{"id":"b265c3cd-e975-454a-8258-d5d1ccbd917b","arxiv_id":"2509.18588","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"UniECG combines ECG interpretation and text-to-ECG generation in one model by fine-tuning a language model and aligning its output tokens with a pretrained ECG diffusion generator.","lead":"UniECG is a single AI model that can look at an ECG and write an evidence-based explanation, and can also create an ECG waveform when given a text description. It is aimed at medical students who need interactive examples for learning ECG interpretation, not at clinical diagnosis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage-2 alignment is trained on full ground-truth reports but deployed on short user prompts; no quantitative evidence that mapped embeddings yield faithful ECGs.","rationale":"The reader's weakest assumption identified the alignment transfer as the point most likely to fail. I agree, and refine it with a concrete mechanism: the training uses ground-truth full reports in the context, making the alignment task easier and less representative of deployment. The exposure bias and distribution shift are testable and directly target the generation claim. The paper's lack of quantitative generation metrics means this concern is not answered by existing evidence. Since the reader's verdict is CONDITIONAL, and my concern reinforces the need for additional evidence rather than proving failure, the verdict should remain CONDITIONAL. The proposed test would settle whether the concern lands, by isolating the effect of prompt style and measuring generation quality objectively.","tokens_in":5382,"tokens_out":5712,"duration_ms":44692,"concrete_test":"Construct a test set of 100 short textual conditions (e.g., 'atrial fibrillation', 'normal sinus rhythm, HR 80') paired with their corresponding full MIMIC-IV-ECG reports. For each condition, generate ECG signals using UniECG under two conditions: (a) given the full report as the prompt (training-like), and (b) given only the short condition (realistic user input). Then: (1) run a pre-trained ECG classifier on the generated signals to measure diagnostic accuracy per condition; (2) compute distributional similarity (e.g., Fréchet distance on ECG feature embeddings) between generated signals and real MIMIC-IV-ECG signals with the same label. If performance for short-condition prompts is substantially worse than for full-report prompts, the alignment does not transfer to user-like inputs, undermining the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of high-fidelity, text-conditioned ECG generation depends entirely on the latent-space alignment in §2.2. Eq. (3) minimizes MSE between the mapper output and the frozen DiffuSETS text encoder embedding Tψ(y) for the same report y. However, the training prompts are constructed from full MIMIC-IV-ECG reports: the query is 'Generate an ECG of ...' and the answer is 'This is an ECG of ...' where the ellipsis is the full textual report. The [ECG] tokens are therefore trained on hidden states computed after the LLM has already seen the complete report. Ld can be minimized by attending to and copying the report content already in context. At inference, users provide short conditions (e.g., 'atrial fibrillation'), and the LLM must generate its own answer prefix autoregressively; the resulting [ECG] hidden states are conditioned on a self-generated, abbreviated prefix, outside the training distribution. This exposure-bias and distribution-shift problem is not addressed. Moreover, the paper provides no quantitative evaluation of generation fidelity or condition alignment, only a single qualitative figure (Fig. 2, right panel). If the mapping generalizes poorly to short, self-generated prompts, the generated ECGs will be condition-mismatched or low-fidelity, directly falsifying the headline generation capability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"UniECG is a two-stage ECG multimodal large language model. In stage 1, the LLM is fine-tuned on ECG signal-image-text triplets to produce evidence-based interpretations. In stage 2, the LLM is frozen, special [ECG] tokens are added, and a lightweight mapper is trained to align the hidden states of those tokens with the text-encoder embeddings of DiffuSETS, a pretrained text-conditioned ECG diffusion model. At inference, the model chooses between interpretation and generation based on the user input; generation is performed by feeding the mapped [ECG] token states into the frozen DiffuSETS decoder. The paper reports quantitative understanding results on the ECG-Grounding benchmark scored by DeepSeek V3.1, and qualitative generation results in a single dialogue example.","tokens_in":5724,"tokens_out":4434,"duration_ms":35459,"significance":"If the generation capability were quantitatively supported, a single model that both interprets real ECGs and synthesizes signal-level ECG waveforms from textual prompts would be a meaningful step beyond prior ECG MLLMs, which are understanding-only. The two-stage