{"id":"d9b9d152-d168-4606-9d4a-125eda300fbd","arxiv_id":"2504.21209","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GenClean uses a variational autoencoder with frequency and normalisation adapters to clean artefacts in medical pulse waveforms without manual labels, claiming robust cross-patient and cross-disease generalisation.","lead":"This paper introduces GenClean, a label-free framework that detects and removes artefacts from medical pulse waveforms using a variational autoencoder trained on 180,000 unlabelled blood-pressure samples. It reports strong generalisation across patients, hospitals, diseases, and signal types, plus a real-time plugin for the ICM+ monitoring software.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed accuracy/F1 depends on artefact labels whose provenance and validation are not described; without an annotation protocol, the central generalisation numbers cannot be interpreted.","rationale":"The reader's weakest_assumption is exactly the load-bearing point, so I agree. I considered other candidates: the five-patient MIMIC cohort and absence of error bars weaken the cross-disease generalisation claim but do not invalidate the internal patient-level result; the discrepancy between Fig. 3b (~19 ms) and Table 2 (39 ms) is a minor reporting issue; the real-time ICM+ plugin is an engineering demonstration with measured overhead and does not depend on test labels. The label question, by contrast, sits under every reported performance number in the abstract and results. If test labels are not reliable, each of accuracy, F1, sensitivity, specificity, and the hypertension-event reduction is uninterpretable. The paper's own Methods text offers only 'experts-labelled data' without a protocol, and the heuristic filtering of training data makes it plausible that the test labels could be generated by the same simple rules, which would make the evaluation circular. The proposed check — blinded re-annotation with agreement statistics and recomputation on consensus labels — directly settles whether the concern lands. Since the reader already conditions acceptance on addressing this issue, my recommendation does not change the verdict category; I would keep it conditional, with label provenance and an independent-annotation check as explicit acceptance conditions. If those checks failed, the paper would need substantive revision or withdrawal of the headline accuracy/F1 claims.","tokens_in":12087,"tokens_out":4686,"duration_ms":47849,"concrete_test":"Ask the authors to release the annotation protocol and a random sample of raw waveform segments with labels, and run a blinded re-annotation study: take roughly 300 internal and 50 MIMIC-III 10-second segments (stratified by predicted clean/artefactual), have two clinicians with stated artefact definitions label them independently, and compute Cohen's kappa. Then recompute accuracy, sensitivity, specificity, and F1 using consensus labels, with disagreements adjudicated. In parallel, check whether the released heuristic rules (ABP range 0–300 mmHg, peak-to-peak >15 mmHg, PPG 0.5–3 Hz bandpower threshold) reproduce the original labels; if heuristic-label agreement is high, the evaluation is likely capturing heuristic filtering rather than expert artefact judgement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim — at least 90% accuracy and 0.89 F1 on the internal held-out set and 95.6% accuracy on MIMIC-III — is only meaningful if the artefact/non-artefact labels used for evaluation are reliable and independent of the training pipeline. The paper never supplies this. In Results/Clinical Datasets, the held-out set is described only as 'balance labelled (non-artefactual: artefactual = 1:1)'; in Methods/Artefact Cleaning Framework, evaluation is said to use 'the artefactual sample and our experts-labelled data' with no protocol, no label definitions, no inter-rater agreement, and no indication whether labels were produced per patient, per segment, or per waveform region. MIMIC-III labels are not described at all. This matters doubly because the training/validation pipeline is not label-free in a neutral sense: Methods/Heuristic rules explicitly filters training data using domain thresholds (ABP 0–300 mmHg, peak-to-peak >15 mmHg, PPG bandpower outside 0.5–3 Hz), and Methods/Artefact Cleaning Framework sets the artefact threshold at the 90th percentile of validation-set reconstruction error. If the held-out labels were derived from the same heuristics or from a single non-blinded annotator, the accuracy and F1 numbers measure consistency with those heuristics rather than clinically meaningful artefact detection. The small cross-disease cohort (five MIMIC-III patients, no error bars) compounds