{"id":"9d1c7978-b91f-4d31-8eff-75744110682b","arxiv_id":"2608.12944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A shared cardiac encoder pretrained with delay-aware cross-modal JEPA on ECG, PPG, and PCG beats modality-specific self-supervised baselines on 25 downstream tasks.","lead":"This paper trains one AI model to understand the heart from three different sensors at once (ECG, PPG, and PCG) by predicting hidden heart states and learning when each signal arrives. A generalist reader might care because a single cardiac model could make wearables, stethoscope screening, and hospital monitors share one powerful representation without needing more labeled data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PPG-EXT pretraining may overlap PulseLM PPG downstream tasks; if so, the Table 1 PPG gains are leakage rather than cross-modal transfer.","rationale":"The paper is well structured: the two-stage curriculum is sensible, the ablations in Table 4 are informative, the delay aligner is visualized in Figure 8, and the ECG and PCG results are externally benchmarked. I do not object to the JEPA framing itself, and Eq. (1) is an idealization that the empirical claim does not strictly require to be literally true. The most load-bearing risk is not the conceptual latent-state assumption but the integrity of the PPG evaluation. The pretraining corpus PPG-EXT is itself a MIMIC PPG benchmark, and the downstream PPG tasks MIMIC AF and PPG Arrhythmia come from PulseLM, a MIMIC-based benchmark. Without an explicit patient-level or window-level exclusion, the paper's largest PPG improvements could reflect pretraining on the test distribution. This is especially hard to audit because code and checkpoints are not released and Tables 1 and 2 lack seed-level error bars, while Table 3 is the only table reporting variability. The reader's Eq. (1) concern is real but secondary: even if the shared-latent assumption were wrong, the method could still learn useful sensor-specific features through cross-modal prediction, and the empirical claim would stand on its own. Leakage, if present, directly invalidates the empirical claim on PPG. Because the ECG and PCG claims rely on separate corpora, the headline would be downgraded rather than fully rejected pending the overlap check. I therefore keep the reader's CONDITIONAL verdict but attach an explicit data-disjointness condition; if the overlap check confirms contamination, the PPG portion of the abstract and the Table 1 claims should be retracted.","tokens_in":20945,"tokens_out":8105,"duration_ms":82880,"concrete_test":"Obtain the PPG-EXT (MIMIC-III-Ext-PPG) patient and record identifiers and the PulseLM split files for the seventeen PPG tasks, especially MIMIC AF and PPG Arrhythmia. Count how many downstream test windows share a source patient or source recording with the PPG-EXT pretraining set. Then retrain CardioState-JEPA with any overlapping PPG windows removed from pretraining, or at minimum re-probe with a checkpoint pretrained only on non-overlapping PPG, and recompute the Table 1 classification and regression averages. If the MIMIC AF and PPG Arrhythmia AUROCs or the 80.4 and 9.1 averages drop materially, the PPG headline gain is contamination. If no overlap is found, this concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing claim is that one frozen shared encoder transfers across ECG, PPG, and PCG (Tables 1-3). The PPG portion of that claim depends on a pretraining/evaluation boundary that the paper never establishes. Section 4.1 states that Stage I pretrains on PPG-EXT [29], which is MIMIC-III-Ext-PPG, with 4.6M training recordings in Table 5, and that downstream PPG tasks use pre-processed data from PulseLM [34]. Table 1's PPG classification tasks include MIMIC AF and PPG Arrhythmia, which are standard MIMIC-derived PPG benchmarks. PulseLM is a PPG-text benchmark that very plausibly derives these tasks from the same MIMIC-III-Ext-PPG corpus. The only disjointness statement in Section 4.2, 'patient-disjoint splits,' refers to the linear-probe train/test split; nothing in the paper states that PPG-EXT pretraining windows were excluded from the PulseLM downstream tasks, or that pretraining and downstream patient IDs are disjoint. If the same recordings or patients appear on both sides, the frozen encoder has already seen the downstream PPG inputs during pretraining, so the 8.2-point PPG classification gain and the 