{"id":"514836a1-3571-466d-be83-a941604ec767","arxiv_id":"2506.08357","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A unified model with AdaIN style injection and a Swin/U-Net backbone converts ECG, PPG, and ABP waveforms in all six directions, beating single-direction baselines on public data.","lead":"MD-ViSCo is a single neural model that converts between ECG, PPG, and arterial blood pressure waveforms in any direction, replacing the usual practice of training a separate model for each source-to-target pair. Across two public datasets it reports lower error and higher correlation than the previous single-direction baselines, and its blood pressure outputs meet AAMI and BHS standards.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AAMI/BHS clinical compliance is measured under a calibration-based split and reduces to BP-regressor accuracy, so the headline clinical claim is not established for calibration-free use.","rationale":"I read the central claim as having two parts: (1) a single approximation model can handle all six waveform-conversion directions with fidelity at least comparable to per-direction baselines, and (2) the generated ABP waveforms meet AAMI/BHS clinical compliance. The evidence for part (1) is reasonably strong: patient-level splits for the approximation stage, five-seed experiments, ablations, and public datasets. Part (2) is the load-bearing weak point. The reader's weakest assumption exactly identifies it: the refinement stage's calibration-based split and the linear SBP/DBP transformation mean the clinical numbers measure a per-patient-calibrated BP regressor, not the unified model's calibration-free waveform generation. I do not see an internally inconsistent derivation in the waveform-conversion setup itself; the concern is about what the AAMI/BHS results can honestly claim. Therefore I agree with the reader's CONDITIONAL verdict and recommend no change: the paper should be published only if the clinical claims are disclosed as calibration-dependent, or re-evaluated calibration-free. I did not identify a separate fatal flaw in the core multi-directional conversion result.","tokens_in":19914,"tokens_out":4677,"duration_ms":56364,"concrete_test":"Recompute Table 3 and Figure 5 on the untouched calibration-free test set—the original patient-level held-out patients, not the re-split used for refinement finetuning—using the refinement model trained only on the calibration-free training split or frozen after pretraining. If AAMI ME/SD or BHS grades fall below the reported Pass/Grade B, the abstract's clinical-compliance claim is unsupported as stated; if they hold, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central clinical claim—generated ABP waveforms satisfying AAMI and BHS—rests on two premises not met by the reported evaluation. First, Section 3.5.1 states that the original calibration-free test set is further split into train/validation/test for refinement finetuning, 'where the same patient may appear in both the training and test sets' (Table 2). The AAMI results in Table 3 and BHS results in Figure 5 therefore reflect a calibration-based setting, not the calibration-free setting the abstract advertises. Second, Section 3.4 defines the final ABP waveform via a linear transformation using predicted SBP and DBP: y_ABP = y_ABP_apx * (SBP_pred − DBP_pred) + DBP_pred. As a result, the generated waveform's SBP and DBP exactly equal the refinement model's predictions, so AAMI ME/SD and BHS MAE percentages evaluate the BP regressor, not the morphology produced by the approximation stage. The conclusion even concedes that the refinement step 'may involve task- or dataset-specific finetuning.' Hence the clinical-compliance claims are only as strong as a calibration-free BP regressor, and they do not validate waveform morphology under AAMI/BHS.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MD-ViSCo, a two-stage framework for multi-directional vital sign waveform conversion among ECG, PPG, and ABP. The approximation stage uses a 1D U-Net with a Swin Transformer and AdaIN to generate a locally normalized target waveform from any single source waveform and a target domain indicator. The refinement stage predicts SBP and DBP from the source waveform and optional patient demographics, then rescales the normalized ABP waveform to mmHg via a linear transformation. The authors evaluate on PulseDB and UCI against NabNet, PPG2ABP, PatchTST, and P2E-WGAN, reporting MAE/PC, physiological feature fidelity, and AAMI/BHS compliance, and claim that one unified model matches or outperforms six separate uni-directional baselines while satisfying AAMI and BHS Grade B for ABP generation.","tokens_in":20118,"tokens_out":4978,"duration_ms":60253,"significance":"If the waveform-conversion results hold, the unified framework is a practically useful contribution: it removes the need to train and maintain six separate source-to-target models and is backed by five-seed experiments, a patient-level calibration-free split for the approximation stage, multiple evaluation metrics, ablations, and public code. The physiological feature fidelity results and multi-directional-versus-uni-directional ablation are informative. The clinical-compliance claim, however, is currently over-stated because the AAMI/BHS evaluation is performed under a calibration-based split and, through the linear transformation, reduces to BP-regressor accuracy rather than waveform morphology. The core waveform-generation contribution is sound enough to warrant a major revision rather than rejection, provided the clinical claims are re-scoped or re-evaluated.","major_comments":[{"comment":"The AAMI/BHS evaluations are reported on a calibration-based refinement split: the original patient-disjoint test set is further split into finetune-train, finetune-val, and finetune-test, and the text explicitly states that \"the same patient may appear in both the training and test sets.