ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation

Zhuoyuan Li, Xidong Wu, Hanyu Cui, Jiansong Gao, Rongrong Fu, Mingke Yan, Haotian TangarXiv 2026 · arXiv

TL;DR

ECGJev is a novel framework for 12-lead ECG interpretation that replaces messy free-text generation with a precise "decision decoder." By using a question-conditioned bridge to extract relevant features from both waveforms and spectrograms, it achieves state-of-the-art performance across multiple benchmarks and demonstrates an impressive ability to identify medical concepts it wasn't explicitly trained to recognize.

What Problem Does This Paper Address?

Interpreting an ECG is essentially a series of targeted diagnostic questions: "Is the heart rate regular?" or "Is there evidence of an old myocardial infarction?"

Current AI approaches generally fall into two camps, both with significant drawbacks. Classification heads are fast and accurate but rigid; they can only predict a fixed set of labels and can't handle new or complex textual queries. On the other hand, Generative ECG-LLMs are flexible enough to answer open questions, but they often "hallucinate" or get tripped up by how a prompt is worded. They also require complex post-processing to turn a paragraph of text into a simple "Yes" or "No" decision.

The authors of ECGJev sought a middle ground: a model that is as flexible as an LLM but as decisive and quantifiable as a classifier.

Methods

The core innovation of ECGJev is the use of a decision decoder—specifically a 0.8B parameter model called StartLux-Decision. Unlike standard LLMs that predict the next word in a sentence, this decoder is pre-trained to output probability distributions over specific options (like "True" vs. "False").

The architecture consists of three main parts:

  1. Dual-View Encoding: The model looks at the ECG in two ways simultaneously. It uses a frozen ECGFounder model for raw time-domain waveforms and a ViT-B/16 model for lead-wise spectrograms (frequency-domain).
  2. Question-Conditioned Bridge: This is the "brain" of the feature extraction. Instead of passing all ECG data to the decoder, a lightweight "resampler" uses the diagnostic question and clinical guidance to pick out the most relevant parts of the signal. It compresses hundreds of data tokens into just 16 "soft tokens" tailored to the specific query.
  3. Symbol-Decoupled Scoring: To prevent the model from simply guessing "True" because it likes the letter 'A', the authors use a technique that randomizes which letter corresponds to which answer during training. This forces the model to rely on the actual clinical meaning rather than position bias.
Fig. 1. ECGJev connects two views of the same recording to a diagnostic question: (a) frozen waveform and spectrogram-image encoders; (b) question-conditioned compression into 16 ECG tokens; (c) direct option scoring; and (d) description alignment followed by the disease-question loss. The clinical guidance is a textual description of the queried disease. Token counts illustrate CPSC2018.

Datasets

The model was evaluated on four major ECG benchmarks:

  • CPSC2018: 9 diagnostic labels.
  • PTB-XL: 5 diagnostic superclasses.
  • Chapman: 4 mutually exclusive rhythm classes.
  • PTB: A smaller dataset with 10 shared diagnostic codes, used to test the model's ability to generalize with limited data.

Key Insights and Ideas

  • Conditioning is Key: The ablation study showed that compressing the ECG signal based on the question significantly improves performance. It allows the model to focus on, for example, the P-wave when asking about atrial rhythm while ignoring it when asking about ventricular morphology.
  • Bypassing the "Textual Detour": By scoring options directly, ECGJev avoids the "lexical frequency priors" that plague generative models (where a model might say "Yes" simply because "Yes" is a common word in its training data).
  • Zero-Shot Medical Reasoning: Perhaps the most exciting finding is that the model can identify concepts it wasn't trained on. Because it understands the textual description of a condition (e.g., "ST depression"), it can look for those features in the ECG signal even if it never saw a label for "ST depression" during training.

Key Results

ECGJev outperformed both specialized classification heads and much larger generative LLMs across the board:

  • Superior Accuracy: It achieved the highest Macro-F1 scores on CPSC2018 (80.46%), PTB-XL (71.68%), and Chapman (96.86%).
  • Robustness on Small Data: On the small-sample PTB dataset, where generative LLMs often collapse and just guess the most frequent label, ECGJev maintained a respectable 35.02% F1 score—more than double its competitors.
  • Zero-Shot Success: As shown in the concept probing results, the model achieved high AUROC scores (up to 0.895) for unseen concepts like ST depression and inverted T waves, proving it actually "understands" the relationship between medical text and cardiac signals.
Table 1. Classification comparison across four ECG benchmarks (%, mean ± SD over three runs). F1 denotes macro-F1; Acc. denotes label-wise binary accuracy on multilabel tasks and four-class accuracy on Chapman. Bold/underline: best/second-best score per column.

Limitations and Future Work

While the model is highly effective, the authors noted that its sensitivity is still lower for very rare conditions, such as ST elevation (which had only 4 positive instances in the test set).

Future work will focus on refining how clinical queries are formulated and strengthening the model's ability to judge highly detailed feature descriptions. The goal is to move toward even more reliable, "interpretable" diagnostic decision-making that clinicians can trust in a high-stakes environment.

Summary updated

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.
BibTeX
@misc{li2026ecgjev,
  author = {Zhuoyuan Li and Xidong Wu and Hanyu Cui and Jiansong Gao and Rongrong Fu and Mingke Yan and Haotian Tang},
  title = {{ECGJev: Question-Conditioned Option Scoring for 12-Lead ECG Interpretation}},
  year = {2026},
  eprint = {2611.02041},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02041},
}
Figure