this: 95.6% over five patients is a single anecdotal estimate. The code release supports training reproduction but does not include label generation for the test sets, so the numbers cannot be independently checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GenClean, a label-free variational-autoencoder-based artefact cleaning framework for pulsatile medical time series. The method is trained on 180,000 ten-second arterial blood pressure (ABP) segments from 150 traumatic brain injury patients and is claimed to generalize across patient-level distribution shifts, across disease cohorts (MIMIC-III), and to other signal modalities (PPG). The authors report at least 90% accuracy and 0.89 F1-score on an internal held-out set, 95.6% accuracy on MIMIC-III ABP, 85.8% accuracy on WESAD PPG, and a 22% reduction in detected hypertension events after cleaning. They also demonstrate integration into the ICM+ monitoring software with modest CPU and memory overhead and provide code.","tokens_in":12426,"tokens_out":5043,"duration_ms":53088,"significance":"If the reported results hold, GenClean would be a valuable contribution: it is among the first label-free artefact-cleaning methods to explicitly target patient- and site-level generalisation, it uses a large in-house training corpus, the code is made available, and the real-time ICM+ integration addresses a practically important deployment constraint. The honest limitations section and the explicit investigation of distribution shifts are also strengths. However, the central evaluation numbers depend on artefact labels whose generation is not described, so the current evidence is not sufficient to certify the generalisation claims.","major_comments":[{"comment":"The evaluation labels that support the central claims are never described. The internal held-out set is only said to be 'balance labelled (non-artefactual: artefactual = 1:1)', and the evaluation is said to use 'our experts-labelled data', but the paper gives no annotation protocol, no definition of what constitutes an artefact at the segment or waveform level, no inter-rater agreement, and no information about annotator blinding. For the MIMIC-III cross-disease cohort, the label generation is not described at all. Since accuracy, F1, sensitivity, and specificity are all computed against these labels, the headline numbers in Fig. 2d and Table 1 cannot be interpreted without knowing whether the labels are reliable and independent of the heuristic rules used in training.","section":"Methods, Artefact Cleaning Framework; Results, Clinical Datasets"},{"comment":"The cross-disease generalisation claim rests on only five MIMIC-III patients and a single pooled accuracy of 95.6%, reported without the number of test segments, a per-patient breakdown, confidence intervals, or significance tests. The hypertension-event reduction is a raw change from 2134 to 1661 events over the same five patients. This sample size is too small to support the broad cross-disease and cross-site generalisation statement made in the abstract and discussion; the authors should report per-patient metrics, confidence intervals, and the total number of annotated segments.","section":"Results, Generalisation on Cross-cohort; Table 1"},{"comment":"For PPG, the Methods define artefact classification by a band-power heuristic (segments with more than 30% of total signal power outside 0.5–3 Hz are classified as artefacts). If the WESAD test labels are generated with this same criterion, the reported 85.8% accuracy measures agreement with that heuristic rather than generalisation to independently established PPG artefacts. The paper should state explicitly how the WESAD test labels were obtained, whether they are independent of the training heuristics, and whether the same heuristic was used for both DaLiA training filtering and WESAD evaluation.","section":"Methods, Heuristic rules; Results, WESAD (PPG)"},{"comment":"The artefact threshold is set as the 90th percentile of reconstruction error on the validation set, and the loss is computed before the rescaling step of the adaptive normalisation layer. These choices are reasonable in principle, but the paper reports no sensitivity analysis with respect to the threshold percentile or the validation-set composition. Without such an analysis, it is unclear whether the claimed accuracy and F1 are robust to this hand-set threshold or whether they would change materially under plausible variations in validation-set artefact prevalence.","section":"Methods, Artefact Cleaning Framework; Methods, Training methods"}],"minor_comments":[{"comment":"The sentence 'One of the patients (Test 5) consistently exhibited lower performance in both metrics, with