10.9 to 9.1 regression MAE gain in Table 1 would reflect memorization rather than cross-modal transfer. The ECG and PCG results are less exposed to this specific issue, since MIMIC-IV-ECG versus PTB-XL/CPSC/CSN and BMD-HS versus CirCor/CinC are separate corpora, so the central claim would survive in weakened form; but the abstract's PPG figure is a major pillar of the headline. This is a concrete, checkable threat to the evaluation protocol rather than a conceptual disagreement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CardioState-JEPA proposes a two-stage joint-embedding predictive architecture that learns a single shared cardiac representation across ECG, PPG, and PCG. Stage I performs intra-modal masked latent prediction on large unimodal corpora (MIMIC-IV-ECG, PPG-EXT, BMD-HS); Stage II uses paired recordings for delay-aware cross-modal prediction with a learned, physiology-anchored delay aligner. The frozen shared encoder is then evaluated via linear probing on 25 downstream tasks, reporting average AUROC gains of 8.2 points for PPG classification, 18.8 points for PCG murmur detection, and 15.5 points for ECG classification over the best self-supervised signal baselines. The paper includes ablations over pretraining modalities, learning objectives, and auxiliary losses, as well as sensitivity analyses for loss weights, and visualizes the learned delay alignment and the shared latent space.","tokens_in":21309,"tokens_out":6099,"duration_ms":59344,"significance":"If the reported results hold, the paper makes a substantial contribution by demonstrating that heterogeneous cardiac signals can mutually supervise a single foundation model, a direction that differs from the current per-modality cardiac foundation models. Strengths include the two-stage curriculum, the explicit supervision of the delay aligner with independent physiological anchors (R-peak-to-S1 and pulse arrival time), the breadth of downstream evaluation (17 PPG tasks, 18 ECG settings, 2 PCG tasks), and the detailed ablations and loss-weight sensitivity analyses. The principal caveat is the potential pretraining/evaluation overlap for PPG, which directly affects a headline claim; the ECG and PCG results rest on separate corpora and are less exposed. The absence of released code/checkpoints and of seed-level variance for most tables are secondary but important reproducibility concerns.","major_comments":[{"comment":"The paper does not establish disjointness between the PPG pretraining corpus (PPG-EXT, 4.6M recordings) and the PPG downstream tasks obtained from PulseLM (e.g., MIMIC AF, PPG Arrhythmia, BIDMC RR, UQVital). The statement in A.6 that downstream segments are 'never seen during pretraining' is an assertion without supporting evidence, and the 'patient-disjoint splits' in Section 4.2 refer only to the linear-probe train/test split. If the same recordings or patients appear in both, the reported 8.2-point average PPG classification gain and the regression MAE improvement (10.9 to 9.1) in Table 1 reflect memorization rather than cross-modal transfer. Please either provide explicit evidence of non-overlap (e.g., recording IDs, patient IDs, or a formal exclusion protocol) or repeat the PPG evaluation with a pretraining set that provably excludes the downstream corpora.","section":"Section 4.1, Table 1, and A.6"},{"comment":"Only the PCG results in Table 3 report mean±std over three seeds; the headline PPG and ECG numbers are single values. Since the protocol is a frozen encoder plus a linear probe, rerunning with multiple seeds is computationally inexpensive, and without variance estimates the reader cannot judge whether differences such as PPG Arrhythmia 96.8 vs. 95.8, or the ECG gap between CardioState-JEPA and MoCo-v3, are significant. Please report mean±std (or confidence intervals) over at least three seeds for all reported numbers, or clearly state that the table entries are single runs and include a statistical significance analysis for the main comparisons.","section":"Tables 1, 2, and 4"},{"comment":"The evidence that cross-modal training produces a modality-invariant yet task-relevant shared space rests on t-SNE plots and silhouette values computed on 'held-out co-recorded' samples, but the paper does not specify which paired dataset these samples come from, how the held-out split is constructed, or the number of samples, and it reports no variance for the silhouette values. Because