{"id":"d4326b22-72ec-448c-a1d1-7544045705be","arxiv_id":"2412.14373","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A BPE-based tokenizer lets an LLM generate clinical text directly from quantized ECG signals, matching two-stage encoder methods with roughly 3x faster training and 48% of the data.","lead":"ECG-Byte turns a 12-lead heart recording into short text-like tokens, so a language model can be trained end-to-end to answer questions about the ECG. It matches slower two-stage encoder pipelines using about half the data and 3 times less training time, while keeping tokens traceable back to the original signal.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No behavioral control demonstrates that ECG tokens influence generated text; the paper's own attention analysis (Fig. 4) shows attention concentrated on text, so the central conditional-NLG claim is unsecured.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the paper does not demonstrate that the generated text is conditioned on ECG content. My independent reading of Section 5.5, especially Figure 4 and the surrounding discussion, reinforces this: the model's own attention is concentrated on text, and the authors explicitly note that attention visualizations may not indicate contribution to the output. The absence of any behavioral control (text-only, shuffled, or random-token) means the headline result in Table 1 is compatible with the model largely ignoring ECG tokens and answering from question priors and generic clinical language. This is not an external-consensus disagreement; it is an internal evidential gap in the central claim. The efficiency numbers and the interpretability of the token-to-signal mapping are credible and would survive even if conditioning were weak, but the paper's stated contribution of 'conditional NLG' requires the ECG to matter. The proposed control directly settles whether the ECG tokens carry signal into the generated text. I agree with the reader's CONDITIONAL verdict and see no reason to change it; the concern is not new, but it is real and unresolved.","tokens_in":26505,"tokens_out":2495,"duration_ms":23762,"concrete_test":"Run the trained ECG-Byte model under three controls on the ECG-QA PTB-XL test set: (1) replace the ECG token sequence with randomly sampled tokens from the same vocabulary, (2) shuffle the ECG tokens within each instance (or permute across leads), and (3) remove ECG tokens entirely, leaving only [SIG START] and [SIG END] with no inserted content. Measure BLEU-4, Rouge-L, Meteor, and BertScore for each control. If any control matches the full-model scores within one standard deviation, the conditional-NLG claim fails; if shuffled or random controls drop substantially, conditioning is demonstrated. As a complementary check, compare generated outputs for identical questions paired with different ECGs from different patients.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is conditional NLG: generated text should depend on the ECG signal. The paper never reports a text-only, shuffled-ECG, or random-token control. Without such a control, the competitive BLEU and Rouge scores in Table 1 could arise from question priors and template answers alone. Section 5.5 (Fig. 4) reports that attention is concentrated on the textual portion of the input, and the authors acknowledge that vision-language models often show minimal attention to visual input. Attention is not explanation, but combined with the absence of a behavioral control, it makes the conditioning premise the least secured part of the argument. The efficiency and interpretability claims are independently supported, but if the ECG tokens are not actually used by the model, then the method is not a generative ECG language model, and the comparison to two-stage baselines does not measure what the abstract claims.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ECG-Byte, a byte-pair-encoding (BPE) tokenizer that converts 12-lead ECG signals into discrete tokens by quantizing normalized amplitudes into a 26-symbol alphabet and then applying BPE merges. The tokens are fed directly into a decoder-only LLM (Llama-3.2-1B, with LoRA) together with a textual question, replacing the conventional two-stage pipeline of self-supervised ECG encoder pretraining followed by LLM finetuning. The authors report competitive or better NLG metrics on ECG-QA (MIMIC-IV and PTB-XL) compared with their implemented two-stage baselines, while using about 48% of the data and roughly 3x less total training time. They also provide ablations over LLM choice, sequence length, ECG segment length, number of merges, and model size, plus qualitative token-to-ECG mappings and attention visualizations. A central claimed advantage is interpretability: ECG tokens can be mapped back to the original signal. The paper concludes that a rule-based compressor can replace a learned ECG encoder in generative ECG language modeling.","tokens_in":26695,"tokens_out":5511,"duration_ms":49403,"significance":"If the central claim holds, the paper would make a useful contribution: it challenges the prevailing two-stage SSL-encoder paradigm for ECG-to-text generation and offers a simpler, faster, and more interpretable alternative. The experiments are generally well structured: results are reported as means over 5 seeds with standard deviations; ablations cover