no value above 0.6' is ambiguous: it is not clear whether this refers to the previous method or to GenClean. If it refers to GenClean, it appears to contradict the earlier claim of at least 90% accuracy for the generalised model; please clarify.","section":"Results, Distribution Shift Exploration"},{"comment":"The processing time is reported as '~19 ms for 120 Hz' and 'longest processing time reaching ~26 ms (150 Hz)' in the results, while Table 2 lists 'Processing Time 39 ms' for a ten-second segment. Please reconcile these numbers and specify whether they measure the same workload.","section":"Real-time results and Table 2"},{"comment":"In the ELBO equation, the text says 'p(x|z) is the encoder, q(z|x) is the decoder', but in the standard VAE notation used in the equation q(z|x) is the encoder and p(x|z) is the decoder. Please correct this description.","section":"Methods, Artefact Cleaning Framework"},{"comment":"There are several typographical and formatting issues: 'trauma brain injury' should be 'traumatic brain injury', 'V AE' has inconsistent spacing, the WESAD description contains 'neural' instead of 'neutral', and some references are incomplete (e.g., Ref. 28 reports 'Sensors 23, nan'). Please proofread the manuscript.","section":"General editorial"},{"comment":"The caption and text state that both MSE and MAE were compared as reconstruction error metrics, but only the MSE result is described in detail. Please report the MAE comparison or state explicitly that it was used only as a check.","section":"Figure 3c"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-label provenance is the make-or-break issue for this manuscript; I would ask the authors to provide a detailed annotation protocol, inter-rater measures, and explicit statements about label independence from the heuristic rules. It would also strengthen the paper to expand the MIMIC-III validation beyond five patients or at least provide per-patient intervals, and to check the 'first artefact-cleaning method' novelty claim against the existing ICU artefact-cleaning literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know this paper is a credible extension of the DeepClean line, not a breakthrough. It adds a frequency adapter and reversible instance normalisation to a VAE artefact cleaner, trains on 150 TBI patients, and shows cross-patient, cross-disease, cross-signal results, plus a real-time plugin in ICM+. The engineering is neat and the code is on GitHub.\n\nWhat's actually good: the scale of training data (180k segments) is a real step up from earlier single-patient work; the generalisation matrices for the previous method vs their generalised version make the patient-shift problem concrete; and the real-time metrics (39 ms per segment, +6% CPU) are exactly the practical numbers a clinician needs to judge deployability. The paper also owns its limitations (paediatrics, arrhythmia OOD) in plain language.\n\nNow the soft spots, in proportion. The stress-test note is right: the artefact labels that everything rests on are never described. For the internal test set we get 'balanced labelled' and 'experts-labelled data' with no protocol, no inter-rater check, no segment-vs-beat definition. For MIMIC-III, five patients, and no label description at all. This is not circularity — the threshold is set on the validation set, not on test labels — but it means the 95.6% accuracy could be measuring agreement with a single annotator or even the same heuristics used during training filtering. That is a documentation gap, not evidence of fraud, but it undercuts the headline numbers.\n\nSecond, five MIMIC-III patients is a single anecdotal estimate. No confidence intervals, no patient-level error bars, and the 22% hypertension-event reduction is reported as a raw percentage. For a paper claiming generalisation, that is a weak load-bearing pillar. The PPG transfer (85.8% on WESAD) is nice, but the label source for WESAD is again unspecified.\n\nSo: the architecture and the real-time story deserve a serious referee. The evaluation protocol needs rigorous revision, not a desk reject. If the authors document the labelling protocol, expand the MIMIC-III cohort, and report per-patient confidence intervals, this becomes a solid contribution. As it stands, I would treat the strong generalisation claims as promising but not yet supported.