acceptodds
Under review as a conference paper at ICLR 2027

LEAP-LLM: Preserving Lead-Time Structure for Question-Conditioned ECG-Language Reasoning

Abstract

The electrocardiogram (ECG) records cardiac electrical activity across multiple leads and is used for detecting cardiac abnormalities. Recent ECG-language models enable large language models (LLMs) to interpret cardiac signals and answer clinically relevant questions. However, existing ECG-language models typically encode the ECG independently of the clinical question, forcing the ECG representation to retain broad and potentially irrelevant signal information. Moreover, these models often aggregate information across leads which can obscure lead-specific details needed for a particular question. To tackle these limitations, we introduce LEAP-LLM, an ECG-LLM architecture for Lead-aware ECG Alignment with Prompt conditioning. LEAP-LLM uses the clinical question to guide ECG feature extraction while preserving explicit lead and temporal structure. Task-trained instruction representations distill information from the question and guide two complementary branches. The Lead branch captures question-relevant temporal patterns within each ECG lead, while the Temporal branch summarizes relevant information across leads at each temporal position. Together, they provide the LLM with compact, question-relevant ECG information while preserving lead and temporal structure. LEAP-LLM reduces modality-token count by 89.0% and prefill FLOPs by on average relative to comparable ECG-LLMs. Across four ECG benchmarks, LEAP-LLM achieves state-of-the-art performance, including a 3.4-percentage-point improvement in G12EC over the strongest prior ECG-specific method. We further show that the model adaptively emphasizes the clinically relevant anatomical lead groups for different cardiac abnormalities, with lead importance patterns consistent with established diagnostic knowledge, providing interpretable evidence that the predictions are grounded in clinically meaningful ECG structure rather than black-box representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Learning, Time-Series–Language Modeling, ECG-Language Models, Cross-Modal Alignment, Medical AI

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uJ3yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LEAP-LLM is an ECG-LLM interface for question-conditioned interpretation. The authors argue that existing models encode the ECG independently of the clinical question, forcing representations to retain broadly irrelevant information, and that cross-lead aggregation obscures lead-specific detail a given question needs. Instead of encoding first and fusing later, the clinical question guides ECG feature extraction while explicit lead and temporal structure is preserved: task-trained instruction representations distill the question into queries that drive two complementary branches — a Lead branch capturing question-relevant temporal patterns within each lead, and a Temporal branch summarizing information across leads at each temporal position. The resulting compact, question-conditioned tokens are fed to the LLM. Validation covers four ECG benchmarks, token-count and prefill-FLOP reduction, and an analysis of lead-importance patterns.

Critical strengths
  • The efficiency result is substantial and practically meaningful. The reported reduction in modality tokens (89.0%) and prefill FLOPs (7.2×) addresses a real deployment bottleneck for ECG-LLMs, where long signal token sequences dominate inference cost. This is the kind of gain that matters independently of accuracy, and it is the main reason I view the paper as above clear rejection.
  • The two-branch decomposition maps cleanly onto clinical reading practice. Separating within-lead temporal evidence from cross-lead evidence at each time step is a natural factorization, and the reported anatomical selectivity of lead attention (inferior MI attending to II, III, aVF) is a reassuring sanity check that the interface learned something structurally meaningful rather than a shortcut.
Critical weaknesses
  • The efficiency claim is established only in the single-question regime, which is not the realistic use case. Because feature extraction is conditioned on the question, every new question about the same recording requires re-extraction; the 89% token reduction and 7.2× prefill gain therefore describe per-question cost, not per-recording cost. In multi-turn consultation or multi-question review of one trace — the setting that actually motivates an ECG-LLM — the advantage shrinks and may reverse relative to a question-agnostic encoder that extracts once and reuses. The paper does not quantify this anywhere, so a headline contribution rests on an unstated assumption. This is the weakness I would most want addressed, as it determines whether the claimed gain survives in practice.
  • The paper is not positioned against the prior work it most directly continues, making the novelty hard to judge. Its motivating claim — that cross-lead aggregation loses the localized, lead-specific detail a question needs — is precisely what fine-grained waveform-level ECG-text alignment was developed to preserve, and the ECG-language alignment stage sits on top of a line of text-supervised ECG pretraining using LLM-enhanced clinical descriptions (e.g., ESI, Yu et al.,; FG-CLEP, Li et al). Generalist ECG-MLLMs supporting flexible input and waveform localization are also closely related to this design. GEM and PULSE are discussed, but neither strand is, so it remains unclear whether question-conditioned selection is necessary on top of what these methods already retain.
Questions for authors
  1. What is the total cost when K questions are asked about the same recording? Could you report prefill FLOPs and latency as a function of K, and identify the K at which the advantage over a question-agnostic encoder disappears?
  2. How does LEAP-LLM compare to generalist ECG-MLLMs that already support flexible input and waveform localization with multi-task understanding?
  3. Is question-conditioned extraction necessary, or does fine-grained waveform-level alignment already retain the lead-local detail you argue is lost? Are there questions your method answers that alignment-based representations cannot?
  4. The lead-importance analysis is correlational — attention indicates where information is concentrated, not causal importance. Could you provide an intervention study (e.g., masking or replacing the identified leads) to establish a causal link to predictions?
6 comments