design—decoupling understanding from generation via latent-space alignment to a frozen diffusion decoder—is an elegant way to add generation without catastrophic forgetting, and it builds on prior work (GILL, DiffuSETS) in a sensible way. The authors explicitly position the model as an educational aid rather than a clinical diagnostic system, which is appropriate. However, the paper's central novelty is the unified model, and the evidence for the generation half is currently anecdotal. The understanding evaluation also needs more rigor before the trade-off against GEM can be interpreted. Strengths include the clear two-stage formulation, the use of publicly available datasets, and the promise of released code and checkpoints.","major_comments":[{"comment":"The alignment loss is trained on full MIMIC-IV-ECG reports: the query is 'Generate an ECG of ...' and the answer is 'This is an ECG of ...' with the complete report filling the ellipsis. The [ECG] hidden states are therefore computed after the LLM has already seen the full report in context. At inference, users provide short prompts (e.g., 'atrial fibrillation'), and the LLM must generate its own answer prefix autoregressively before emitting [ECG] tokens. This is a distribution shift and an exposure-bias problem that is not addressed. The claim that minimizing MSE between mapped [ECG] states and DiffuSETS text embeddings transfers to short, self-generated prompts is an untested assumption. Please provide quantitative generation results on both full-report conditions and short-prompt conditions, including condition-alignment metrics, signal fidelity, and clinical correctness.","section":"§2.2, Eq. (3)"},{"comment":"The understanding results are point estimates from a single scoring run by DeepSeek V3.1, with no validation of the LLM judge, no error bars, and no repeated sampling. UniECG is below GEM on most metrics, including Diagnosis Accuracy (66.80 vs. 72.30), ECG Feature Grounding (69.44 vs. 73.78), and Average (63.98 vs. 69.59). The statement that UniECG 'largely retains' its interpretation ability is therefore not quantitatively supported; at best the tables show a trade-off. Report mean±std over multiple runs or provide a statistical test, and include a human-annotation or independent-judge validation for the metrics.","section":"§3.2, Tables 1–2"},{"comment":"The paper's headline claim includes 'high-fidelity ECG signals with fine-grained temporal features', but the only support is a single qualitative figure. There is no quantitative evaluation of generation: no signal fidelity metrics (e.g., morphology error, SNR, Fréchet distance), no condition-alignment metrics, no diversity measure, and no comparison to the underlying DiffuSETS baseline. Since generation is half of the claimed contribution and is the primary novelty over prior ECG MLLMs, this is a load-bearing omission. Please add a generation benchmarking section with quantitative metrics and a comparison to DiffuSETS (and ideally to the training data distribution).","section":"§3.2, Fig. 2"}],"minor_comments":[{"comment":"The notation in Eq. (2) conditions on s_{1:t}, but the stage-2 training is described as text-only and no signal s appears in the surrounding text. The conditioning variable should be the text tokens of the query/answer prefix; please correct.","section":"§2.2, Eq. (2)"},{"comment":"The ECG-Grounding benchmark is scored by DeepSeek V3.1; please specify whether this judge was validated against human ratings, and report the prompt used for scoring.","section":"§3.1"},{"comment":"The introduction mentions 'GPT-5' as an example of a unified model, but this model is not established in the literature as of the submission date. Please cite concrete released models (e.g., GPT-4V, LLaVA, BLIP-3) or remove the reference.","section":"§1"},{"comment":"The abstract says the model is 'a step toward interactive ECG education', but the paper contains no educational-efficacy evaluation. Please soften the educational claim or add a relevant user study.","section":"Abstract/§4"}],"recommendation":"major_revision","confidential_remarks":"The generation capability is essentially a learned interface to DiffuSETS, which is prior work from the same research group; the novelty is the unified understanding-and-generation interface, not the ECG generation itself. The ECG-Grounding benchmark used for understanding is also from the same group. Independent evaluation would significantly strengthen the paper. As a journal submission, the paper's empirical depth is thin; it reads more like a workshop or short-conference paper. The lack of quantitative generation evaluation is the main blocker for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a neat but under-evidenced idea. The paper combines ECG explanation and generation in one LLM by adding [ECG] tokens and learning a mapper from their hidden states to the text-embedding space of the frozen DiffuSETS diffusion model. As far as I know, no prior ECG model does both in one pass, so the setup is a real contribution. The two-stage training is clearly described, and freezing the LLM while only adding tokens in stage 2 is sensible. Understanding results are moderately positive: they beat PULSE by a wide margin and match GEM on lead coverage, though they trail on most other dimensions.