\n\nRecommendation: send to peer review with a major-revision request focused on evaluation details. It will be cited as evidence that label-free cleaning works across sites, so it should be held to a high standard.","headline":"A useful, well-scoped engineering extension of DeepClean with a real-time deployment story, but the evaluation is under-specified (labels, cohort size, error bars) and needs major revision before the generalization claims carry their weight.","tokens_in":12990,"tokens_out":4586,"would_cite":true,"duration_ms":45161,"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 label-free variational autoencoder trained on 180,000 ten-second blood-pressure samples cleans artefacts across patients, cohorts, and signal types, and runs in real time.","keywords":["artefact cleaning","medical pulsatile time series","label-free learning","variational autoencoder","distribution shift","cross-cohort generalisation","real-time monitoring","arterial blood pressure"],"falsifier":"Collect a sample of the held-out arterial-blood-pressure segments and have two independent clinicians annotate each as artefactual or clean; if inter-rater agreement is low or the model's errors concentrate on segments clinicians disagree about, then the reported 90% accuracy and 0.89 F1-score are not anchored to a reliable ground truth. Alternatively, feed the model a set of clean but atypical rhythm segments (e.g., from patients with atrial fibrillation): if it flags most of them as artefacts, the claimed generalisation to unseen physiology fails.","tokens_in":11929,"feed_emoji":"🩺","tokens_out":10255,"duration_ms":95019,"temperature":0.7,"pith_summary":"The paper claims that its label-free framework, GenClean, can handle the patient-level distribution shifts that typically break supervised cleaners. Trained on 180,000 ten-second arterial-blood-pressure samples from 150 trauma-brain-injury patients, with no artefact labels, the model flags any segment whose variational-autoencoder reconstruction error exceeds a validation-set threshold. The authors report at least 90% accuracy and 0.89 F1-score on held-out patients from the same centre, 95.6% accuracy on an unrelated intensive-care-unit cohort, and 85.8% accuracy when the same recipe is moved to photoplethysmography. They also demonstrate real-time operation inside a bedside monitoring system, at about 19–39 ms per ten-second segment.","feed_headline":"Label-free model cleans blood-pressure traces across patient groups","feed_subtitle":"Trained on 180,000 segments, it hits 95.6% accuracy on new ICU data and runs on a bedside CPU.","key_machinery":"The load-bearing object is a one-dimensional convolutional variational autoencoder trained with the standard variational-autoencoder objective (the evidence lower bound), with the artefact decision made from mean-squared reconstruction error. Two modules wrap the autoencoder: a frequency adapter that resamples any input rate to the training rate and back, and a reversible instance-normalisation layer that standardises each segment's mean and variance before encoding and restores them after decoding. The reconstruction error is thresholded at the 90th percentile of the validation set, and segments above the threshold are flagged as artefacts. A latent dimension of roughly 20 is reported as the best balance between reconstruction fidelity and class separation. This machinery learns a probabilistic prior over clean pulsatile waveforms and classifies by deviation from that prior, requiring no labels.","core_discovery":"GenClean's central discovery is that a variational autoencoder trained only to reconstruct clean pulsatile waveforms doubles as a generalised artefact detector: a segment is labelled artefactual when its reconstruction deviates from the input beyond a threshold, and this rule transfers to new patients, new recording sites, different sampling rates, and even a different sensor modality. The paper demonstrates this by training on one hospital's trauma cohort and testing on held-out patients and on an external intensive-care arterial-blood-pressure cohort, where it reports 95.6% accuracy, and by adapting the same training to a wrist-worn photoplethysmography dataset, where it reports 85.8% accuracy. The authors argue that the generalisation comes from three design choices working together: resampling inputs to a common frequency, reversible instance normalisation that strips patient-specific mean and variance before reconstruction, and a loss computed before rescaling so one global threshold works across patients.","pith_inferences":["Inference: the 90th-percentile validation threshold ties the operating point to the artefact prevalence of the validation set; deploying in a unit with a different artefact rate would require recalibrating the threshold, which the paper does not address.","Inference: because the variational autoencoder reconstructs the most probable clean waveform, rare but physiologically normal patterns such as arrhythmias may be flagged as artefacts; a targeted test on arrhythmia segments would reveal whether this