We thank the reviewer for the detailed feedback. We quantify the efficiency trade-off across multiple questions per recording and clarify how question-conditioned extraction complements fine-grained ECG–text alignment, supported by controlled ablation and question-swap analysis. We also clarify how lead-token interventions demonstrate selective dependence on anatomically relevant representations beyond attention visualization.

Question 1 (Weakness 1) : Cost of multiple questions about one recording

We report total prefill FLOPs and latency across multiple questions per recording. LEAP-LLM has the lowest prefill FLOPs through four questions and the lowest latency through eight questions among the evaluated settings, GEM becomes more efficient beyond these respective points as it reuses question-independent ECG tokens.

Evaluation Protocol: For each model, we cache every part of the prompt that is identical across questions about the same recording and recompute only the rest for each new question. For GEM, ECG-R1 and PULSE, the cached part includes the ECG and image tokens, which do not depend on the question. LEAP-LLM's ECG tokens are extracted with the question, so only its system prompt is cached. We vary K∈{1,2,4,8,16,32}K \in \{1, 2, 4, 8, 16, 32\} and report the total prefill cost of one recording with all KK questions. Latency is the median of 10 runs at batch size 1 on one NVIDIA L4. ECG-to-image rendering for GEM and PULSE is not counted, which favours those baselines.

Table R6. Prefill cost of one recording with KK questions

ModelK=1K=2K=4K=8K=16K=32
Prefill TFLOPs ↓
LEAP-LLM (ours)3.286.0811.6722.8745.2690.04
GEM10.6911.9014.3019.1228.7548.01
ECG-R113.4814.6416.9521.5830.8349.34
PULSE43.5244.8347.4452.6763.1484.06
Prefill latency (ms) ↓
LEAP-LLM (ours)1702825009301,7593,461
GEM3814666309451,5832,873
ECG-R15386328141,2101,9663,511
PULSE1,1341,2241,4141,8322,6854,339

Table R6 shows that LEAP-LLM has the lowest prefill FLOPs among the compared models at K=1,2,4K=1,2,4 and the lowest latency at K=1,2,4,8K=1,2,4,8. GEM becomes more efficient at larger question counts as it reuses its question-independent ECG representation. LEAP-LLM remains faster than ECG-R1 and PULSE at every evaluated question count, including K=32K=32.

(continued in the next comment)

Question 2/3 (Weakness 2): Comparison with other ECG methods and benefit of question conditioning

Comparision with General ECG-MLLM

We have compared LEAP-LLM with GEM, PULSE and ECG-R1 on downstream performance and inference efficiency (Tables 1–2 Main Manuscript). We also identified anyECG-chat [1] as closely related work that extends ECG-MLLMs to flexible inputs and waveform localization. Compared with these broader capabilities, LEAP-LLM focuses on compact, question-dependent representations for 12-lead ECG interpretation. Both LEAP-LLM and anyECG-chat incorporate lead and temporal information, but anyECG-chat’s MLP connector produces question-independent ECG features, whereas LEAP-LLM conditions extraction on the question. Flexible input handling is also compatible with our approach, variable-duration and reduced-lead recordings could be accommodated through segmentation, padding and appropriate masking. Our controlled ablation supports the benefit of question conditioning under matched settings. Our extraction module could also be incorporated into such generalist models, provided their encoders expose lead- and time-resolved features, complementing their broader capabilities with adaptive evidence selection.

Comparison with ECG-Text Alignment methods and Importance of Question Conditioning We thank the reviewer for highlighting ESI [1] and FG-CLEP [2]. We have expanded Related Work to clarify their relationship to LEAP-LLM. Our experiments support the benefit of question-conditioned extraction under a fixed token budget.