\n\nThe soft spots are real and load-bearing on the generation side. There is no quantitative evaluation of generated ECGs: no fidelity, diversity, condition-match, or clinical correctness metrics, and no comparison to DiffuSETS itself, despite it being the same group's prior work. A single qualitative example with atrial fibrillation is not enough to support \"high-fidelity\" or \"remarkable generation ability.\"\n\nThe stress-test concern is valid. Eq. (3) aligns mapper outputs to the text embedding of the full MIMIC-IV-ECG report after the LLM has already seen that full report in context. At inference, users give short prompts like \"atrial fibrillation,\" so the model generates a short prefix and then emits [ECG] tokens conditioned on that prefix, which is outside the training distribution. The paper does not address this exposure-bias/distribution-shift problem. If the mapper generalizes poorly, the generated ECG will not match the condition. This is testable and should be checked with varying prompt lengths.\n\nThe understanding evaluation uses DeepSeek V3.1 as an unvalidated judge, with no error bars or human agreement. That weakens the quantitative claims, though the relative improvements over PULSE are large enough to be plausible.\n\nReproducibility is also a problem: no code or checkpoints are available, and the base LLM and key hyperparameters are unspecified. That makes verification harder.\n\nTo the paper's credit, it is transparent about being an educational aid, not a clinical tool, and it acknowledges notable limitations in the conclusion. The citation pattern leans heavily on the same group's prior work, but that is not itself a flaw—it just makes the missing comparison to DiffuSETS more conspicuous.\n\nI would send this to peer review rather than desk reject. The unified-model idea is worth exploring, and a referee can push for generation metrics and an analysis of short-prompt generalization. As submitted, the central generation claim is unsupported; the paper reads like a promising technical note, not a finished system.","headline":"Reasonable two-stage idea, but the generation claim rests on one qualitative example and a distribution shift the paper never addresses.","tokens_in":6181,"tokens_out":3364,"would_cite":false,"duration_ms":26859,"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":"UniECG proposes the first unified ECG model that can both produce evidence-based interpretations of ECG signals and images, and generate signal-level ECGs from text descriptions.","keywords":["ECG","multimodal large language model","text-to-ECG generation","ECG interpretation","latent space alignment","diffusion model","two-stage training","medical education"],"falsifier":"Generate ECGs from a set of textual reports with known reference classes (e.g., atrial fibrillation, left bundle branch block, heart rates 60 vs 120 bpm), then run a pretrained ECG classifier on the generated signals and compare its label agreement and morphology statistics (QRS duration, RR irregularity) against real MIMIC-IV-ECG waveforms with the same text conditions. If the classifier cannot distinguish conditions or morphology metrics deviate substantially from real data, the alignment has not transferred to the diffusion decoder.","tokens_in":5274,"feed_emoji":"🫀","tokens_out":3926,"duration_ms":32898,"temperature":0.7,"pith_summary":"This paper introduces UniECG, a single multimodal large language model that attempts to do both directions of the ECG task: given an ECG signal or image, it produces an evidence-based diagnostic explanation, and given a textual description, it generates a signal-level 12-lead ECG waveform. The proposal is a two-stage training recipe: first the model is fully fine-tuned to interpret ECG data, then its newly added [ECG] tokens are aligned in latent space with the text encoder of a frozen diffusion-based ECG generator (DiffuSETS), so the same LLM can emit tokens that the diffusion decoder turns into waveforms. The paper reports that understanding performance stays in the same range as the current grounding-focused model while gaining generative ability absent from previous ECG multimodal models. The motivation is interactive ECG education and case-based learning, not clinical diagnosis.","feed_headline":"One model interprets ECGs and synthesizes new ones from text","feed_subtitle":"UniECG answers evidence-based questions about real traces and generates signal-level ECGs from text descriptions for case-based learning.","key_machinery":"The load-bearing mechanism is latent-space alignment: the LLM vocabulary gains n special [ECG] tokens, and after the model is frozen, only these token embeddings are fine-tuned (Eq. 2). At inference, the final-layer hidden states of the [ECG] sequence