder (LoRA); one forward pass scores all answer options
Inputs
ECG
Approach
ECG · 12 lead · ECG QA · ECG LLM · option scoring · LoRA

Leaderboard rows· 1

All leaderboards
TaskDatasetMetricValueRank
ECG rhythm classificationChapman-Shaoxing 12-lead ECG database, figshare release (Zheng 2020)Macro-F1 ↑96.86#1 / 1

Datasets used· 2

Papers· 1

Authors· 7

Cite this model

Cite the publication recorded for this model. Check it against the publisher before submitting.

Li, Z., Wu, X., Cui, H., Gao, J., Fu, R., Yan, M., & Tang, H. (2026). ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation (arXiv 2611.02041). arXiv 2611.02041v1. https://arxiv.org/abs/2611.02041
Li, Zhuoyuan, Xidong Wu, Hanyu Cui, Jiansong Gao, Rongrong Fu, Mingke Yan, and Haotian Tang. 2026. "ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation (arXiv 2611.02041)." arXiv 2611.02041v1. https://arxiv.org/abs/2611.02041.
Li, Z., Wu, X., Cui, H., Gao, J., Fu, R., Yan, M. and Tang, H. (2026) 'ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation (arXiv 2611.02041)', arXiv 2611.02041v1. Available at: https://arxiv.org/abs/2611.02041
Li, Zhuoyuan, et al. "ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation (arXiv 2611.02041)." arXiv 2611.02041v1, 2026. https://arxiv.org/abs/2611.02041.
Z. Li et al., "ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation (arXiv 2611.02041)," arXiv 2611.02041v1, 2026, [Online]. Available: https://arxiv.org/abs/2611.02041.
@misc{li2026ecgjev,
  title         = {{ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation (arXiv 2611.02041)}},
  author        = {Li, Zhuoyuan and Wu, Xidong and Cui, Hanyu and Gao, Jiansong and Fu, Rongrong and Yan, Mingke and Tang, Haotian},
  year          = {2026},
  howpublished  = {arXiv 2611.02041v1},
  eprint        = {2611.02041},
  archivePrefix = {arXiv}
}

Source metadata: 2026

Record updated Nov 5, 2026

History

  1. Added to the ChatPPG research hub