ESI learns ECG representations through contrastive and captioning objectives using ECGs paired with LLM-generated descriptions [2]. FG-CLEP aligns ECG segments with individual report tags through tag-specific representations and uses LLMs to supplement waveform information missing from reports [3]. We agree that such representations can retain local waveform detail. However, LEAP-LLM addresses the subsequent step, selecting and aggregating that information before it reaches the LLM. In particular, FG-CLEP uses report tags to guide alignment, whereas LEAP-LLM uses the question at inference time to determine which ECG evidence to emphasize.

This distinction matters when constructing a compact ECG representation. Passing all (N x L) lead-time features directly to the LLM would increase its input length and computational cost, while including information irrelevant to the current question. Compressing these features reduces this cost but requires deciding which information to retain. LEAP-LLM uses the question to guide this compression, producing (N+L) tokens that emphasize relevant evidence while retaining separate lead-specific and time-indexed representations.

Our controlled ablation directly tests whether this conditioning helps. With the encoder, LLM backbone, lead-time structure and training setup held fixed, adding question conditioning improves performance by 3.1 percentage points on CSN and 3.8 points on G12EC (Table R3, rows b–c, in our response to Reviewer zX9Y). The question-swap analysis further shows that attention follows the queried anatomical territory, while lead-token interventions demonstrate selective dependence on anatomically relevant lead representations.

A sufficiently informative question-agnostic representation could support the same answers. Our contribution is to make evidence selection adaptive when the interface must compress the ECG into a limited number of tokens. This is particularly relevant to questions requiring localized evidence, as demonstrated by our territory-specific MI analysis. Encoders such as ESI or FG-CLEP could complement this approach: their aligned representations would still require an interface to a generative LLM, where our method could be applied.

(continued in the next comment)

Question 4: Intervention evidence beyond attention

We agree that attention alone does not establish causal importance. Our lead-token intervention study show that the model’s answers depend selectively on anatomically relevant lead representations by testing whether changing the selected lead evidence affects the model's answer. We show that replacing leads associated with the queried MI territory reduces gold-answer likelihood substantially more than replacing unrelated or randomly selected leads, providing evidence beyond attention visualization.

Evaluation Protocol: We replace selected patient-specific lead tokens with the mean representation of the same leads in a reference cohort, immediately before they enter the LLM, and measure the reduction in gold-answer likelihood.

Table R7. Reduction in gold-answer likelihood (%) after replacing lead tokens

Replaced lead tokensInferior MIAnterior MI
Inferior (II, III, aVF)50.55.5
Anterior (V1–V4)6.855.0
Three random leads17.217.6

Replacing the matching territory reduces likelihood 9.2× and 8.1× more than replacing the other territory, and about 3× more than replacing three random leads. These interventions show selective functional dependence on anatomically relevant lead representations at the LLM interface, which the attention analysis alone could not establish. This experiment is presented in our manuscript (Section 5.2, with details in Appendix H). We have revised the manuscript to clarify the distinction between attention visualization and intervention-based evidence.

References

[1] Li, Haitao, et al. "anyecg-chat: A generalist ecg-mllm for flexible ecg input and multi-task understanding, 2025a." URL https://arxiv. org/abs/2506.00942.

[2] Yu, Guo and Sano. ECG Semantic Integrator (ESI): A Foundation ECG Model Pretrained with LLM-Enhanced Cardiological Text. TMLR, 2024. https://arxiv.org/abs/2405.19366

[3] Li et al. FG-CLEP: Fine-grained Contrastive Learning for ECG-Report Alignment with Waveform Enhancement. arXiv:2505.11939, 2025. https://arxiv.org/abs/2505.11939

Reviewer_ntn7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LEAP-LLM creates an ECG-language interface that keeps explicit lead and time axes. The approach uses a PatchTST-style encoder that processes the 12 leads into 99 patches, with embeddings for lead, time and lead-group. The implementation uses 128 instruction tokens in instructions. Training is two-stage on ECGInstruct and ECG-Grounding. Authors report SOTA results on four ECG-Bench datasets (PTB-XL, CODE-15%, G12EC, CSN), 89.0% fewer modality tokens and 7.2x fewer prefill FLOPs. Authors also claim interpretability evidence.

Critical strengths

A clear, well-motivated architectural idea with a careful efficiency and intervention analysis. The efficiency appendix is thorough.

Critical weaknesses

The claims are not isolated by proper ablation studies: hypothesis is that (i) conditioning the ECG tokens on the question selects relevant evidence, (ii) explicit lead/time structure matters, and (iii) the lead/time/group embeddings are relevant. The only ablation is Lead-only vs Temporal-only vs Full.

