{"id":"43c35699-4069-41c9-9831-9d36fe9f94c8","arxiv_id":"2507.01045","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A multimodal foundation model for ECG and PPG, pretrained on 1.7 million individuals, is reported to transfer across lead configurations and sensing modalities on diagnosis, demographics, blood pressure, outcome prediction, and ECG QA tasks.","lead":"A machine learning model was pretrained on ECG, PPG, and text data from about 1.7 million people, then adapted to diagnose heart conditions, estimate blood pressure, predict outcomes, and answer ECG questions. The authors report that this single model transfers across different devices and lead configurations better than task-specific models trained from scratch, though some of the reported comparisons in the paper contradict that claim.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own numbers contradict its headline: in the CVD diagnosis section, CSFM macro-F1 is below the best baseline on CinC17, PTB-XL, and SimBand (0.634 vs 0.677, 0.328 vs 0.357, 0.357 vs 0.398), while the text claims it substantially outperforms.","rationale":"The paper's strongest claim is that CSFM 'consistently outperforms' one-modal-one-task models. The 'Cardiovascular Disease Diagnosis' paragraph, which is the first and most direct test of that claim, reports CSFM macro-F1 values below the comparison model on all three datasets. This is not a gap in evidence; it is the reported evidence pointing in the opposite direction. The accompanying sentence claiming that the CSFM series 'substantially outperforms' conventional learning strategies cannot be true simultaneously. I therefore do not think the central claim can be accepted as written. The reader's channel-agnostic concern is real, but secondary: an explicit architecture description would help, yet even with perfect channel-agnostic transferability the headline 'consistently outperforms' is unsupported by the diagnostic numbers. The CODE-15 overlap-control ambiguity is also serious, but the CVD contradiction is more decisive because no overlap confound is needed to see it. A fixed experimental rerun or a simple textual correction would settle the issue; if the numbers are correct, the central claim is false and rejection is warranted. If they are a typo, the manuscript needs correction before further evaluation.","tokens_in":15978,"tokens_out":7132,"duration_ms":73252,"concrete_test":"Extract the exact macro-F1 values for CSFM and the best non-CSFM baseline from Figure 3a for CinC17, PTB-XL, and SimBand, and independently rerun the PTB-XL 44-class diagnostic experiment using the same subject-wise split (80/10/10) with the released CSFM weights once available. If CSFM's macro-F1 is not above the reported 0.357 baseline, the central claim is false as stated; if the text was a swap, the corrected values must be published before the claim can be evaluated.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing assumption of the central claim is falsified by the manuscript's own reported results. In the 'Cardiovascular Disease Diagnosis' paragraph (Results, Figure 3a), the authors write that on CinC17 CSFM achieved macro-F1 0.634 (95% CI: [0.558, 0.710]) versus 0.677 ([0.656, 0.699]); on PTB-XL 0.328 ([0.296, 0.361]) versus 0.357 ([0.338, 0.377]); and on SimBand 0.357 ([0.324, 0.391]) versus 0.398 ([0.279, 0.516]). In all three cases the CSFM value is lower than the comparison value, yet the same paragraph asserts that 'In most cases, our CSFM model series substantially outperforms conventional learning strategies,' and the Abstract and Conclusion claim that CSFM 'consistently outperforms traditional one-modal-one-task approaches.' If these numbers are accurate, the flagship diagnostic evidence directly refutes the central claim. If they are a typographical swap, the manuscript must be corrected and Figure 3a and Table 2 checked for consistency. This is not a matter of statistical uncertainty: the confidence intervals for CinC17 and PTB-XL do not overlap, so the comparison is unambiguously against CSFM. This concern is more decisive than the channel-agnostic question because, even granting the architecture-level transferability, the paper's own numbers fail the headline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CSFM, a family of transformer-based cardiac sensing foundation models pretrained with masked modeling on heterogeneous ECG, PPG, and text data from roughly 1.7 million individuals across MIMIC-III-WDB, MIMIC-IV-ECG, and CODE. The authors claim that CSFM consistently outperforms conventional one-modal-one-task models across five downstream scenarios: cardiac disease diagnosis, demographic information recognition, vital sign measurement, clinical outcome prediction, and ECG question answering, while also transferring across ECG lead configurations (12-, 6-, 2-, 1-lead) and across ECG-only, PPG-only, and combined modalities. The paper further reports that CSFM embeddings outperform hand-crafted features and features from prior ECG/PPG foundation models, and that the model supports cross-modality waveform reconstruction and augmentation. The central claim is that a single pretrained backbone can serve multiple tasks and input configurations without task- or device-specific architectural changes.","tokens_in":16459,"tokens_out":3866,"duration_ms":36633,"significance":"If the central claims were fully supported, the work would be a substantial contribution: a large-scale, multi-modal, channel-agnostic foundation model for cardiac signals with broad downstream applicability would be of high practical value, especially for settings with limited lead configurations or wearable-only data. The paper's strengths include its very large pretraining corpus, the breadth of downstream evaluations against multiple baselines, and the explicit comparison with existing foundation models and hand-crafted features. However, the paper's own reported numbers in the cardiovascular disease diagnosis section directly contradict the headline claim of consistent superior performance, and the CODE-15 mortality evaluation appears to involve a subset of the pretraining data. These issues are load-bearing because they concern the central claim of the paper, and they must be resolved before the findings can be accepted.","major_comments":[{"comment":"The reported macro-F1 values for CSFM are lower than the best conventional baseline on all three diagnostic datasets: CinC17 0.634 (95% CI [0.558, 0.710]) versus 0.677 ([0.656, 0.699]), PTB-XL 0.328 ([0.296, 0.361]) versus 0.357 ([0.338, 0.377]), and SimBand 0.357 ([0.324, 0.391]) versus 0.398 ([0.279, 0.516]). The text immediately following states that 'In most cases, our CSFM model series substantially outperforms conventional learning strategies,' and the Abstract and Conclusion claim that CSFM 'consistently outperforms traditional one-modal-one-task approaches.' This is a direct self-contradiction: the confidence intervals for CinC17 and PTB-XL do not overlap, so the comparison is statistically unambiguous. The authors must either correct the reported numbers or substantially temper the claims; as written, the flagship diagnostic evidence fails the paper's central claim.","section":"Results, Cardiovascular Disease Diagnosis (Wearable ECG, PPG, 12-Lead ECG)"},{"comment":"The evaluation of 1-year mortality prediction on CODE-15 is potentially circular because CODE-15 is described as 'public small version of CODE-Full' and CODE-Full is one of the pretraining datasets. If CODE-15 is a subset of CODE-Full, then the pretraining data contains the very recordings (or subjects) used for downstream validation. The sentence 'we ensured that no training subjects in CODE-Full is available in validation/testing subset of CODE-15' is unclear and grammatically ambiguous; if the intended meaning is that CODE-15 subjects were excluded from pretraining, this needs a precise explanation of how that was possible given that CODE-15 is a subset of CODE-Full. Without such clarification, the reported AUC of up to 0.844 on CODE-15 cannot be interpreted as an external validation result.","section":"Results, Clinical Outcome Prediction (ECG, PPG)"},{"comment":"The claim that CSFM is 'channel-agnostic, enabling direct transfer learning without the need to reinitialize input-specific layers' is load-bearing for the entire multi-lead and multi-modality generalization story, yet the paper provides no architectural specification of how the model accepts variable channel counts or arbitrary subsets of channels. The main text does not describe how leads are tokenized, how channel masking is applied at inference time, or how positional encodings handle different numbers of channels. An explicit description of the mechanism, or an ablation demonstrating that the model does not degrade simply because of input-size handling, is needed to support the transfer claims in Figure 4 and Table 2.","section":"Transfer from 12-Lead to Fewer-Lead Settings"},{"comment":"The text in the 'Transfer from 12-Lead to Fewer-Lead Settings' paragraph states that 'CSFMs consistently outperform conventional approaches,' but Table 2 shows that CSFM-Tiny is not consistently superior: for example, in the 6-lead 100% condition CSFM-Tiny achieves macro-F1 0.268, which is below ResNet1d18's 0.290, and in the 2-lead 100% condition CSFM-Tiny's 0.267 is below ResNet1d18's 0.272. The claim of consistent superiority is therefore not supported by the table's own numbers, even setting aside the diagnostic results in Figure 3a. The authors should either revise the claim or report the comparison in a way that is faithful to the full set of results.","section":"Table 