multiple LLMs, sequence lengths, ECG lengths, and merge counts; and the code is promised to be released. The efficiency comparison in Table 2 is a genuine strength, and the interpretability analysis through token-to-signal mapping is more informative than the usual latent-feature visualizations. However, the paper's load-bearing claim is conditional NLG: that the generated text actually depends on the ECG tokens. That claim is not established by the current experiments because no text-only, shuffled-ECG, or random-token control is reported. The paper's own attention analysis (Figure 4) shows that attention is concentrated on the text portion, and the authors explicitly acknowledge that this pattern is common in vision-language models.","major_comments":[{"comment":"The central claim is that ECG-Byte enables conditional NLG, where generated text depends on the ECG tokens. However, no experiment demonstrates that the ECG tokens influence the output. The training objective in Eq. (3) conditions on Context that includes XID, but the paper never reports a control with the ECG tokens removed, replaced by random tokens, or shuffled. Given that Figure 4 shows the model attending primarily to the text portion of the input, and Section 5.5 cites prior VLM work showing minimal attention to visual input, the competitive BLEU and Rouge scores in Table 1 could be driven by question priors and template answers. Please add a text-only or shuffled-ECG control and report whether the metrics change materially. If they do not change, the conditional-NLG framing in the abstract and Section 5.1 should be weakened to avoid overclaiming.","section":"Section 4, Section 5.1, Section 5.5"},{"comment":"The sentence \"Table 1 demonstrates ECG-Byte's effectiveness, showing competitive or superior performances across all metrics and datasets\" is not supported by the table. On ECG-QA PTB-XL, the LMIM baseline achieves higher BLEU-4 (15.14 ± 0.28 vs 13.93 ± 0.21) and higher Meteor (29.64 ± 0.30 vs 29.17 ± 0.31) than ECG-Byte, even though ECG-Byte is superior on Rouge-L and BertScore. The text should be revised to \"competitive on most metrics and datasets, superior on some,\" with a clear statement of where ECG-Byte loses. This overstatement appears again in the cross-dataset transfer results in Table 7, where ECG-Byte is often worse than the two-stage baselines.","section":"Section 5.1, Table 1"}],"minor_comments":[{"comment":"The constants ϵ1 = 0.5, ϵ2 = 10⁻⁶, and the alphabet size |A| = 26 are presented without justification. Please add a sentence on how sensitive the results are to these choices or acknowledge them as hyperparameters that were not systematically tuned.","section":"Section 3.2, Eq. (1)"},{"comment":"The main experiments use num_merges = 3500, but the ablation in Table 8 shows that num_merges = 2500 gives higher BLEU-4 (15.10 vs 13.93) and Meteor (30.12 vs 29.17) on ECG-QA PTB-XL. Please explain why 3500 was chosen for the main results, or report the best-performing setting.","section":"Section 4.1 and Table 8"},{"comment":"The T = 2500 row uses only 97,244 training instances instead of the 400,000 used for T = 250 and T = 500. This confounds the effect of ECG segment length with dataset size. The text should state this limitation explicitly and avoid attributing the T = 2500 improvement solely to richer ECG context.","section":"Table 6"},{"comment":"The attention visualization is qualitative and only described for one lead. Please specify how many examples were inspected and whether the pattern was consistent, or provide a quantitative summary of attention allocated to ECG vs text tokens across the evaluation set.","section":"Section 5.5, Figure 4"},{"comment":"The statement \"we focus on one lead due to the uniformity of attention across encoded signal tokens\" is asserted without supporting evidence. Please provide a quantitative comparison of attention across leads or remove the claim.","section":"Section 5.5"},{"comment":"There are typographical artifacts in headings, e.g., \"T raining\" in Section 5.3 and \"T oken\" in Section 5.5. These should be corrected in a final proofreading pass.","section":"Headings"}],"recommendation":"major_revision","confidential_remarks":"The missing behavioral control is the key issue. It is fixable and should be required before publication: a text-only or shuffled-ECG control would directly test whether the model actually conditions on ECG tokens. If such a control shows no drop in metrics, the paper would no longer support its conditional-NLG claim, which would be a major problem for its scope. The paper otherwise has solid experimental infrastructure, including multi-seed reporting and ablations, and the efficiency claim is credible. I would not reject outright, but the revision must address the control experiment and the overstated 'superior' wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work on ECG-language modeling. ECG-Byte is a BPE tokenizer for raw ECG signals that lets you skip the SSL encoder and feed the LLM directly. The paper's real contribution is showing that this rule-based compressor can be trained end-to-end and still land in the same ballpark as two-stage SSL baselines on the ECG-QA tasks, while cutting training time by about 3x and using roughly half the data. That efficiency claim is concrete, measured, and reproducible with the released code. The token-to-signal mapping analysis is also a nice addition: you can see BPE merging P waves, QRS complexes, and T waves, which is more interpretable than any latent vector.