Novelty over ITFormer or OpenTSLM approaches is not evaluated / discussed.

The headline numbers are selectively framed: HeartLLM (most efficient baseline and has fewer prefill FLOPs than LEAP) is excluded. Total FLOPs are higher for LEAP. "Comparable ECG-LLMs" is never defined in the abstract. End-to-end latency for 256 tokens is only marginally lower than GEM and PULSE and peak memory is lower than GEM, making the practical benefit the prefill and TTFT. On accuracy, the abstract reports the better of two LEAP variants per metric. "State-of-the-art on four benchmarks" and "outperforms across all benchmarks" are not substantially backed. The interpretability claim in the abstract ("evidence that predictions are grounded in clinically meaningful ECG structure") also exaggerates the analysis: authors only look at attention weights for lead importance.

Questions for authors

Please compare results to OpenTSLM and ITFormer.

Please define "comparable ECG-LLMs" in the abstract.

Minor: "Equation equation 4" (Sec. 3.4) and "presented in 4" (App. B.3); "Aquisition" (App. A); "(Hancock et al., 2009)." after a full stop in B.1

8 comments

We thank the reviewer for the thoughtful comments. We add controlled component ablations and comparisons with ITFormer and OpenTSLM, and clarify the evidence supporting interpretability through question-swap analysis and lead-token interventions. We also qualify the interpretability claims, specify the abstract's efficiency comparison set, and correct the reported writing issues.

Weakness 1: Controlled ablation of individual design components

We agree that the Lead-only, Temporal-only, and Full comparisons do not isolate the contributions of question conditioning, lead-time structure, and positional embeddings. Our new controlled ablation shows that lead-time structure, question conditioning, and positional embeddings each provide performance gains when added sequentially and in isolation, while holding the encoder, LLM backbone, and training setup fixed.

Table R4. Controlled ablation with the same encoder, LLM backbone and training setup. Accuracy (%).

VariantBackboneSQPCSN Acc ↑G12 Acc ↑
(a)Vicuna-7B78.271.2
(b)Vicuna-7B✓82.374.2
(c)Vicuna-7B✓✓84.579.2
(d)Vicuna-7B✓✓87.681.2
(e)Vicuna-7B✓✓85.478.0
(f) LEAP-LLMVicuna-7B✓✓✓90.685.2

*S: lead-time structure. Q: question conditioning. P: lead and time positional embeddings.

Starting from a question-agnostic baseline, introducing explicit lead-time structure improves accuracy by 4.1 percentage points on CSN and 3.0 points on G12EC. Adding question conditioning provides further gains of 3.1 and 3.8 points, respectively, with lead-time structure. Adding lead and time positional embeddings then yields gains of 5.2 and 7.2 points. Comparing variants (c) and (f), adding the dual-branch structure while retaining question conditioning and positional embeddings improves accuracy by 6.1 percentage points on CSN and 6.0 points on G12EC. Removing question conditioning or positional embeddings from the full model also reduces accuracy on both datasets. The full model achieves the highest accuracy among all evaluated variants, supporting the complementary benefits of all three components.

These comparisons provide evidence for the benefit of each component. We have added this ablation study to Section 4.3 of the revised manuscript.

Question 1 (Weakness 2) : Comparison with ITFormer and OpenTSLM

Thank you for the suggestion. We have added comparisons with ITFormer and OpenTSLM at the 3B parameter scale. LEAP-LLM achieves stronger performance on both PTB-XL and CODE-15% than ITFormer trained on the same instruction-tuning data and OpenTSLM evaluated using its released ECG checkpoint.

Evaluation Protocol Given the compute constraints and limited rebuttal period, we conducted these comparisons at the 3B parameter scale, including a 3B variant of LEAP-LLM (Llama-3.2-3B-Instruct backbone). We trained ITFormer on the same instruction-tuning data as LEAP-LLM and evaluated OpenTSLM using its released ECG checkpoint. All models were evaluated on the same PTB-XL test split (2,082 examples) and CODE-15% test split (1,400 questions).

ITFormer. ITFormer was originally developed and trained on EngineMT-QA, an aero-engine time-series question-answering dataset [1]. We therefore retrain it on ECG data using the authors' implementation with Qwen2.5-3B-Instruct. Specifically, we use the same 356,423 instruction-tuning examples as LEAP-LLM-3B, drawn from the PTB-XL, ECG-QA, and CODE-15% training splits. Each ECG is resampled to 600 samples per lead and standardized per lead, and the examples are converted to ITFormer's question-answer format. Following its training procedure, we first pretrain the time-series encoder on the training ECGs and then train the interface with the encoder and LLM frozen.