\" Section 4.3 and the abstract nevertheless present the AAMI pass and BHS Grade B without this caveat. Because AAMI/BHS are device-approval standards that assume evaluation on subjects not used for calibration, the headline clinical-compliance claim is not established for the calibration-free setting that the paper otherwise emphasizes. The conclusion even concedes that the refinement step \"may involve task- or dataset-specific finetuning.\" Please either report AAMI/BHS on the original patient-disjoint test set with the refinement model frozen after training on the train split, or clearly qualify all clinical-compliance claims as calibration-based in the abstract and Section 4.3.","section":"Section 3.5.1, Table 2"},{"comment":"The linear transformation y_ABP = y_ABP_apx · (SBP_pred − DBP_pred) + DBP_pred, combined with the local min-max normalization of y_ABP_apx, makes the generated waveform's systolic and diastolic values identically equal to the refinement model's outputs SBP_pred and DBP_pred. Consequently, the AAMI/BHS metrics in Table 3 and Figure 5 evaluate the accuracy of the BP regressor, not the morphology generated by the approximation model. Morphology is assessed separately by MAE/PC and feature fidelity, but the text in Section 4.3 and the abstract overstates what the clinical-compliance experiments test. Please rephrase these claims to state that the AAMI/BHS evaluation validates the amplitude (BP) estimates, and clarify that the waveform shape itself is not evaluated by these standards.","section":"Section 3.4, Eq. (5)"}],"minor_comments":[{"comment":"The notation for the Swin Transformer encoder is inconsistent: z = SwinT_E(h) appears after B(h, s) = \\tilde{h} is defined, so the relationship between the bottleneck B and SwinT_E is unclear; please align the names and make the flow x → h → \\tilde{h} → \\hat{y} unambiguous.","section":"Section 3.3, Eq. (1)"},{"comment":"Please report the number of unique subjects used for the AAMI evaluation, since the AAMI standard requires at least 85 subjects and the table currently lists only ME and SD.","section":"Table 3"},{"comment":"The y-axis scales differ across panels because ECG/PPG are in normalized units while ABP is in mmHg; please add explicit annotations or a supplementary common-scale figure to prevent misleading cross-panel comparisons.","section":"Figure 3"},{"comment":"The global normalization bounds for PulseDB (2.34–286.58 mmHg) appear extreme; please clarify whether these are raw min/max values and whether outlier removal was applied before computing the bounds.","section":"Appendix 6.1.1"},{"comment":"The WCL similarity threshold symbols (\\tau_s, T_s) are used in Eq. (6) and Eq. (7) before being defined; please define all hyperparameters in the main text or refer explicitly to Table 8 when the symbols first appear.","section":"Appendix 6.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's code availability, five-seed protocol, and patient-level split for the approximation stage are strong points. The main risk is that the clinical-compliance headline is tied to the calibration-based refinement split and to BP-regressor accuracy via the linear transformation. If the authors can re-run the AAMI/BHS evaluation on the original patient-disjoint test set or substantially qualify the claims, the paper could become acceptable; I would ask for that revision before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is real and worth your time: a single model that converts between any two of ECG, PPG, and ABP, all six directions. The architecture (AdaIN style conditioning inside a 1-D U-Net with a Swin Transformer) is a sensible new combination, and their table correctly shows no prior work covers all three as both source and target. The evidence for the central claim is solid: five-seed runs, patient-level splits for the approximation stage, multiple metrics, ablations, and code on GitHub. Multi-directional training consistently beats separate per-direction models. That part deserves credit.\n\nThe soft spots are exactly where the stress-test note lands. The AAMI/BHS results come from a calibration-based split, not the calibration-free split the abstract implies. Section 3.5.1 says the original test set is further split for finetuning, with the same patient possibly in train and test. Table 2 makes this clear, but the abstract and Section 4.3 present the results without that qualifier. Second, the linear transformation in Section 3.4 means the generated ABP waveform's SBP and DBP always equal the refinement model's predicted values. So the AAMI/BHS numbers are really grading the BP regressor, not the waveform morphology. The paper is honest about the finetuning in the conclusion, but the headline oversells clinical readiness.