{"id":"62856c6b-58a5-454d-855d-18eaec724281","arxiv_id":"2608.03690","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 7M-parameter graph ECG foundation model pre-trained with random lead dropout matches 12-lead models on full input and beats zero-padded baselines on most datasets with 1-2 leads.","lead":"LAEF is a small ECG model that represents heart recordings as flexible graphs, so it can diagnose from one wearable lead or the full clinical 12 without retraining. Read this to see whether a single-lead smartwatch reading can inherit the accuracy of models trained on millions of hospital ECGs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 'on par at 12 leads' is contradicted by Table 5: LAEF is significantly worse than the best baseline on most datasets, so the claimed no-trade-off lead-agnostic advantage is overstated.","rationale":"Reader's weakest_assumption (simulated vs real point-of-care domain shift) is real and acknowledged in Limitations; but it is an external-validity caveat, not an internal contradiction. A more load-bearing problem is that the paper's own significance markers undercut the 'on par at L=12' claim. The central value proposition is a single lead-agnostic model that works in both clinical and POC settings; if full-lead performance is significantly worse on most benchmarks, the 'does not sacrifice performance' claim is false and the trade-off needs disclosure. This does not invalidate the reduced-lead results; LAEF still outperforms zero-padded baselines at 1-2 leads. But it changes how the paper should be read: the advantage is not 'free' and the abstract overstates. The paper is otherwise thorough (ablations, CKA, lead-importance), and no circularity is apparent. Code/checkpoints not released, which also supports conditional status. Verdict stays CONDITIONAL/UNCHANGED.","tokens_in":42258,"tokens_out":13119,"duration_ms":126912,"concrete_test":"Using Table 5's L=12 rows and the stated bootstrap procedure (n=1000), count how many of the 18 datasets have LAEF's AUROC significantly below the best baseline (absence of asterisk while below best). If the count is >3, replace the abstract 'on par' with 'competitive but significantly worse on a substantial minority/majority,' and report the 12-lead trade-off explicitly. Ideally rerun the bootstrap with released checkpoints/test splits to confirm the printed markers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims LAEF is 'on par with specialized 12-lead baselines over 12× larger at full lead availability,' and Section 4.1 says lead-agnostic pretraining 'does not sacrifice performance.' This is the load-bearing link between the clinical 12-lead and point-of-care value propositions. Table 5/Table 1 reports bootstrap asterisks for scores 'not statistically significantly worse than the best.' At L=12, LAEF carries no asterisk and is below the best score on most of the 18 datasets (e.g., PTB-XL All 91.5 vs 94.6; Georgia 87.6 vs 92.0; EchoNext 79.7 vs 83.0; SPH 96.3 vs 98.2; ZZU pECG 88.1 vs 90.8; CPSC-Extra 85.8 vs 88.1). If an unstarred score denotes a significant deficit under the paper's own bootstrap procedure, the 'on par' claim fails on roughly 13/18 datasets, not just a few. Even if 'on par' is read loosely as 'competitive', the paper's stronger claim that no performance is sacrificed is contradicted by its own significance markers. The reduced-lead gains may therefore come with a real 12-lead cost that the abstract understates; this is a correctness/evidence issue, not an external-domain-shift issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LAEF, a 7M-parameter lead-agnostic ECG foundation model that represents an ECG of any lead count as a variable-size spatiotemporal graph of per-lead temporal segments. A GAT encoder processes this graph, and the model is pre-trained on 9.2M 12-lead ECGs with masked node modelling and stochastic lead sampling so that reduced-lead inputs are processed natively without zero-padding. The authors evaluate LAEF against seven ECG foundation models on 18 downstream datasets at 12, 2, and 1 lead(s), reporting macro AUROC. They claim that LAEF is on par with much larger specialized 12-lead models at full lead availability and outperforms zero-padded alternatives on 17/18 and 14/18 datasets at one and two leads, respectively, with an average +3.2 AUROC gain. They also present representation-stability (CKA) analyses, a large lead-importance study over 164 conditions, and extensive ablations of the codebook, masking ratio, topology, encoder, and training dynamics.","tokens_in":42688,"tokens_out":6500,"duration_ms":71863,"significance":"If the claims were fully supported, LAEF would be a practically valuable contribution to point-of-care ECG diagnostics, because it directly addresses the fixed-12-lead limitation of existing foundation models. The evaluation is unusually broad: 18 datasets, 7 baselines, paired random-lead seeds, bootstrap significance testing, and detailed ablations. The paper also provides a transparent account of architecture, pretraining stages, hyperparameters, and computational cost, and it explicitly states the limitation that no real point-of-care data were used. The reduced-lead results are strong and the lead-agnostic graph formulation is a plausible mechanism. However, the headline claim of 12-lead parity is not supported