- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
LaUER is a language-anchored ECG representation learning framework intended to support both disease detection and interpretation from a single representation. The premise is that classification-oriented representations align with coarse disease concepts while generation-oriented ones emphasize fine-grained linguistic expression, even though cardiologists draw on the same underlying knowledge and differ only in expression granularity. To this end, ECG images are introduced alongside signals, since images are the format clinicians actually read. A multi-granularity pretraining strategy is built over clinical text: semantic-granularity learning anchors signal and image representations to report-level text via asymmetric cross-modal alignment; concept-granularity learning uses a standardized cardiovascular disease query bank as anchors with a query-guided multi-label readout over fused signal-image tokens; token-granularity learning trains a token-level decoder on the same fused tokens for report generation. Validation covers linear probing, zero-shot classification, and report generation.
Critical strengths
- The premise is well-articulated and directly testable. Framing detection and interpretation as two readouts over shared knowledge at different linguistic granularities is a clean hypothesis, and it is the kind of claim whose success or failure is informative regardless of the outcome. This framing is what makes the paper worth engaging with rather than dismissing, and it is the main reason I view it as above clear rejection.
- Introducing the ECG image modality on clinical grounds is a sensible, underexploited choice. Most ECG-text work uses signals only, yet images are what cardiologists actually read; the reported linear-probing and zero-shot gains suggest the fused representation is at least competitively discriminative.
Critical weaknesses
- The unified-representation claim is conceded by the paper itself and then left unexamined. The abstract states that report generation is only "close to" state of the art while detection exceeds it, which means the single representation serves the discriminative readout better than the generative one — precisely the asymmetry the thesis set out to eliminate. The paper does not analyze why the generative side lags, nor whether the concept-level (discriminative) and token-level (generative) objectives interfere, for example through gradient conflict or through the shared fused-token bottleneck being pulled toward label-relevant rather than description-relevant content. Nor do the ablations isolate each granularity's marginal contribution, so it is not established that all three stages are necessary or even mutually reinforcing. Since "one representation for both" is the entire contribution, a result that supports it only halfway, without diagnosis, does not sustain the claim.
- The framework is not positioned against the prior work it most directly extends. Text-anchored ECG pretraining using machine-generated or LLM-enhanced cardiological descriptions (e.g., ESI, Yu et al., TMLR 2024) established that clinical text can supervise signal encoders, and subsequent work aligned localized waveform segments with individual report tags (e.g., FG-CLEP, Li et al., 2025). LaUER's concept and token granularities are essentially alignment stages of this kind, so the marginal contribution of the multi-granularity curriculum over that line is unclear. This matters substantively rather than bibliographically: it has been reported that a large share of ECG reports omit detailed waveform descriptions, which bounds how much interpretive signal clinical text can supply as an anchor — a limitation directly relevant to a language-anchored design, and one that the same line of work addresses through text enhancement. MERL and MELP are cited, so the gap is conspicuous.
Questions for authors
- Why does report generation only approach state of the art while detection exceeds it? Do you observe interference between the concept-level discriminative objective and the token-level generative objective, e.g., degraded generation when the concept stage is trained longer?
- How do your concept- and token-granularity anchoring stages differ from prior fine-grained alignment between localized waveform segments and individual report tags? What does the semantic stage add beyond that?
- Given that many ECG reports omit detailed waveform descriptions, how does this bound the clinical text used as an anchor? Did you filter, augment, or otherwise enhance the report corpus?
- Can you provide a per-stage ablation showing the marginal contribution of each granularity, and confirm that the progressive ordering of the three stages is necessary rather than incidental?
- The manuscript is short for a three-stage pretraining framework. Are the fusion strategy, the query bank construction, and the image encoder adequately ablated in the appendix?
11 comments
Reply to Weakness 1 and Question 1 (Part B)
Continue to reply W1/Q1.
Regarding the performance difference between report generation and disease detection, the revised manuscript adds a paired evaluation against frontier multimodal large language models on the same cases. We randomly sample 200 records from the MEETI (Zhang et al., 2025) test set. LaUER, DeepSeek-V4.1-Flash, and GPT-6.1-Sol each generate a report and an interpretation for every record in this same set of 200 records. The reports and interpretations are then independently scored by the same six cardiologists using GEM's seven five-point clinical criteria. As shown in this response's Table R2, LaUER's mean score is at least 4.0 on every criterion. Findings novelty reaches 4.7, the highest among the three models, while analytical accuracy and reasoning quality both reach 4.5, within 0.2 points of both frontier models. These results show that LaUER performs well in report generation compared with frontier multimodal large language models and offers clinical value.