\n\nThis is a strong systems paper for the physiological-signal-synthesis community, and the unified conversion result is a genuine consolidation. The clinical compliance claims need a major revision: either re-evaluate on the calibration-free split or clearly state the calibration-based condition up front, and explicitly note that the AAMI/BHS scores reflect the regression module's accuracy, not the shape fidelity of the generated waveform. Reproducibility would improve with a commit hash and one-command eval, but that is minor.\n\nI would send this to peer review, not desk reject it. The central contribution is novel and empirically supported; the clinical claims need tightening, not scrapping. A good referee will help the authors frame the result honestly.\n\nFor what it's worth, I'd bring this to reading group and cite it in my own work on waveform synthesis.","headline":"The unified multi-directional conversion claim holds up, but the AAMI/BHS clinical compliance headline is not established for calibration-free use—it rests on a calibration-based split and reduces to the BP regressor's accuracy.","tokens_in":20685,"tokens_out":2584,"would_cite":true,"duration_ms":31223,"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 that one unified model can generate any target waveform among ECG, PPG, and ABP from any single source waveform, with lower average error than six task-specific baselines and AAMI/BHS-compliant ABP output.","keywords":["vital sign waveform conversion","unified model","ECG","PPG","ABP","AdaIN","Swin Transformer","AAMI/BHS standards"],"falsifier":"On a strictly patient-disjoint test set with no calibration finetuning, compute BHS grades from the systolic and diastolic values detected directly on the generated ABP waveform rather than from the regressor's predictions, and check whether the pass rates still meet the AAMI/BHS thresholds; if they drop, the compliance claim is confined to the calibrated setting.","tokens_in":19653,"feed_emoji":"🩺","tokens_out":6248,"duration_ms":74626,"temperature":0.7,"pith_summary":"The paper claims that a single model can replace the usual collection of six separate source-to-target waveform converters: from ECG, PPG, or ABP as input, it can generate any of the other waveform types. The payoff is practical: one architecture, one training run, and one deployed model instead of several task-specific pipelines, which matters in hospital settings where the available sensor can change. On two public ICU datasets, the model reports an 8.8% lower mean absolute error and a 4.9% higher Pearson correlation averaged over all conversion directions and single-direction baselines. For generated ABP waveforms, the paper reports compliance with the AAMI standard and Grade B on the BHS standard, which are clinical thresholds used for blood-pressure measurement devices.","feed_headline":"A single model generates any vital-sign waveform from any other","feed_subtitle":"MD-ViSCo reports 8.8% lower mean absolute error than six single-direction baselines and AAMI-compliant ABP waves.","key_machinery":"The workhorse is Adaptive Instance Normalization (AdaIN) injected through a shallow 1-D U-Net with a Swin Transformer bottleneck. AdaIN re-centers and re-scales the feature map's mean and standard deviation using a style vector $\\gamma, \\beta$ learned from the one-hot target choice, so the same weights generate every target modality while the Swin Transformer supplies global temporal context. The refinement stage uses a PatchTSMixer encoder for the source waveform, a DistilBERT text encoder for patient demographics, a weighted contrastive loss that aligns clinically similar samples, and a linear transformation $\\hat{y}_{ABP} = \\hat{y}_{apx} \\cdot (\\hat{y}_{SBP} - \\hat{y}_{DBP}) + \\hat{y}_{DBP}$ that rescales the normalized waveform into mmHg.","core_discovery":"The central discovery is that multi-directional waveform conversion can be treated as a style-transfer problem. All waveforms are locally min-max normalized so that amplitude scales match, all modalities are pooled into one training set, and a one-hot target code is mapped to a style embedding that Adaptive Instance Normalization uses to shift and scale internal feature statistics. The resulting single model generates normalized waveforms in any target modality. A refinement stage then converts the normalized ABP waveform into millimeters of mercury by multiplying by the predicted systolic-to-diastolic range and adding the predicted diastolic pressure, using patient demographics encoded as text. On the tested datasets, this one-model design matches or beats per-direction baselines, and the authors report that the generated ABP waveforms satisfy AAMI and BHS Grade B.","pith_inferences":["The paper leaves implicit that its AAMI/BHS compliance scores