OpenTSLM. We evaluated the released llama-3.2-3b-ecg-sp checkpoint [2], which was trained on ECG-QA-CoT. We followed its ECG input format, providing the 12 leads as separate time series sampled at 100 Hz, and used its prompt template.

Table R5. Comparison with ITFormer and OpenTSLM at the 3B scale (%).

ModelBackbonePTB-XL F1 ↑PTB-XL AUROC ↑PTB-XL HL ↓CODE-15% F1 ↑CODE-15% AUROC ↑CODE-15% HL ↓
ITFormer-3BQwen2.5-3B-Instruct12.350.028.126.561.721.5
OpenTSLM-3BLlama-3.2-3B21.650.930.714.552.916.3
LEAP-LLM-3BLlama-3.2-3B-Instruct71.280.412.783.588.75.8

LEAP-LLM-3B achieves the highest F1 and AUROC and the lowest Hamming loss on both datasets. With matched instruction-tuning data, it exceeds ITFormer by 30.4 AUROC points on PTB-XL and 27.0 points on CODE-15%. The released OpenTSLM checkpoint also shows limited AUROCs of 50.9 and 52.9, respectively.

(continued in the next comment)

Context for baseline selection. Our initial literature review found that independent studies reported performance close to or below random baselines for these models on certain ECG tasks, which informed our decision not to include them in the original comparisons. Specifically, ARTIST [3] fine-tunes 4B versions of ITFormer and OpenTSLM on its own instruction-tuning data and reports ECG-QA F1 scores of 49.91 and 41.00, respectively, compared with a random-guess F1 of 50.00. ECG-Reasoning-Benchmark [4] reports initial-diagnosis accuracy for OpenTSLM-3B of 54.8% on PTB-XL and 51.6% on MIMIC-IV-ECG, whereas PULSE, GEM, and ECG-R1 achieve 80.9-85.4% and 72.1–80.2%, respectively.

These independent findings align with our results and suggest that general-purpose time-series LLMs are not well suited to ECG interpretation than ECG-specific models.

We have clarified these limitations in the revised manuscript in the Related works sections and included these experiments in Appendix D.

a) Distinction with ITFormer

Our core methodological contribution is question-conditioned, structure-preserving dual-axis alignment, which is motivated by the native structure of the ECG signal. Unlike sequential alignment architectures such as ITFormer, which collapse channel and temporal dimensions into fused instruction tokens, LEAP-LLM preserves explicit lead-indexed and time-indexed representations at the LLM interface. This distinction is particularly important for ECG interpretation, where anatomical lead identity and temporal waveform characteristics provide complementary diagnostic information. Maintaining this explicit structural correspondence is precisely what enables the LLM to selectively leverage anatomical lead identity during interpretations (as confirmed by our intervention results (section 5.2 Manuscript)).

B Distinction from OpenTSLM

OpenTSLM, encodes each ECG lead as a sequence of temporal patches and projects these patches into the LLM’s embedding space. The LLM receives the resulting tokens alongside the clinical question, so the same ECG yields the same input patch representations across questions. LEAP-LLM uses the question during alignment itself. Its two branches produce complementary representations, so the question can influence which ECG features are emphasized before they reach the LLM. This distinction matters for ECG interpretation, where the significance of a waveform change depends on both its timing and the leads in which it appears.

Weakness 3: Qualifying the efficiency, accuracy and interpretability claims

We agree that the original wording overstated the scope of our findings. We have clarified the efficiency comparison set and emphasized prefill and TTFT gains, while qualifying the interpretability claim and adding lead-token interventions to complement the attention analysis, in the revised manuscript.

Efficiency. HeartLLM is excluded from this headline comparison because its substantially lower predictive performance places it at a different accuracy-efficiency trade-off from these ECG-specific models. It remains included in the full efficiency table, where we acknowledge its lower FLOPs and memory usage. We also clarify that the main practical gains are in prefill and TTFT, with more modest improvements in end-to-end generation latency because those are driven by equivalent decoding costs of the Large Language Model.

Interpretability. We agree that attention alone does not establish that predictions depend on the attended leads. We have the lead-token intervention analysis (Section 5.2 of the Manuscript) that shows that replacing leads associated with the queried MI territory reduces gold-answer likelihood substantially more than replacing unrelated or randomly selected leads. Together with the question-swap analysis (Manuscript Appendix G; Table R1 in Response to Reviwer zX9Y ), this supports two conclusions: the question guides where the model attends, and its answer depends on the corresponding lead representations. We have revised the abstract accordingly and added the intervention methods and results to Section 5.2 and Appendix H.