These results also demonstrate the practical value of the proposed framework. Clinical ECG data are distributed across independent hospital systems. A single institution faces challenges in aggregating large-scale data to train a dedicated report-generation model from scratch and must also account for the computational costs of training and ongoing inference. LaUER has 267.6M parameters and completes three-stage pre-training on the MEETI corpus of 784,680 records using two RTX PRO 6000 GPUs, enabling the same fused representation to support both disease detection and report interpretation. Together with the stage-wise ablations in this response's Table R1, these results show that the model can acquire competitive report-generation capability while largely preserving classification performance. This provides a viable path for local deployment at individual institutions and addresses our central question of whether one representation can support two readouts.
Regarding the lack of isolated marginal contributions for each granularity, this response's Table R1 and the revised manuscript's Table 6 already report stage-wise marginal changes. LaUER is inspired by how humans understand natural language. At the language level (SGL), we first establish an overall association between things and language. At the concept level (CGL), we associate individual concepts with language and learn to distinguish them. At the token level (TGL), we learn to describe things based on our understanding of concepts. CGL therefore relies on trained semantic-granularity representations, and TGL requires trained concept-granularity representations. We compare the contributions of successive stages along LaUER's cumulative SGL, SGL+CGL, and SGL+CGL+TGL training sequence. We explain this further in Q4.
Regarding whether longer concept-stage training degrades generation, we do not train CGL to convergence in isolation and then freeze it. The CGL loss remains active during TGL, and the two objectives are jointly optimized throughout. Overfitting does require attention during training. The training-dynamics figure, provided as the revised manuscript's Figure 6 in the revised manuscript's Appendix D.2, is compiled from training logs and shows how the two modules remain balanced. It reports changes in concept zero-shot performance and generation quality across CGL and CGL+TGL training. The zero-shot values for CGL-1 through CGL-5 on the left are evaluated separately on downstream datasets using the CGL checkpoint retained after each epoch. Those for CGL/TGL-1 through CGL/TGL-5 on the right use the CGL+TGL checkpoint after each epoch under the same downstream evaluation protocol. AR_PPL on the right is measured on the validation set after each training epoch. The right axis reports AR_PPL, for which lower is better, so an upward curve indicates improved generation. During CGL, concept zero-shot AUROC rises rapidly from 24.0% to 82.26% and largely saturates. After TGL begins, it remains at 82.2%–82.3%, while validation AR_PPL falls from 2.60 to 2.35. Together, the curves show that CGL establishes the discriminative readout and TGL continues to improve generation without disrupting it.
(Response Table R1 and R2 are presented in Part C.)
Reply to Weakness 1 and Question 1 (Part C)
Continue to reply W1/Q1.
Response Table R1. Stage-wise ablation results for classification readouts and generative capability. AUROC (%) is averaged equally over the six downstream tasks.
| Stage | Semantic-based ZS | Concept-based ZS | Linear 1% Signal | Linear 1% Image | Linear 1% Fusion |
|---|---|---|---|---|---|
| SGL | 66.17 | N/A | 74.68 | 77.38 | N/A |
| SGL+CGL | 66.13 | 82.94 | 74.38 | 77.26 | 80.49 |
| SGL+CGL+TGL | 66.11 | 82.72 | 74.52 | 77.26 | 80.92 |
SGL/CGL/TGL results are averaged over three pre-training seeds (42, 66, and 20260915). The between-seed standard deviations of concept zero-shot AUROC are 0.90 and 1.41 for SGL+CGL and SGL+CGL+TGL, respectively. SGL has no trained fusion module or concept head, so the corresponding entries are N/A.