SensSmartTech, VitalDB, and EPHNOGRAM are used in Stage II pretraining, the held-out samples must be provably disjoint from the pretraining paired data for the silhouette claim to be meaningful. Please specify the dataset, split, and sample size, and consider adding a quantitative cross-modal correspondence test (e.g., cross-modal retrieval or k-NN classification of one modality against another) to directly support the shared-representation claim.","section":"Section 4.5 and A.9"}],"minor_comments":[{"comment":"The sensitivity analysis varies only lambda_cross, lambda_delay, and lambda_state; tau_max and sigma are free parameters that are not covered. Please include a sensitivity analysis for these or justify the chosen values.","section":"A.5 and Eqs. (3)-(4)"},{"comment":"The symbol H is overloaded: it denotes the encoded token sequences (e.g., H_m) and also the Smooth-L1 loss in Eq. (2). Please use a different symbol for the loss (e.g., L_huber) to avoid ambiguity.","section":"Section 3.4 and Eq. (2)"},{"comment":"The alignment kernel in Eq. (4) uses a softmax over 'in-band target tokens,' but the paper does not define how the in-band set is determined. Please specify the window size or selection rule.","section":"Section 3.5 and Eq. (4)"},{"comment":"The caption reports silhouette values of 'sil + 0.12' and 'sil -0.01' while the text reports 0.121 and -0.006; please make the numbers consistent.","section":"Figure 3 and Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations, and the ECG/PCG results are likely to survive even if the PPG overlap concern is confirmed. The central issue is whether the PPG headline numbers can be verified as contamination-free; if the authors can show disjointness or re-run with a clean pretraining set, the paper would be suitable for publication after revision. I would also suggest the editor consider whether code/checkpoint release should be a condition for acceptance given the foundation-model claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CardioState-JEPA is a serious attempt at a shared ECG/PPG/PCG foundation model, and I don't know of a prior JEPA-style latent-prediction approach with a learned physiological delay across these modalities. The two-stage curriculum is sensible, the delay aligner is supervised by independent beat-to-S1 and pulse-arrival anchors rather than downstream labels, and the frozen-encoder linear probe over 25 tasks is a clean protocol. The ablations over modality subsets, SSL objectives, and the three auxiliary losses are more thorough than most papers in this space, and the reported limitations are honest.\n\nThe thing to check before believing the headline numbers is a possible train/eval overlap in the PPG arm. Stage I pretrains on PPG-EXT (MIMIC-III-Ext-PPG, 4.6M recordings) and the downstream PPG tasks use PulseLM preprocessed data. MIMIC AF and PPG Arrhythmia are standard MIMIC-derived PPG benchmarks, PulseLM is plausibly built on the same corpus, and the paper never says that pretraining recordings, windows, or patients were excluded from the downstream tasks. The only 'patient-disjoint splits' statement refers to the linear-probe train/test split. If the same recordings appear on both sides, the 8.2-point PPG classification gain and the regression improvements are memorization, not cross-modal transfer. This is directly checkable and needs an explicit answer. The ECG and PCG results are less exposed to this because MIMIC-IV-ECG vs PTB-XL/CPSC/CSN and BMD-HS vs CirCor/CinC are separate corpora, so the central claim survives in weakened form, but the abstract's PPG figure is a major pillar.\n\nOther soft spots are more minor. Tables 1 and 2 lack seed-level error bars; Table 3 has them for PCG only. No code or checkpoints are released, which matters here because the delay alignment and the exact PPG preprocessing are hard to audit otherwise. The PCG improvement of 18.8 AUROC points is large enough that I'd want replication. The Eq (1) shared latent cardiac state is a modeling prior rather than a proved fact; the downstream gains support it indirectly, but they don't distinguish 'shared cardiac physiology' from 'a convenient joint embedding' without additional evidence like the delay alignment visualization.