by the paper's own significance markers, and the point-of-care wording overstates what simulated lead reduction can show. These issues are fixable by a careful rewrite of the claims, but they affect the central message.","major_comments":[{"comment":"The claim that LAEF is \"on par\" with specialized 12-lead baselines, and the statement in §4.1 that \"lead-agnostic pre-training does not sacrifice performance at full lead availability,\" are contradicted by the paper's own bootstrap markers. In Table 1, an asterisk denotes a score that is not statistically significantly worse than the best. At L=12, LAEF has an asterisk on only 3 of 18 datasets (PTB-XL Sub, CPSC2018, PTB), while on the majority it is below the best score without an asterisk; examples include PTB-XL All (91.5 vs 94.6), Georgia (87.6 vs 92.0), EchoNext (79.7 vs 83.0), SPH (96.3 vs 98.2), and ZZU pECG (88.1 vs 90.8). Thus the paper's own significance procedure indicates a real, often significant, 12-lead performance cost on most datasets. The abstract and §4.1 should be rewritten to quantify this trade-off and to distinguish \"broadly competitive in ranking\" from \"no sacrific","section":"Abstract and §4.1, Table 1/Table 5"},{"comment":"The reduced-lead experiments simulate point-of-care conditions by dropping leads from clinical 12-lead ECGs and fine-tuning all models on 12 leads only. The manuscript explicitly acknowledges in Limitations that real point-of-care data could not be included and that noise profile, electrode placement, and acquisition bandwidth remain unaddressed. Given this, the abstract's phrase \"point-of-care-oriented diagnostics\" and \"direct point-of-care-oriented diagnostics\" is stronger than the evidence. I recommend rephrasing these as \"simulated reduced-lead robustness\" or adding an explicit caveat in the abstract and conclusions. This is not an internal inconsistency, but it affects how far the headline result can be generalized.","section":"§4 Evaluation protocol and Limitations"}],"minor_comments":[{"comment":"Typo: \"with with a single randomly sampled lead\" contains a duplicated \"with.\"","section":"Abstract"},{"comment":"The per-lead macro AUROC values are reported as [0.854, 0.857], differing only in the third decimal, and then Z-scored into \"relative importance.\" Please clarify over which dimension the Z-scoring is computed and how such tiny AUROC differences yield stable category-level importance patterns; this would make the analysis easier to interpret.","section":"Figure 4 / Lead importance analysis"},{"comment":"The dataset table appears misaligned in the submitted text (e.g., the CODE row and the Unlabelled/Labelled subrows), making it hard to verify which datasets are used for SSL pretraining versus supervised fine-tuning. Please check the table formatting.","section":"Table 3"},{"comment":"LAEF's memory footprint at L=12 (up to 5858 MB) is much larger than several baselines due to explicit graph indexing; the table caption notes this, but the main text's emphasis on \"scales naturally\" should mention that the indexing overhead at full lead count is a limitation, even though throughput and latency at L≤2 are good.","section":"§8.2 / Table 7"},{"comment":"The causal statement that the reduced-lead advantage is \"primarily driven by LAEF's lead-agnostic architecture\" is supported only by a dissociation between CKA stability and classification AUROC. Since the paper labels this as \"suggests,\" please consider softening the contribution bullet that presents this as a demonstrated mechanism rather than a post-hoc interpretation.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"This is a substantial empirical study with an unusually broad benchmark and a genuinely useful idea. The main problem is overclaiming: the abstract and §4.1 assert 12-lead parity that the paper's own bootstrap asterisks deny. This is correctable by rewriting the claims and discussing the trade-off explicitly. I do not see a need to rerun experiments, but the revised manuscript must reconcile the text with Tables 1 and 5 and temper the point-of-care language. With those changes, the paper would be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know about this paper: it's a genuine advance in handling variable-lead ECGs, but the abstract overstates the 12-lead performance. The lead-agnostic graph architecture is new and the reduced-lead results are strong, but Table 5 shows LAEF is significantly worse than the best baseline on 14 of 18 datasets at L=12, contradicting the 'on par' claim.\n\nWhat's new: LAEF treats any lead subset as a variable-size spatiotemporal graph, with complete intra-lead edges and time-aligned inter-lead edges. Stochastic lead sampling during masked node modeling pretraining lets it natively handle 1-2 leads without zero-padding. The evaluation is broad: 18 datasets, 7 baselines, paired seeds, bootstrap significance testing, extensive ablations. The lead-importance analysis across 164 conditions is post-hoc and not used for fitting, so no circularity. The self-citations are for datasets and a baseline model, which is legitimate.