Response Table R2. Cardiologist ratings on a five-point scale (mean ± standard deviation) for LaUER and two frontier multimodal large language models (the revised manuscript's Table 5)
| Model | Params | Analytical Relevance | Analytical Accuracy | Analytical Completeness | Reasoning Quality | Findings Novelty | Clinical Value | Overall Satisfaction |
|---|---|---|---|---|---|---|---|---|
| DeepSeek-V4.1-Flash | Undisclosed | 4.7±0.46 | 4.6±0.57 | 4.8±0.40 | 4.7±0.49 | 4.5±0.66 | 4.6±0.60 | 4.7±0.51 |
| GPT-6.1-Sol | Undisclosed | 4.7±0.45 | 4.6±0.60 | 4.7±0.51 | 4.7±0.52 | 4.4±0.70 | 4.6±0.57 | 4.6±0.60 |
| LaUER (ours) | 267.6M | 4.3±0.52 | 4.5±0.48 | 4.2±0.60 | 4.5±0.67 | 4.7±1.03 | 4.0±0.85 | 4.3±0.64 |
Note. The evaluation uses 200 MEETI test records (selection seed 42), independently scored by six cardiologists. All three models use the same cases, raters, and scoring criteria, namely GEM's seven five-point criteria (Lan et al., 2025). The prompt templates are provided in the revised manuscript's Appendix C.5. The parameter counts of the frontier models are undisclosed; LaUER has 267.6M parameters. Bold indicates the highest value among the three models.
Please see the revised manuscript's Figure 6 in the revised manuscript's Appendix D.2 for the changes in concept zero-shot AUROC and validation AR_PPL during CGL→TGL training. The left axis shows concept zero-shot AUROC (%), the right axis shows validation AR_PPL (lower is better), and the gray dashed line marks the boundary between CGL and TGL.
Reply to Weakness 1 and Question 1 (Part A)
We thank the reviewer for accurately summarizing our central hypothesis and for the constructive feedback. The reviewer's description of two readouts of the same knowledge at different language granularities captures our intention. To facilitate verification, we have anonymously released the code, model weights, etc. The project code is available at https://github.com/anonymous-LaUER/LaUER_main. Model weights and configuration files are available at https://huggingface.co/anonymous-LaUER/LaUER. The ECG image data required for downstream evaluation are available at https://huggingface.co/datasets/anonymous-LaUER/LaUER-downstream-images, and the supervision annotations for the pretraining corpus are available at https://huggingface.co/datasets/anonymous-LaUER/LaUER-supervision-v4.
W1 & Q1. Examining the unified-representation claim, the generation gap, and objective interference
We thank the reviewer for this valuable question, which will help us clarify the main message in the revision. The reviewer accurately describes our claim. Detection and interpretation are two readouts of the same knowledge at different language granularities. Our question is whether a single representation can support both readouts. Following the progression of language understanding from coarse to fine, we use semantic-granularity learning (SGL), concept-granularity learning (CGL), and token-granularity learning (TGL) to learn a unified representation. In CGL, disease concepts serve as queries that read concept-related ECG information from the fused token sequence through cross-attention. A shared affine classifier maps the updated concept-query states to logits, each of which passes independently through a sigmoid to give the probability of the corresponding disease. In TGL, the decoder reads the same fused token sequence to generate reports and interpretations autoregressively. Classification requires concept-related information that distinguishes diseases, whereas generation requires fine-grained sequence information for clinical text. The two tasks therefore place different demands on the shared representation. We acknowledge this difference and investigate the cost at which one representation can support two readouts. This positioning appears in the abstract and introduction, but was not stated explicitly enough. W1 could consequently be read as rejecting the unified-representation claim. We will make the distinction explicit in the revision. Below, we examine one representation with two readouts through representation visualizations and the marginal changes of SGL, CGL, and TGL.
Representation visualizations of the modules. As shown in the revised manuscript's Appendix D and the main text, the concept and token readouts act on the same fused representation. The concept-tracing layer and the sixth decoder layer exhibit different attention distributions for the same cases, but both read the same signal–image fusion tokens. The two readouts address the same cardiac pathological state, expressing it through class discrimination or natural language, respectively. This explains how the same representation can support both readouts. It differs fundamentally from using two independent models and underlies our description of the representation as unified.
Marginal changes of the modules. We first consider the hypothesis that a severe conflict between the concept-level discriminative and token-level generative objectives would damage both classification and generation. The actual results in this response's Table R1 do not support this hypothesis. After TGL, concept zero-shot performance changes by only −0.22%, while fused-representation linear probing and cross-dataset zero-shot performance improve by 0.43% and 0.32%, respectively (this response's Table R1 for linear probing; the revised manuscript's Tables 33 and 34 in the revised manuscript's Appendix E.3 for cross-dataset zero-shot evaluation). We observe no evidence that the discriminative objective crowds out generative information or that the generative objective destroys discriminative information. We interpret this as the two forms of supervision sharing the same fused representation while acting along different readout directions. They are compatible under our training recipe.
(Please continue to read Part B.)