\n\nBottom line: this is a well-designed paper with one real evaluation risk. I'd send it to a serious referee, but the referee should require the authors to rule out the PPG overlap and release artifacts. If the PPG column is cleaned up, the cross-modal claim and the architecture are worth citing; right now I'd hold off citing until the overlap answer is public.","headline":"A thoughtful shared-cardiac-encoder paper whose PPG results may be contaminated by pretraining/evaluation overlap; worth peer review but needs the overlap ruled out.","tokens_in":21836,"tokens_out":4109,"would_cite":false,"duration_ms":36784,"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":"CardioState-JEPA claims that ECG, PPG, and PCG are time-delayed renderings of one shared cardiac state, and that a single frozen encoder trained with masked latent prediction plus delay-aware cross-modal prediction can learn that state…","keywords":["cardiac foundation model","joint-embedding predictive architecture","self-supervised learning","ECG","PPG","PCG","cross-modal alignment","delay-aware pretraining"],"falsifier":"Retrain Stage II on paired recordings whose cross-modal pairs have been shuffled so the target signal comes from a different patient while delay supervision still uses the source-only anchor; if the reported downstream gains persist, the cross-modal objective is not learning shared physiology, and if they collapse, the shared-state reading is supported.","tokens_in":20774,"feed_emoji":"🫀","tokens_out":10089,"duration_ms":81551,"temperature":0.7,"pith_summary":"This paper tries to establish that ECG, PPG, and PCG are not three separate signal families but three time-delayed renderings of one hidden cardiac state, and that a single model can learn that state. It introduces CardioState-JEPA, a two-stage joint-embedding predictive architecture that first performs masked latent prediction on abundant unimodal data and then delay-aware cross-modal prediction on scarce paired data. Evaluated as a frozen encoder on 25 downstream tasks, the shared encoder raises average PPG classification by 8.2 AUROC points, PCG murmur detection by 18.8 points, and ECG classification by 15.5 points over the best self-supervised single-signal baseline. If correct, this makes heterogeneous cardiac signals mutually supervising sources for one cardiac foundation model, reducing the need for labeled or clinically annotated ECG data.","feed_headline":"Shared cardiac encoder beats per-sensor models by 18.8 AUROC","feed_subtitle":"One frozen model pretrained on ECG, PPG, and PCG lifts PPG, murmur, and ECG scores across 25 tasks.","key_machinery":"The load-bearing object is the latent cardiac state $c(t)$ of Eq. (1), with each modality modeled as a rendering $x_m(t)=O_m(c(t-\\tau_m(t)), u_m(t))$. The carrying mechanism is delay-aware cross-modal prediction: a learned delay head estimates per-token offsets $\\tau_{m\\to n}$, a Gaussian kernel gathers the target token at the shifted time, and the shared predictor matches source predictions to those aligned targets, with delay supervision anchored to R-peak-to-S1 and pulse-arrival-time intervals. The shared predictor and momentum encoder keep both prediction objectives in one space. This machinery forces the encoder to remove sensor-specific appearance and temporal offset so that downstream linear probes read shared cardiac physiology rather than waveform morphology.","core_discovery":"On the paper's own terms, the discovery is that a single shared Transformer pretrained with intra-modal masked latent prediction followed by delay-aware cross-modal prediction yields a cardiac code that is largely invariant to sensing modality while remaining discriminative for downstream tasks. After Stage I, codes cluster by sensor; after Stage II the modality silhouette falls from 0.121 to -0.006, and downstream performance improves. The same frozen encoder attains an average PPG classification AUROC of 80.4, an average ECG AUROC of 84.1 across 18 settings, and 97.9 and 66.8 AUROC on CirCor murmur and CinC2016 abnormal heart sound detection. The authors read this as evidence that the model has aligned ECG, PPG, and PCG in latent cardiac time rather than by raw timestamp.","pith_inferences":["Beyond the paper's claims: if the shared-latent view is correct, the delay estimates themselves become a free physiological measurement, so pulse transit time and electromechanical delay could be read directly from a single ECG-PPG or ECG-PCG pair.","The same two-stage recipe could transfer to other synchronized physiological families, such as impedance cardiography, ballistocardiography, or