are passed through an encoder–decoder Transformer mapper with learnable queries, and mean squared error (Eq. 3) pulls its outputs into the embedding space of DiffuSETS's frozen text encoder. DiffuSETS — a diffusion model that generates 12-lead ECGs conditioned on text reports — then decodes these aligned embeddings into waveforms. The trick is that the second training stage needs no ECG data at all, only MIMIC-IV-ECG text reports converted to '","core_discovery":"The central claim is that one model can be made capable of both evidence-based ECG interpretation and text-conditioned ECG signal generation, without sacrificing either. The two-stage design first teaches ECG-to-text reasoning by full-parameter fine-tuning on the ECG-Grounding dataset, then freezes the LLM and teaches text-to-ECG by adding trainable [ECG] tokens whose hidden states are projected by a lightweight mapper toward DiffuSETS's frozen text-encoder embeddings; the pretrained diffusion decoder then synthesizes the signal. The paper shows quantitative understanding scores (e.g., overall average 63.98 vs 69.59 for the current best comparison) and qualitative generation examples, arguin","pith_inferences":["An obvious next step the paper does not perform is quantitative evaluation of generations (e.g., classifier agreement with the conditioning text, beat-level morphology metrics); the current evidence for 'high fidelity' is qualitative.","The same two-stage latent-alignment recipe could transfer to other physiological time series where a text-conditioned diffusion model exists, such as PPG or EEG, turning a multimodal LLM into a signal synthesizer without waveform training.","Because the generation stage is trained only on reports, the model's generative behavior is bounded by the diversity of MIMIC-IV-ECG text; novel conditions not covered by report phrasing may not be synthesizable.","If the alignment lands in low-density regions of the frozen embedding space, generations could look plausible yet be clinically inconsistent — a failure mode a user-facing educational tool would need to guard against."],"forward_implications":["A single model can switch between explaining a real ECG and generating a synthetic one based solely on user wording, enabling dialogue-style case-based learning.","Signal-level (not image-only) outputs mean generated ECGs can be fed to downstream deep-learning models, potentially supporting data augmentation for rare or under-represented cardiac conditions.","The generation skill is acquired without waveform supervision, since the frozen diffusion decoder supplies signal fidelity while the LLM learns only when and how to emit [ECG] tokens.","The decoupling means understanding fine-tuning and generation alignment can be done independently; one can upgrade the generator without retraining interpretation.","If the approach holds, ECG education tools could let a student request 'an ECG with atrial fibrillation and a heart rate of 110' and then question the model about the trace it just produced."],"fun_headline_variants":["ECG explainer and generator in one unified model","Two-way ECG: explains traces, generates from text","UniECG: learn ECG by explaining and generating","One model for ECG explanations and synthetic signals","From ECG to text and back: UniECG does both"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The generation half of the central claim rests on the assumption that aligning [ECG] token hidden states with DiffuSETS's frozen text-encoder embeddings (using only text, no waveforms) yields embedding-space regions from which the diffusion decoder produces condition-faithful, clinically plausible ECG signals; the paper presents only qualitative support for that transfer.","fun_headline_variants_meta":{"raw":{"variants":["ECG explainer and generator in one unified model","Two-way ECG: explains traces, generates from text","UniECG: learn ECG by explaining and generating","One model for ECG explanations and synthetic signals","From ECG to text and back: UniECG does both"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1257,"prompt_tokens":699,"completion_tokens":558,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":443,"completion_tokens_details":{"reasoning_tokens":481}},"tokens_in":443,"tokens_out":558,"duration_ms":4593,"temperature":1.0,"reasoning_tokens":481,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:41:43.830673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate ECGs from a set of textual reports with known reference classes (e.g., atrial fibrillation, left bundle branch block, heart rates 60 vs 120 bpm), then run a pretrained ECG classifier on the generated signals and compare its label agreement and morphology statistics (QRS duration, RR irregularity) against real MIMIC-IV-ECG waveforms with the same text conditions. If the classifier cannot distinguish conditions or morphology metrics deviate substantially from real data, the alignment has not transferred to the diffusion decoder.","supporting_citations":[],"review_version":1}