are, in effect, scores for the blood-pressure regressor followed by a linear rescale, not for the generated waveform's own morphology; a stricter reading would grade systolic and diastolic values detected directly on the raw generated ABP waveform.","An untested extension is that the same style-transfer design could absorb additional waveform types, such as respiration, as new target domains by enlarging the one-hot code and reusing the normalization and AdaIN machinery.","Because inference requires only one input waveform, the model is a natural fit for wearable or ward monitoring where sensor availability varies; the paper does not quantify performance when the single input waveform is noisy or missing, which a deployment study could add."],"forward_implications":["Only one trained model needs to be deployed for conversions among ECG, PPG, and ABP, so adding a conversion direction costs no new architecture or separate maintenance pipeline.","Multi-directional training improves even per-direction error relative to uni-directional training, with reported average MAE reductions of 5.4% for ECG generation, 3.1% for PPG generation, and 4.0% for ABP generation on PulseDB.","If the AAMI/BHS results hold, the model supports non-invasive-to-invasive ABP monitoring in wards or pre-hospital settings where only PPG or ECG is available.","Encoding patient information as text lets the refinement stage accept different demographic schemas without manual feature engineering, easing portability across datasets with different patient tables."],"supporting_citations":[{"why":"Baseline PPG-to-ABP model that defines the comparison for one conversion direction.","marker":"[16]"},{"why":"Baseline two-stage ABP model and source of the linear transformation that rescales normalized ABP to mmHg using predicted SBP/DBP.","marker":"[17]"},{"why":"Supplies one benchmark dataset with synchronized ECG/PPG/ABP segments and patient demographic tables.","marker":"[15]"},{"why":"Supplies the second benchmark dataset, without demographic information, used to test portability.","marker":"[14]"},{"why":"Style-transfer framework whose domain-conditioned AdaIN design the approximation model adapts.","marker":"[30]"},{"why":"Swin-Unet architecture that the approximation model's transformer U-Net follows.","marker":"[43]"},{"why":"PatchTSMixer backbone used by the refinement model's waveform encoders.","marker":"[35]"},{"why":"DistilBERT text encoder used to embed patient demographics for the refinement model.","marker":"[44]"},{"why":"Provides the weighted contrastive learning objective used to align waveform and patient embeddings.","marker":"[34]"}],"fun_headline_variants":["One model converts any vital sign to any other","MD-ViSCo: single model for all vital-sign waveform swaps","Any-to-any vital sign waveforms with one unified model","Unified vital-sign converter cuts error by 8.8% over baselines","Single model generates ECG, PPG, ABP from any input"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the AAMI/BHS compliance scores, which come from a calibration finetuning split where the same patient can appear in both tuning and test sets and where the waveform's systolic and diastolic values are set by the regression model, should be taken as evidence about the generated waveform itself.","fun_headline_variants_meta":{"raw":{"variants":["One model converts any vital sign to any other","MD-ViSCo: single model for all vital-sign waveform swaps","Any-to-any vital sign waveforms with one unified model","Unified vital-sign converter cuts error by 8.8% over baselines","Single model generates ECG, PPG, ABP from any input"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1646,"prompt_tokens":1007,"completion_tokens":639,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":551}},"tokens_in":623,"tokens_out":639,"duration_ms":6707,"temperature":1.0,"reasoning_tokens":551,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:12:59.407023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a strictly patient-disjoint test set with no calibration finetuning, compute BHS grades from the systolic and diastolic values detected directly on the generated ABP waveform rather than from the regressor's predictions, and check whether the pass rates still meet the AAMI/BHS thresholds; if they drop, the compliance claim is confined to the calibrated setting.","supporting_citations":[{"cited_title":"PPG2ABP: translating photoplethysmo- gram (PPG) signals to arterial blood pres- sure (ABP) waveforms,","cited_arxiv_id":null,"evidence_quote":"Baseline PPG-to-ABP model that defines the comparison for one conversion direction."},{"cited_title":"PulseDB: A large, cleaned dataset based on MIMIC- III and VitalDB for benchmarking cuff-less blood pressure estimation methods,","cited_arxiv_id":null,"evidence_quote":"Supplies one benchmark dataset with synchronized ECG/PPG/ABP segments and patient demographic tables."},{"cited_title":"DistilBERT, a distilled ver- sion of BERT: smaller, faster, cheaper and lighter,","cited_arxiv_id":null,"evidence_quote":"DistilBERT text encoder used to embed patient demographics for the refinement model."}],"review_version":1}