failure mode is real.","Inference: the reported 22% reduction in hypertensive events should not be read as pure error correction; without an independent human-annotated reference, some removed 'events' could be genuine but atypical physiology, so a gold-standard clinical validation is a natural next step.","Inference: the same label-free recipe should transfer to other cardiac-driven pulsatile signals such as intracranial pressure or fetal heart rate, since the only requirement is a learnable waveform prior for clean segments; the paper suggests but does not test this."],"forward_implications":["A model trained on one centre's trauma cohort can be applied directly to arterial blood pressure from a different intensive-care population recorded at a different sampling rate, with reported accuracy above 95% and no retraining.","Removing artefact-contaminated segments cuts the number of counted hypertensive events by about 22% on the external cohort, suggesting that alarms derived from raw waveforms could be reduced substantially.","The same variational-autoencoder recipe transfers to photoplethysmography with accuracy above 85%, so the method is not tied to invasive blood-pressure morphology.","With roughly 19–39 ms of processing per ten-second segment and only a few mFLOPS, the method runs on bedside CPU hardware rather than requiring a GPU server.","The latent space becomes visibly more separated between artefactual and clean segments, so the representation itself could support downstream monitoring tasks beyond cleaning."],"supporting_citations":[{"why":"supplies the earlier label-free variational-autoencoder artefact-rejection method that GenClean builds on and compares against","marker":"[8]"},{"why":"provides the training-on-one-PPG-dataset-and-testing-on-another protocol used for the photoplethysmography transfer experiment","marker":"[12]"},{"why":"the bedside monitoring software used as the host for the real-time feasibility demonstration","marker":"[17]"},{"why":"defines the pressure reactivity index, the clinically relevant metric that motivates the 10-second segment length","marker":"[18]"},{"why":"the external intensive-care waveform database used as the cross-disease, cross-site generalisation test set","marker":"[19]"},{"why":"supplies the systolic and diastolic thresholds used to count hypertensive events before and after cleaning","marker":"[22]"},{"why":"the variational autoencoder and its training objective that form the backbone of the model","marker":"[39]"},{"why":"introduces reversible instance normalisation, the technique used to absorb patient-level mean and variance shifts","marker":"[43]"}],"fun_headline_variants":["Label-free autoencoder scrubs artefact from ABP and PPG in real-time","GenClean: unsupervised artefact removal that works on new patients and sensors","Label-free VAE cleans ABP and PPG traces across cohorts","Real-time artefact filtering without labels, from ICU to wearables","One unsupervised model scrubs pulsatile signals on a bedside CPU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies assume the expert labels for the held-out and external test sets are correct and consistently defined, yet the paper never describes how those labels were generated or validated, so the numbers could be measuring agreement with unverified annotations.","fun_headline_variants_meta":{"raw":{"variants":["Label-free autoencoder scrubs artefact from ABP and PPG in real-time","GenClean: unsupervised artefact removal that works on new patients and sensors","Label-free VAE cleans ABP and PPG traces across cohorts","Real-time artefact filtering without labels, from ICU to wearables","One unsupervised model scrubs pulsatile signals on a bedside CPU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001227,"raw_usage":{"total_tokens":5027,"prompt_tokens":915,"completion_tokens":4112,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":4019}},"tokens_in":531,"tokens_out":4112,"duration_ms":30397,"temperature":1.0,"reasoning_tokens":4019,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:09:45.332751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a sample of the held-out arterial-blood-pressure segments and have two independent clinicians annotate each as artefactual or clean; if inter-rater agreement is low or the model's errors concentrate on segments clinicians disagree about, then the reported 90% accuracy and 0.89 F1-score are not anchored to a reliable ground truth. Alternatively, feed the model a set of clean but atypical rhythm segments (e.g., from patients with atrial fibrillation): if it flags most of them as artefacts, the claimed generalisation to unseen physiology fails.","supporting_citations":[],"review_version":1}