2, transfer learning results"}],"minor_comments":[{"comment":"The word 'reocgnition' appears in the sentence 'demographic information reocgnition' and should be corrected to 'recognition.'","section":"Figure 1 caption"},{"comment":"The phrase 'measured by meas absolute error' should read 'measured by mean absolute error (MAE).'","section":"Results, Demographic Information Recognition"},{"comment":"There is a duplicated phrase in the sentence 'This was compared with that of the Fusion Transformer model introduced in Ohet al.32, as well as with as well as with LLaVA'; one 'as well as with' should be removed.","section":"Results, ECG Question Answering"},{"comment":"The text refers to 'as shown in Figure 3d' and 'as shown in Figure 3e' with a missing closing parenthesis after the second reference, and the sentence structure should be repaired for clarity.","section":"Results, Clinical Outcome Prediction"},{"comment":"The reference 'Ranftlet al.26' is missing a space, and the phrase 'with R-squared value calculated' appears in the Figure 3 caption in a stilted manner; a fuller description of the metric would be clearer.","section":"Methods and captions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claim is contradicted by its own reported diagnostic numbers in Section 'Results, Cardiovascular Disease Diagnosis' (Figure 3a). If the numbers are a typographical error, the authors must correct them and then the central claim may survive; if the numbers are accurate, the claim of consistent superior performance is false. The CODE-15 circularity concern also requires a clear explanation. The paper has substantial scope and data, but the flagship evidence needs to be reconciled before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one for the pretraining recipe, not for the headline. The paper builds CSFM, a transformer masked-model pretrained on ECG, PPG, and text from ~1.7M individuals (MIMIC-III-WDB, MIMIC-IV-ECG, CODE), then evaluates it across diagnosis, demographics, blood pressure, outcome prediction, and ECG QA. The genuinely new bits are the joint multi-modal pretraining and channel-masked pretraining that let one backbone take 12-, 6-, 2-, or 1-lead ECG and ECG/PPG/combined inputs without architectural changes. The cross-modality reconstruction results (PPG-to-ECG, single-lead-to-12-lead) are interesting and go beyond the usual fixed-format ECG foundation models. The evaluation breadth is a real strength, including comparisons to ECG-FM and PaPaGei embeddings.\n\nBut the central claim as written is not supported by the paper's own numbers. In the Cardiovascular Disease Diagnosis paragraph they report CSFM macro-F1 below the comparison baseline on all three datasets: 0.634 vs 0.677 on CinC17, 0.328 vs 0.357 on PTB-XL, and 0.357 vs 0.398 on SimBand. The CIs for CinC17 and PTB-XL don't overlap. Then the same paragraph says CSFM 'substantially outperforms' conventional methods, and the abstract/conclusion say 'consistently outperforms.' That is a direct contradiction, not a statistical quibble. Either the numbers are a typo (then Figure 3a and Table 2 need to be checked for consistency) or the main claim is false. This needs to be fixed before the paper is credible.\n\nOther soft spots: the CODE-15 overlap control is asserted ('no training subjects in CODE-Full is available in validation/testing subset of CODE-15') but the procedure is not detailed; the channel-agnostic property is claimed without architecture-level specification or a dedicated ablation proving variable channel counts at inference; the code and weights are promised but not yet released. None of these are fatal, but they add friction.\n\nDespite the reporting problem, the technical direction is solid and the experiments are extensive. The paper deserves a serious referee; it should not be desk-rejected. But the referee should require a corrected headline claim, a transparent explanation of the CODE-15 exclusion, and a proper description of the channel-agnostic mechanism.\n\nRecommendation: send to peer review, but with a note that the current abstract and conclusion overstate the results. I would not cite it until the numbers are reconciled and code/weights are out.","headline":"A credible multi-modal cardiac foundation model whose headline 'consistently outperforms' is directly contradicted by its own reported macro-F1 numbers in the CVD diagnosis section.","tokens_in":16928,"tokens_out":2485,"would_cite":false,"duration_ms":22527,"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 single transformer pretrained on heterogeneous ECG, PPG, and text from 1.7 million people learns channel-agnostic cardiac representations that outperform bespoke one-task models across diagnosis, demographics, blood pressure, outcomes…","keywords":["cardiac sensing foundation model","masked