Writing corrections

Comparable baselines: We agree that the abstract should name the comparison set. "Comparable ECG-LLMs" refers to the open ECG-LLMs that pass ECG-derived tokens to an LLM backbone for question answering and that we profile in the efficiency analysis: GEM [5], ECG-R1 [6] and PULSE [7]. We have made the necessary changes in the abstract of the revised manuscript to make this distinction clear.

Minor Typographical mistakes: We thank you for the careful reading. We have corrected all four issues: removed the duplicated "Equation" in Sec. 3.4, replaced the incomplete reference in App. B.3 with "Table 4", corrected "Aquisition" to "Acquisition" in App. A, and moved the citation in App. B.1 before the period. We also proofread the manuscript and corrected additional minor typographical errors.

We appreciate the time and effort you have put into reviewing our work. With the discussion deadline approaching, we wanted to follow up on whether the clarifications and additional experiments we provided have resolved your concerns.

We would be happy to address any remaining questions before the deadline. Thank you again for your time and consideration.

Reviewer_zX9YOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LEAP-LLM, a question-conditioned ECG-language framework that preserves lead-specific and temporal structure through task-trained instruction representations and dual-branch cross-attention. The method is evaluated through benchmark comparisons, inference efficiency profiling, architectural ablations, and representation analyses, including attention visualization and targeted token interventions.

Critical strengths

The paper uses a unified profiling protocol with identical input, hardware, batch size, and generation length, separates prefill latency from time-to-first-token, and explicitly notes differing measurement scopes across baselines. This makes the efficiency claims credible and reproducible.

Critical weaknesses
  1. The paper does not specify the exact pretraining sample sizes or provide sufficient independent validation of encoder quality. Moreover, evaluation is restricted to standardized ECG recordings, leaving the robustness of the learned representations under varying acquisition conditions insufficiently established.
  2. The paper lacks pretraining sample sizes and independent encoder validation, and evaluates only standardized ECG recordings, leaving robustness to acquisition variation unestablished.
Questions for authors
  1. Can the authors demonstrate that different clinical questions applied to the same ECG induce clinically appropriate changes in lead–time attention?
  2. What are the exact pretraining sample sizes and dataset versions, and how was the pretrained ECG encoder independently validated?
  3. Can the authors provide controlled comparisons with matched encoders, backbones, and token budgets to isolate the contributions of question conditioning and lead–time structure preservation?
6 comments

We thank the reviewer for the constructive feedback. We address your concerns through controlled component ablations, a question-swap analysis demonstrating question-dependent attention, and independent frozen-encoder validation. We also specify the pretraining datasets and sample sizes and clarify cross-dataset generalization and the limitations regarding acquisition robustness.

Question 1: Question-dependent Lead Attention

We agree that question conditioning should change which leads the model reads, and that the change should follow cardiac anatomy. Our experiment below show that LEAP-LLM redirects its attention toward anatomically relevant leads when the question changes, while the ECG remains fixed.

Evaluation Protocol: To test whether changing the question redirects attention toward the corresponding anatomical lead group, we performed a controlled question-swap analysis while holding the ECG fixed. We constructed the evaluation cohort from the 2,082 ECGs in the held-out PTB-XL fold. We focused this analysis on the 384 single labelled inferior and anterior MI ECGs. Each ECG was paired with two questions: “Is there evidence of myocardial infarction in {inferior/anterior} leads on this ECG?” The questions had identical format and token length, with only the queried territory changing. We take LEAP-LLM's lead-attention scores, average them over time and sum them within two lead groups 1) inferior (II, III, aVF) and 2) anterior (V1–V4). Table R1 reports the share of lead attention each group receives under each question.

Table R1. Lead-group attention share (%) on 384 held-out single-territory MI ECGs

Question askedAttention to inferior leads (%)Attention to anterior leads (%)
Is inferior MI present?79.87.7
Is anterior MI present?9.178.4

Across these 384 ECGs, changing the question moves 70.7 percentage points of attention onto the queried territory in both directions, an 8.8× and 10.2× increase in its share. Figure R1 shows the behaviour on one held-out ECG with inferior MI (PTB-XL 6925). Under the inferior-MI question, attention concentrates on leads II and III and the model correctly answers "yes". Under the anterior-MI question, attention moves mainly to V2 and V3 and the model correctly answers "no".