\n\nThe main soft spot is the one the stress-test flags. There is no control condition that shows the generated text actually changes when the ECG tokens change. No text-only baseline, no shuffled-ECG baseline, no random-token baseline. The paper's own Figure 4 shows attention concentrated on the text portion, and the authors openly discuss that vision-language models often ignore the visual input. Given that, competitive BLEU and Rouge scores alone do not prove the model is reading the ECG. This is not fatal to everything in the paper—the tokenizer itself, the efficiency analysis, and the interpretability mapping all stand on their own—but it unseats the abstract's central claim about conditional NLG. The fix is straightforward: add a text-only or shuffled-signal control in a revision.\n\nTwo smaller issues. First, Section 5.1 says 'competitive or superior across all metrics,' but on PTB-XL the MIM baseline beats ECG-Byte on BLEU-4 (15.14 vs 13.93) and Meteor (29.64 vs 29.17), so that wording should be dialed back. Second, the baselines are mostly self-implemented; they are described well and use standard SSL objectives, but the ECG-Chat comparison is explicitly not apples-to-apples, so the real comparison is to their own reimplementations. That is acceptable for a systems paper, but the reader should keep it in mind.\n\nThe ablations and the honest limitations section are good; the paper does not oversell, and the code is available. This is a solid empirical systems contribution that deserves a serious referee. It is not the last word, but it gives the community a simpler alternative worth testing. I would send it out and ask for the control experiment before accepting.","headline":"A genuinely useful tokenization recipe for ECG-LLMs, with solid efficiency numbers, but the missing conditioning control leaves the central claim under-supported.","tokens_in":27251,"tokens_out":3089,"would_cite":true,"duration_ms":26153,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A rule-based byte-pair tokenizer lets an LLM answer ECG questions end-to-end, matching learned ECG encoders at a fraction of the training cost.","keywords":["ECG language modeling","byte pair encoding","tokenization","end-to-end LLM training","generative ECG analysis","interpretable ECG tokens","natural language generation","time-series tokenization"],"falsifier":"Run the exact ECG-QA PTB-XL experiment with the ECG token block replaced by (a) no ECG tokens at all and (b) the same ECG tokens in random order. If BLEU-4, Rouge-L, Meteor, and BertScore do not drop substantially relative to the reported ECG-Byte numbers, the paper's claim that ECG-Byte conveys signal content is falsified.","tokens_in":26298,"feed_emoji":"🫀","tokens_out":6692,"duration_ms":53860,"temperature":0.7,"pith_summary":"ECG-Byte is a tokenizer that turns raw 12-lead electrocardiogram signals into discrete text-like tokens, so a large language model can be trained end-to-end to answer questions about an ECG. The paper argues that this rule-based compression replaces the standard two-stage pipeline, in which an ECG encoder is first pretrained with a self-supervised objective and only then used to finetune an LLM on latent features. The central claim is that ECG-Byte matches or beats those learned-encoder baselines on natural language generation metrics while training about three times faster and using roughly 48% of the data. A further claimed benefit is interpretability, because each ECG token can be mapped back to the exact signal segment that produced it, allowing attention patterns to be visualized on the original waveform. If this holds, generative ECG-language modeling becomes cheaper and more transparent, and the same recipe may apply to other physiological signals.","feed_headline":"Compression beats pretraining: ECG tokens match learned encoders","feed_subtitle":"A byte-pair tokenizer lets an LLM answer ECG questions with 3x faster training and half the data.","key_machinery":"The central object is ECG-Byte, an adapted byte pair encoding (BPE) tokenizer for ECG signals. It works by normalizing each ECG to the range [0,1], quantizing amplitudes into 26 symbols (the lowercase letters a-z), flattening the 12 leads into one long symbol string, and iteratively merging the most frequent adjacent symbol pairs, exactly the BPE compression algorithm, into new token IDs up to a chosen number of merges. Encoding uses a trie for longest-match tokenization. What carries the argument is that these tokens are inserted directly into the LLM's vocabulary and input sequence, so the model learns to condition generation on the ECG with no separate encoder or projection layer, while the merge history makes every token traceable back to the waveform.","core_discovery":"The paper's discovery is that a purely rule-based compressor, byte pair encoding applied to quantized ECG amplitudes, can carry a 12-lead ECG signal into an LLM's vocabulary well enough for competitive conditional text generation, making a learned, self-supervised ECG encoder unnecessary. ECG-Byte quantizes each normalized sample into one of 26 symbols, flattens the 12 leads into a single symbol string, and runs BPE on