pretraining","ECG","PPG","channel-agnostic transfer","multimodal representation learning","clinical outcome prediction","ECG question answering"],"falsifier":"Take the released checkpoint and fine-tune it on a lead configuration that was not emphasized in pretraining—for example, a three-lead montage such as leads I, aVF, and V3—using the paper's standard pipeline. If performance collapses unless the input projection is reinitialized or the lead ordering is permuted, the channel-agnostic claim is falsified; a simpler check is whether the tokenizer or positional indexing assumes a fixed maximum channel count.","tokens_in":15761,"feed_emoji":"🫀","tokens_out":11725,"duration_ms":107763,"temperature":0.7,"pith_summary":"This paper argues that a single pretrained transformer can replace the standard one-dataset-per-task approach to cardiac monitoring. The authors train a cardiac sensing foundation model (CSFM) with masked modeling on heterogeneous electrocardiogram (ECG), photoplethysmogram (PPG), and text-report data from roughly 1.7 million people, then fine-tune it for five clinical scenarios: disease classification, demographic recognition, blood-pressure measurement, outcome prediction, and ECG question answering. Across those tasks the model consistently outperforms conventional one-modal, one-task networks trained from scratch, and the same checkpoint transfers across 12-, 6-, 2-, and single-lead ECG configurations and across ECG-only, PPG-only, and combined inputs. If the claim holds, one scalable backbone could serve both hospital 12-lead systems and wearables that record a single channel, reducing the need for bespoke models per device.","feed_headline":"One pretrained model spans 12-lead ECG to single-lead PPG","feed_subtitle":"Masked pretraining on 1.7M patients' heart signals and reports beats task-specific baselines across five clinical tasks.","key_machinery":"The load-bearing mechanism is masked modeling along both time and channel dimensions. During pretraining, random segments of the input waveform—and, for multi-channel data, random channels—are obscured and the transformer must reconstruct them from the remaining context, so the model learns to represent cardiac activity without assuming a fixed channel set. This is what the paper calls being channel-agnostic, and it is the property that enables the same checkpoint to be fine-tuned on 12-lead, 6-lead, 2-lead, or single-lead inputs and on ECG or PPG without reinitializing input-specific layers. A dense regression head can be added on top of the pretrained transformer for waveform-level tasks such as blood-pressure reconstruction and cross-modality signal generation.","core_discovery":"The central discovery is that a transformer pretrained generatively on cardiac signals with channel-wise and temporal masking learns representations that are not locked to any particular lead count or sensor type. The paper reports that CSFM embeddings, used either by fine-tuning or as frozen features with a classical classifier, outperform conventional convolutional, recurrent, and autoencoder baselines trained from scratch across the five evaluated scenarios. It further reports that the model transfers from 12-lead pretraining to 6-, 2-, and 1-lead settings without reinitializing input layers, that it works with ECG alone, PPG alone, or both together, and that it can reconstruct PPG-based ECGs and single-lead-to-12-lead ECGs, generating synthetic waveforms that support transferable classification. The authors interpret this as evidence that heterogeneous, multi-modal pretraining yields a versatile cardiac sensing backbone.","pith_inferences":["If the channel-agnostic property is real, then varying which channels appear during pretraining is itself a form of augmentation for lead invariance; an ablation that randomizes the visible channel set on each training step would test this directly.","The paper's result that text-paired pretraining helps single-lead performance even when it shrinks the signal-only dataset suggests clinical text carries cross-lead knowledge; a controlled comparison with text removed from pretraining would isolate that effect.","The same masked-channel recipe could apply to other multi-channel physiological recordings, such as EEG montages or multiparameter ICU waveforms, where device heterogeneity currently forces per-montage retraining."],"forward_implications":["A single deployed checkpoint could serve a 12-lead diagnostic ECG at a hospital and a single-lead wearable at home, because the input layer does not need to be rebuilt for each device.","New sites with very little labelled data could fine-tune from the pretrained backbone; the paper finds fewer-lead fine-tuning with 10% of the training set reaches performance comparable to conventional models trained on 100%.","The embeddings work as