fetal monitoring, whenever a beat-level anchor can be detected.","A practical extension would align a new sensor's tokens to the existing shared code, letting a rare modality inherit the representation of abundant ones without a large paired corpus."],"forward_implications":["A single frozen encoder can serve ECG, PPG, and PCG tasks, so one cardiac foundation model can replace three per-sensor models.","Signal-only cardiac pretraining can approach or match models trained with clinical text or supervised labels on several ECG benchmarks, lowering the need for privileged annotations.","The learned delay aligner tracks physiological timing, so per-token delay estimates are interpretable cardiac timing measurements rather than opaque parameters.","The two-stage curriculum matches the practical data regime: scarce synchronized multi-sensor recordings are reserved for alignment while abundant unimodal corpora carry most of the representation learning."],"supporting_citations":[{"why":"Supplies the joint-embedding predictive architecture that the Stage I and Stage II losses are built on.","marker":"[4]"},{"why":"Provides the 800K-record MIMIC-IV-ECG corpus used for Stage I ECG pretraining.","marker":"[13]"},{"why":"Provides the large PPG-EXT corpus used for unimodal PPG pretraining and for synchronous ECG-PPG pairs.","marker":"[29]"},{"why":"Provides the VitalDB synchronous ECG-PPG recordings used for delay-aware cross-modal alignment.","marker":"[21]"},{"why":"Provides the EPHNOGRAM synchronous ECG-PCG recordings used to align electrical and acoustic modalities.","marker":"[18]"},{"why":"Provides the SensSmartTech trimodal ECG-PPG-PCG recordings used for cross-modal alignment in Stage II.","marker":"[20]"},{"why":"Provides the BMD-HS heart sound corpus used for Stage I PCG pretraining.","marker":"[1]"},{"why":"Provides PTB-XL, the main ECG downstream benchmark with four classification subgroups.","marker":"[40]"},{"why":"Provides the CirCor DigiScope murmur detection benchmark that reports the largest PCG gain.","marker":"[32]"},{"why":"Provides the PulseLM PPG benchmark data covering the seventeen PPG downstream tasks.","marker":"[34]"}],"fun_headline_variants":["Shared heart-signal model tops per-sensor AUROC by up to 18.8","One encoder for ECG, PPG, PCG: beats sensors by 18.8 AUROC","CardioState-JEPA aligns ECG/PPG/PCG, lifts AUROC up to 18.8","Delay-aware model learns shared cardiac code, outshines single-sensor","Cross-modal cardiac AI: shared encoder gains 18.8 AUROC on murmur"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ECG, PPG, and PCG are time-delayed renderings of one shared latent cardiac state that a scalar per-token delay can align; if no such common state is recoverable after tokenization, the cross-modal gains would reflect matched sensor appearance rather than shared cardiac physiology.","fun_headline_variants_meta":{"raw":{"variants":["Shared heart-signal model tops per-sensor AUROC by up to 18.8","One encoder for ECG, PPG, PCG: beats sensors by 18.8 AUROC","CardioState-JEPA aligns ECG/PPG/PCG, lifts AUROC up to 18.8","Delay-aware model learns shared cardiac code, outshines single-sensor","Cross-modal cardiac AI: shared encoder gains 18.8 AUROC on murmur"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1574,"prompt_tokens":1007,"completion_tokens":567,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":451}},"tokens_in":623,"tokens_out":567,"duration_ms":4869,"temperature":1.0,"reasoning_tokens":451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:03:04.604241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain Stage II on paired recordings whose cross-modal pairs have been shuffled so the target signal comes from a different patient while delay supervision still uses the source-only anchor; if the reported downstream gains persist, the cross-modal objective is not learning shared physiology, and if they collapse, the shared-state reading is supported.","supporting_citations":[{"cited_title":"Aston, Pe- ter H","cited_arxiv_id":null,"evidence_quote":"Provides the large PPG-EXT corpus used for unimodal PPG pretraining and for synchronous ECG-PPG pairs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SensSmartTech trimodal ECG-PPG-PCG recordings used for cross-modal alignment in Stage II."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BMD-HS heart sound corpus used for Stage I PCG pretraining."}],"review_version":1}