Figure 6 (revised manuscript, Appendix G) shows an example of how the model attends to different lead groups when asked different questions about the same ECG.

Together, these results show that LEAP-LLM’s lead attention depends on the question and shifts toward the corresponding anatomical territory, rather than remaining fixed for a given ECG.

(continued in the next comment)

Question 2 (Weaknesses 1-2) : Pretraining data and independent encoder validation

1) Generalization across diverse ECG datasets. We acknowledge the reviewer's concern regarding robustness across different acquisition conditions. LEAP-LLM is instruction-tuned on ECG-Instruct and evaluated across four ECG benchmarks: PTB-XL, CODE-15%, G12EC, and CSN. These datasets contain ECG recordings collected from different institutions, patient populations, and acquisition settings, providing a diverse evaluation of the model's generalization capabilities. Importantly, G12EC and CSN are out-of-domain evaluation datasets that are not included in LEAP-LLM's training set, allowing us to assess generalization to previously unseen data distributions. Our strong results provide evidence that LEAP-LLM generalizes beyond its instruction-tuning distribution. Robustness to specific acquisition perturbations remains an important direction for future evaluation and we have addressed that in the limitation section of the manuscript.

2) Encoder pretraining and validation. We thank the reviewer for requesting clarification. Our encoder was pretrained on 805,926 recordings: 17,104 from the PTB-XL v1.0.3 training split [1] and 788,822 from the MIMIC-IV-ECG v1.0 training split [2]. We had validated our pretrained encoder via linear probing on held-out dataset which shows that the encoder captures diagnostic information.

Evaluation Protocol: To independently validate the pretrained encoder, we performed frozen-encoder linear probing on CODE-15% [3], which was excluded from encoder pretraining. We sampled 600 single-label examinations from each of seven classes (4,200 total) and used an 80/20 examination-level split with seed 0, yielding 3,360 training and 840 test examinations. We averaged the encoder features across leads and patches to obtain a 512-dimensional representation per examination. Features were standardized using training-set statistics, and a one-vs-rest logistic regression classifier was fitted on the training split. The encoder remained frozen throughout. We train only the linear classifier.

Table R2. Independent frozen-encoder linear-probe validation on CODE-15% (840 test examinations)

ClassAUROC
Normal ECG0.989
First-degree AV block0.983
RBBB0.988
LBBB0.992
Sinus bradycardia0.995
Sinus tachycardia0.998
Atrial fibrillation0.991
Macro-AUROC0.991

Results: The frozen encoder achieves a macro-AUROC of 0.991, with per-class AUROCs ranging from 0.983 to 0.998. These results demonstrate that its pretrained representations contain diagnostic information accessible to a linear classifier on a dataset excluded from encoder pretraining. We have added these encoder pretraining details and validation results in Appendix B of the manuscript.

Question 3: Controlled ablation of lead-time structure and question conditioning

We thank the reviewer for highlighting the need to distinguish the contributions of our extraction module from differences in the encoder and LLM backbone. We conducted a controlled ablation using the same pretrained encoder, backbone, data and training schedule across all variants showing that preserving lead-time structure and question conditioning improves performance

Starting from a question-agnostic extractor without explicit lead-time organization and positional embeddings, we progressively introduce (i) lead-time structure, (ii) question-conditioned extraction, and (iii) lead and time positional embeddings. Table R3 reports accuracy on CSN and G12.

Table R3. Controlled ablation with the same encoder, LLM backbone and training setup. Accuracy (%).

VariantSQPCSN Acc ↑G12 Acc ↑
(a)78.271.2
(b)✓82.3 (+4.1)74.2 (+3.0)
(c)✓✓85.4 (+3.1)78.0 (+3.8)
(d) LEAP-LLM✓✓✓90.6 (+5.2)85.2 (+7.2)

S: lead-time structure. Q: question conditioning. P: lead and time positional embeddings. Parentheses give the gain over the previous row. Introducing lead-time structure improves accuracy by 4.1 percentage points on CSN and 3.0 points on G12. Adding question conditioning yields a further 3.1 and 3.8 points, respectively. This comparison directly supports the benefit of question-conditioned extraction . Overall, the full model outperforms the baseline by 12.4 points on CSN and 14.0 points on G12. We have added this ablation study to Section 4.3 of the revised manuscript.

These results show the effectiveness of the proposed extraction design.