a frozen feature extractor, so lightweight deployment is possible by feeding them to a classical classifier rather than fine-tuning the whole transformer.","Cross-modality generation means PPG-only devices could be used to synthesize ECG waveforms for atrial-fibrillation screening, and single-lead devices could synthesize 12-lead waveforms for diagnosis.","The same pretrained representations support long-term mortality prediction and short-term ICU false-alarm detection, suggesting one model can cover both risk-stratification and acute-care monitoring."],"supporting_citations":[{"why":"Supplies ECG/PPG waveform segments with matched clinical text reports used in pretraining.","marker":"[22]"},{"why":"Supplies 12-lead ECG recordings with machine-generated reports for pretraining.","marker":"[23]"},{"why":"Supplies the large private 12-lead ECG corpus used for pretraining and mortality-related labels.","marker":"[11,24]"},{"why":"Provides the 44-class public ECG benchmark used for diagnosis across 12-, 6-, 2-, and 1-lead settings.","marker":"[7]"},{"why":"Provides the intraoperative physiological database used for demographic recognition and blood-pressure reconstruction.","marker":"[25]"},{"why":"Provides the ECG question-answering benchmark and the fusion-transformer baseline used for comparison.","marker":"[32]"},{"why":"Provides the ICU false-alarm dataset used to test ECG-only, PPG-only, and combined inputs.","marker":"[9]"},{"why":"Provides the single-lead ECG challenge data used for atrial-fibrillation classification.","marker":"[6]"},{"why":"Provides the public subset used for 1-year mortality prediction from 12-lead ECGs.","marker":"[10]"},{"why":"Defines the transfer protocol for moving from 12-lead to fewer-lead fine-tuning with reduced training fractions.","marker":"[34]"}],"fun_headline_variants":["One model for ECG and PPG across lead counts","Pretrained on 1.7M hearts, transfers across sensors and leads","Generative pretraining makes cardiac model work on ECG and PPG","Multi-modal cardiac foundation model beats single-task baselines","From 12-lead ECG to single-lead PPG: one model transfers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire transfer claim rests on the assumption that the channel-masking pretraining makes the transformer genuinely channel-agnostic, meaning any subset of leads or either modality can be fed in at inference time without reinitializing any input-specific layer; the paper asserts this property but provides no architecture-level specification or ablation that proves it.","fun_headline_variants_meta":{"raw":{"variants":["One model for ECG and PPG across lead counts","Pretrained on 1.7M hearts, transfers across sensors and leads","Generative pretraining makes cardiac model work on ECG and PPG","Multi-modal cardiac foundation model beats single-task baselines","From 12-lead ECG to single-lead PPG: one model transfers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000307,"raw_usage":{"total_tokens":1800,"prompt_tokens":1029,"completion_tokens":771,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":684}},"tokens_in":645,"tokens_out":771,"duration_ms":8282,"temperature":1.0,"reasoning_tokens":684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:37:03.816250+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the released checkpoint and fine-tune it on a lead configuration that was not emphasized in pretraining—for example, a three-lead montage such as leads I, aVF, and V3—using the paper's standard pipeline. If performance collapses unless the input projection is reinitialized or the lead ordering is permuted, the channel-agnostic claim is falsified; a simpler check is whether the tokenizer or positional indexing assumes a fixed maximum channel count.","supporting_citations":[{"cited_title":"Data9, 279 (2022)","cited_arxiv_id":null,"evidence_quote":"Provides the intraoperative physiological database used for demographic recognition and blood-pressure reconstruction."},{"cited_title":"& Choi, E","cited_arxiv_id":null,"evidence_quote":"Provides the ECG question-answering benchmark and the fusion-transformer baseline used for comparison."},{"cited_title":"Neural Inf","cited_arxiv_id":null,"evidence_quote":"Provides the ICU false-alarm dataset used to test ECG-only, PPG-only, and combined inputs."},{"cited_title":"D.et al.Af classification from a short single lead ecg recording: The physionet/computing in cardiology challenge 2017","cited_arxiv_id":null,"evidence_quote":"Provides the single-lead ECG challenge data used for atrial-fibrillation classification."},{"cited_title":"M.et al.Deep neural network-estimated electrocardiographic age as a mortality predictor.Nat","cited_arxiv_id":null,"evidence_quote":"Provides the public subset used for 1-year mortality prediction from 12-lead ECGs."}],"review_version":2}