\n\nSoft spots. The abstract's claim that LAEF is 'on par with specialized 12-lead baselines over 12× larger at full lead availability' does not survive contact with Table 5. On 14 of 18 datasets, LAEF's L=12 score carries no asterisk, meaning it is significantly worse than the best baseline under the paper's own bootstrap procedure. It is best or statistically tied on only four (CODE 15%, PTB-XL Sub, CPSC2018, PTB). So the conclusion that lead-agnostic pretraining 'does not sacrifice performance' is not supported. The reduced-lead advantage is real and clinically relevant, but the 12-lead cost is understated.\n\nAlso, the point-of-care evaluation simulates reduced leads by dropping leads from 12-lead clinical ECGs. The paper acknowledges it could not validate on real wearable data due to acquisition challenges. That's an honest limitation, but it means the headline 1-2 lead advantage may not fully transfer to wearable noise and electrode placement. Code and checkpoints are not yet released, which limits reproducibility.\n\nOverall, this is a solid empirical paper with a credible core finding. The 12-lead overstatement is a load-bearing weakness that should be fixed in revision, but it doesn't sink the main contribution. I'd send it to peer review, with a request that the authors correct the abstract and discuss the actual L=12 results.\n\nFor a reading group, it's worth discussing both the architecture and the gap between abstract claims and tables.","headline":"New lead-agnostic graph architecture with strong reduced-lead results, but the 'on par at 12 leads' claim is contradicted by the paper's own significance markers.","tokens_in":43175,"tokens_out":6933,"would_cite":true,"duration_ms":53725,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims a 7.1M-parameter ECG model that treats leads as graph nodes can match 12-lead foundation models and outperform zero-padded baselines on 1–2 lead inputs.","keywords":["ECG foundation model","lead-agnostic","point-of-care diagnostics","graph attention network","masked node modelling","stochastic lead sampling","12-lead ECG","wearable ECG"],"falsifier":"Collect a labelled dataset of ECG recordings from actual wearable single-lead or two-lead devices (e.g., smartwatch or handheld recorder) with ground-truth diagnoses, then run LAEF and a zero-padded 12-lead baseline on the same recordings. If LAEF's average AUROC advantage over the zero-padded baseline at L=1 or L=2 shrinks to zero or reverses, the claim that lead-agnostic training transfers to real point-of-care settings is disconfirmed; the paper itself states it could not validate on real point-of-care data.","tokens_in":42200,"feed_emoji":"🫀","tokens_out":5589,"duration_ms":54725,"temperature":0.7,"pith_summary":"The paper tries to establish that the lead-count mismatch between clinical 12-lead ECGs and point-of-care 1–2 lead devices is not an input-formatting problem but an architectural one. It introduces LAEF, a 7M-parameter ECG foundation model that represents any ECG as a variable-size spatiotemporal graph, so any subset of leads is a first-class input with no zero-padding. Pre-trained on 9.2M 12-lead ECGs via masked node modelling with stochastic lead sampling, LAEF matches specialized 12-lead foundation models up to 12× larger at full lead availability, and outperforms all zero-padded alternatives on 17 of 18 datasets with one randomly sampled lead and 14 of 18 with two leads, averaging +3.2 AUROC points. This matters because it suggests one model could serve both clinical and wearable cardiac diagnostics without retraining or signal reconstruction per device.","feed_headline":"Lead-agnostic ECG model wins on 1–2 leads, matches 12-lead giants","feed_subtitle":"Graph-based LAEF treats missing leads as missing nodes, beating zero-padded baselines by +3.2 AUROC on average.","key_machinery":"The key machinery is the spatiotemporal ECG graph: each lead is partitioned into S=20 temporal segments, each mapped by a 1D CNN embedder to a node; intra-lead edges form a complete subgraph over all segments of a lead (capturing beat-to-beat and morphological context), while inter-lead edges connect only time-aligned segments across leads (capturing the spatial synchrony of simultaneous cardiac projections). A Graph Attention Network operates on this graph, and because graph construction uses no lead identity, the model scales naturally with the number of active leads, modelling lead absence as missing nodes. Pre-training uses masked node modelling with per-lead codebook prototypes and stoc","core_discovery":"The paper's central claim is that treating ECG leads as graph nodes rather than fixed tensor channels makes a single foundation model natively lead-agnostic: any subset of the standard 12 leads can be processed without architectural modification, without zero-padding, and without knowing lead identity. LAEF splits each lead into 20 temporal segments, connects segments within a lead with full intra-lead edges and time-aligned segments across leads with inter-lead edges, and processes the resulting variable-size graph with a Graph Attention Network. During pre-training, random lead subsets are sampled per ECG, acting as a structured information bottleneck that forces the model to infer global","pith_inferences":["Because LAEF deliberately omits lead identity, it may generalize to non-standard electrode placements, but it also cannot exploit known clinical priors that certain conditions are best seen in specific leads; a variant with optional lead-conditioned embeddings could combine both