a large concatenated corpus to learn merges that become new vocabulary tokens. The LLM is then finetuned with the sequence [BOS] [SIG START] <ECG tokens> [SIG END] <question> <answer> [EOS], with the loss computed only on the answer tokens. Empirically, against two-stage baselines that pretrain contrastive, masked-image, or dual-objective ECG encoders, ECG-Byte is competitive or superior across BLEU-4, Rouge-L, Meteor, and BertScore on both MIMIC-IV and PTB-XL question-answer sets, while requiring about 48% of the data and roughly one-third of the total training time.","pith_inferences":["One test the paper does not run is a text-only or ECG-shuffled control; without it, the competitive scores could partly reflect the LLM answering from question priors or template text rather than from the ECG tokens.","The paper's own attention analysis shows the model concentrating on the textual part of the input, which, if confirmed by a control, would weaken the claim that the ECG tokens are what drive the generated answer.","A practical extension is to benchmark ECG-Byte tokens against other discrete time-series encodings, such as symbolic aggregation approaches, for generative tasks; the paper proposes refined quantizers as future work but does not compare them.","If the conditioning-control issue is resolved, the same compression approach may generalize to EEG, PPG, or EMG language modeling, where interpretable reversible tokens could help clinicians audit model reasoning."],"forward_implications":["A dedicated self-supervised ECG encoder is not a necessary component for competitive ECG-to-text generation; a rule-based tokenizer suffices.","Training time and data requirements for ECG language models can be cut to roughly one-third and one-half, respectively, without sacrificing NLG quality.","Because ECG tokens are reversible, attention maps and token-usage statistics become interpretable in terms of P waves, QRS complexes, and T waves.","The approach works across different base LLMs and ECG segment lengths, with full 10-second ECGs giving the highest scores in the paper's ablations.","The same BPE-on-quantized-signal recipe can be applied to other physiological time series for generative tasks."],"supporting_citations":[{"why":"Introduces the byte pair encoding algorithm that ECG-Byte adapts for ECG compression.","marker":"(Gage, 1994)"},{"why":"Adapts BPE to subword tokenization in NLP, providing the tokenizer design that ECG-Byte extends to signals.","marker":"(Sennrich et al., 2016)"},{"why":"ECG-Chat, the main two-stage ECG-language baseline and the source of the MIMIC-IV ECG pretraining data used for comparison.","marker":"(Zhao et al., 2024)"},{"why":"Provides the ECG-QA dataset whose question-answer pairs are the evaluation benchmark.","marker":"(Oh et al., 2023)"},{"why":"Supplies the PTB-XL dataset used for cross-dataset evaluation and transfer tests.","marker":"(Wagner et al., 2020)"},{"why":"Demonstrates converting continuous values to discrete tokens, motivating the quantization step.","marker":"(Chen et al., 2022)"},{"why":"Earlier work applying quantization and BPE to compress ECGs for BERT classification, extended here to generative NLG.","marker":"(Tahery et al., 2024)"},{"why":"State-of-the-art contrastive ECG method reimplemented as the L_M_ERL baseline that ECG-Byte must match.","marker":"(Liu et al., 2024a)"},{"why":"Source of the [SIG START] and [SIG END] special-token design for marking the ECG segment in the input sequence.","marker":"(Liu et al., 2023)"},{"why":"LoRA, the parameter-efficient finetuning method used in all reported experiments.","marker":"(Hu et al., 2021)"}],"fun_headline_variants":["BPE tokenizer for ECG signals: no pretrained encoder needed","Rule-based ECG tokenizer matches learned encoding for LLM QA","End-to-end ECG language modeling via byte-pair encoding","Tokenizer replaces ECG encoder: train 3x faster, use half data","ECG tokenization without SSL: faster, lighter, still accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the LLM's answers are actually driven by the ECG tokens; nothing in the experiments removes or scrambles the ECG input to prove this, and the attention maps show most of the model's attention falling on the text.","fun_headline_variants_meta":{"raw":{"variants":["BPE tokenizer for ECG signals: no pretrained encoder needed","Rule-based ECG tokenizer matches learned encoding for LLM QA","End-to-end ECG language modeling via byte-pair encoding","Tokenizer replaces ECG encoder: train 3x faster, use half data","ECG tokenization without SSL: faster, lighter, still accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000415,"raw_usage":{"total_tokens":2162,"prompt_tokens":982,"completion_tokens":1180,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":1091}},"tokens_in":598,"tokens_out":1180,"duration_ms":11449,"temperature":1.0,"reasoning_tokens":1091,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:17:04.612585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact ECG-QA PTB-XL experiment with the ECG token block replaced by (a) no ECG tokens at all and (b) the same ECG tokens in random order. If BLEU-4, Rouge-L, Meteor, and BertScore do not drop substantially relative to the reported ECG-Byte numbers, the paper's claim that ECG-Byte conveys signal content is falsified.","supporting_citations":[],"review_version":1}