References [1] Wagner et al. PTB-XL, a large publicly available electrocardiography dataset. Scientific Data, 2020. v1.0.3

[2] Gow et al. MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset. PhysioNet, 2023. v1.0

[3] Ribeiro et al. CODE-15%: a large scale annotated dataset of 12-lead ECGs. Zenodo, 2021

Reviewer_zX9YOpenReview ↗

Post-Rebuttal Comment

I appreciate the authors' additional experiments and clarifications. The pretraining details and independent encoder validation satisfactorily address my second question. The question-swap analysis and new ablations also provide useful evidence.

However, my main concern regarding the methodological contribution remains unresolved.

  1. Question-dependent attention vs. ECG-dependent evidence selection. The question-swap experiment demonstrates that attention shifts toward the anatomical leads explicitly specified in the question. While this confirms the presence of question conditioning, it does not establish that the model selects evidence based on the actual ECG findings, rather than merely learning a rigid question-to-lead association.

  2. Component effectiveness vs. methodological novelty. The new S/Q/P ablation shows that adding components improves accuracy, but it fails to isolate the benefit of the proposed dual-branch design over simpler alternatives. Notably, adding standard positional embeddings alone yields a further +5.2 points on CSN and +7.2 points on G12EC. This strongly suggests that a substantial portion of the improvement stems from conventional structural encoding, rather than demonstrating a unique advantage of the dual-branch architecture itself. Moreover, ITFormer already employs instruction tokens, question-guided channel/temporal attention, and structural positional encodings. A matched comparison against standard question-conditioned attention—using the same encoder, LLM, and output-token budget—is still missing.

Thus, while the rebuttal successfully strengthens the empirical evidence, it does not establish a sufficiently distinct methodological contribution beyond existing instruction-guided time-series alignment methods.

I therefore maintain my current rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response

We sincerely thank all reviewers for their constructive feedback and the Area Chair for overseeing the discussion. Following the rebuttal, we have incorporated additional experiments, analyses, and clarifications into the revised manuscript. These revisions strengthen the evaluation of our design, clarify its relationship to existing approaches, and examine the benefits and computational tradeoffs of question-conditioned ECG interpretation. Below, we summarize the major revisions. All new and modified text is highlighted in blue in the revised manuscript.

Main Paper

Abstract and Related Work

We refined the scope of our accuracy, efficiency, and interpretability claims. We expanded the discussion of ESI and FG-CLEP and clarified how question-conditioned evidence selection complements ECG–text representation learning.

Controlled Ablations and Simpler Alternatives

We added component ablations separating the contributions of lead–time structure, question conditioning, and positional embeddings. In our follow-up response, we further provided:

  • Extended component combinations: showing the isolated benefit of the dual-branch structure

  • A matched cross-attention baseline: showing the benefit of the dual-branch design over standard question-conditioned attention under matched experimental settings.

Structural and Functional Analysis

We brought together three complementary analyses:

  • Varying the ECG under a fixed question,
  • Varying the question for a fixed ECG, and
  • Intervening on patient-specific lead tokens. Together, they support evidence selection that responds to both the ECG and the question, and selective reliance on clinically relevant representations for downstream prediction.

Appendix Additions and Extensions

Encoder Pretraining and Validation (Appendix B)

  • We specified pretraining datasets, versions, and sample counts, and added independent frozen-encoder validation on CODE-15%, achieving 0.991 macro-AUROC.

Additional Baseline Comparisons (Appendix D)

  • We added comparisons with ITFormer and OpenTSLM at the 3B scale, including retraining ITFormer on the same instruction-tuning data. LEAP-LLM achieves stronger results on both PTB-XL and CODE-15%. We also clarified the architectural distinction from ITFormer in our responses.

Multi-Question Efficiency (Appendix E.5)

  • We extended the efficiency evaluation to multiple questions per recording, allowing each model to reuse shared prompt content through KV caching. This analysis explicitly accounts for LEAP-LLM’s need to recompute question-conditioned interface outputs. We also clarify the tradeoff at larger question counts

Question-Dependent Lead Selection (Appendix G)

  • We added the question-swap protocol, cohort details, quantitative results, and visualization, complementing the functional intervention evidence in Appendix H.

Limitations and Minor edits

  • We also addressed acquisition-robustness limitations, clarified comparisons with generalist ECG-MLLMs, and corrected the reported writing issues.

Collectively, these additional results and analysis reinforce our original findings and further establish the benefits of the dual branch design catered to ECG interpretation. We appreciate the reviewers’ suggestions, which have helped us strengthen the manuscript, and welcome any remaining questions or clarifications.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.