properties.","The paper's reduced-lead evaluation simulates point-of-care inputs by dropping leads from 12-lead clinical ECGs; the authors list validation on real wearable recordings as future work, and that validation would settle whether the +3.2 AUROC advantage survives real-world noise, electrode shifts, and bandwidth differences.","The lead-importance analysis, which is stable at the population level but structured by category, could be used to choose which single lead a smartwatch or patch should prioritize for a given screening target—for example, inferior ischemia shows elevated relative importance in lead III.","The graph formulation suggests a direct extension to non-standard or mixed-device configurations, such as combining a smartwatch limb lead with a chest-patch lead, since the model accepts arbitrary lead subsets without identity information."],"forward_implications":["A single pretrained model can serve both full 12-lead clinical diagnostics and 1–2 lead wearable inference, eliminating the need for per-device retraining or 12-lead reconstruction.","Reduced-lead inference cost scales with active leads: at L=1–2 LAEF reports the best throughput and latency among evaluated models, supporting resource-constrained point-of-care deployment.","Lead-agnostic pre-training does not sacrifice full-lead performance: LAEF matches baselines over 12× larger at L=12, so lead robustness and clinical utility are not in tension.","Zero-padding is not merely inefficient; the representation analysis indicates it corrupts reduced-lead inputs for fixed-12-lead models, which is why their AUROC collapses even when their internal representations appear stable.","Single-lead performance is stable across all 12 standard leads (AUROC range <0.003) while retaining category-specific lead preferences, suggesting the model can be deployed on any single-lead device without knowing which lead is being used."],"supporting_citations":[{"why":"Supplies the Graph Attention Network encoder that processes the variable-size spatiotemporal graph.","marker":"Velickovic et al. 2017"},{"why":"Provides the masked prediction of discrete hidden units and iterative codebook refinement scheme adapted here to per-lead ECG segments.","marker":"Hsu et al. 2021"},{"why":"Defines the iterative masked-modelling ECG pre-training lineage and supplies the Ribeiro-dev development set and protocol used in ablations.","marker":"Coppola et al. 2025"},{"why":"Provides the benchmark baselines, fine-tuning protocol, and 18-dataset evaluation framework that the paper extends to reduced-lead inference.","marker":"Al-Masud, Alcaraz, and Strodthoff 2025"},{"why":"Representative zero-padding and lead-masking approach that LAEF is explicitly designed to supersede; supplies the comparison point for reduced-lead robustness.","marker":"Oh et al. 2022"},{"why":"Documents inter-lead redundancy in 12-lead ECGs, the physiological premise for stochastic lead sampling during pre-training.","marker":"Mason et al. 2024"},{"why":"Provides the CODE dataset, the largest pre-training source of 12-lead ECGs and a downstream evaluation benchmark.","marker":"Ribeiro et al. 2020"},{"why":"Provides PTB-XL, the standard multi-label ECG benchmark whose diagnostic, form, rhythm, sub, and super labels are used across several evaluation datasets.","marker":"Wagner et al. 2020"}],"fun_headline_variants":["Graph-based ECG model adapts to any lead set, beats 12-lead baselines","Lead-agnostic ECG: 7M params match 12-lead giants on 1-2 leads","ECG foundation model treats leads as nodes, wins on sparse leads","Missing leads no problem: LAEF graph model improves AUROC by 3.2","Small ECG model beats zero-padding on single-lead diagnostics"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reduced-lead evaluation assumes that deleting leads from a 12-lead clinical ECG faithfully mimics a point-of-care recording; if real wearable signals differ enough in noise, electrode placement, or bandwidth, the reported 1–2 lead advantage may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Graph-based ECG model adapts to any lead set, beats 12-lead baselines","Lead-agnostic ECG: 7M params match 12-lead giants on 1-2 leads","ECG foundation model treats leads as nodes, wins on sparse leads","Missing leads no problem: LAEF graph model improves AUROC by 3.2","Small ECG model beats zero-padding on single-lead diagnostics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000222,"raw_usage":{"total_tokens":1315,"prompt_tokens":795,"completion_tokens":520,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":412}},"tokens_in":539,"tokens_out":520,"duration_ms":4829,"temperature":1.0,"reasoning_tokens":412,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:39:55.544725+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a labelled dataset of ECG recordings from actual wearable single-lead or two-lead devices (e.g., smartwatch or handheld recorder) with ground-truth diagnoses, then run LAEF and a zero-padded 12-lead baseline on the same recordings. If LAEF's average AUROC advantage over the zero-padded baseline at L=1 or L=2 shrinks to zero or reverses, the claim that lead-agnostic training transfers to real point-of-care settings is disconfirmed; the paper itself states it could not validate on real point-of-care data.","supporting_citations":[],"review_version":1}