{"id":"37135553-ae24-4042-b7ce-10c749701d4d","arxiv_id":"2509.11606","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Fine-tuning Wav2Vec2 on diffusion-generated and augmented biosignals improves abnormal heart sound classification across single-channel, PCG+ECG, and multichannel PCG data, with reported state-of-the-art benchmark numbers.","lead":"This paper trains a Wav2Vec2 transformer to flag abnormal heart sounds, using diffusion models and signal augmentation to create extra training data. The method reports top results on single-channel and synchronized PCG-plus-ECG benchmarks, and competitive results on a noisy six-channel wearable vest dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Subject-level leakage through recording-level splits and an unspecified synthetic-data boundary is the load-bearing risk to the SOTA claim.","rationale":"The reader's weakest assumption is exactly the split hygiene and synthetic-data boundary, and I agree that this is the most load-bearing point. The paper's central claim is state-of-the-art performance on three benchmarks; all of those numbers depend on an evaluation protocol that prevents the model from seeing the same patients in training and test. The paper does not state that the split is patient-level, and given the multi-recording-per-patient structure of CinC, this is not a pedantic detail. Additionally, the relationship between the diffusion generator and the test set is under-specified: Section 2.1.1 says synthetic data is generated from training-a, while Section 3.2.1 describes generating synthetic PCG from Icentia ECG for single-channel models. If the generator is trained on the full training-a before the split, or if 'generated from training-a' means the test portion is used as conditioning, then the test set has indirectly influenced training. These are addressable but currently unresolved. The paper otherwise contains considerable methodological detail, reproducible training schedules, and internally consistent augmentation gains; I see no reason to suspect fabrication. The conditional verdict is appropriate: the reported numbers should not be taken at face value until the split and generation boundaries are clarified. I therefore recommend no change to the reader's CONDITIONAL verdict.","tokens_in":16771,"tokens_out":4418,"duration_ms":52233,"concrete_test":"Obtain the exact split code or subject IDs from the authors and check for patient overlap between the 60/20/20 train/validation/test partitions of CinC 2016 and training-a. Also verify that the diffusion generator for training-a was trained only on recordings from the training partition. Independently re-run the pipeline with a strict patient-level 60/20/20 split (no subject in more than one partition) and the same augmentation/synthetic protocol; if the subject-level MCC drops by more than 0.05 or accuracy by more than 2% relative to the reported numbers, the original metrics are inflated by leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the lack of a stated patient-level split for the CinC 2016 experiments. The paper says only 'A 60-20-20 split is used' (Section 2.1.1). CinC has 3153 recordings from 764 patients; if the split is at the recording level, multiple recordings from the same patient will appear in both training and test sets. Because subject-level metrics are computed by averaging fragment predictions per subject, the model can exploit patient-specific acoustics/identity rather than generalize to new patients, inflating the reported 92.48% accuracy / 0.8283 MCC and the training-a 93.14%/0.8380 results. The same section states 'Synthetic data is generated from training-a and used only in models trained on this subset,' but does not specify whether the diffusion generator's training data was restricted to the training split of training-a. If the generator saw test-subject recordings (directly or via training on the full training-a before splitting), the test set is indirectly observed during training. Section 3.2.1 also says synthetic PCG for single-channel models was generated from Icentia ECG, while Section 2.1.1 says synthetic data is generated from training-a—this ambiguity matters because the multimodal experiments may rely on synthetic data derived from training-a. Without explicit patient-disjoint splits and a clear statement that the generative model never observed test subjects, the central SOTA claim is not falsifiable from the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scalable Wav2Vec 2.0-based classifier for abnormal heart sound detection, fine-tuned on augmented and diffusion-generated synthetic biosignals. It reports state-of-the-art subject-level results on the CinC 2016 single-channel PCG dataset (accuracy 92.48%, UAR 93.05%, MCC 0.8283), on the training-a PCG+ECG subset (accuracy 93.14%, MCC 0.8380), and on a multichannel mPCG vest dataset (accuracy 77.13%, MCC 0.5082). The method combines traditional signal augmentations with WaveGrad/DiffWave synthetic signals, and is evaluated on single-channel, multimodal, and multichannel inputs. The central claim is that the proposed pipeline enables transformer-based models to achieve state-of-the-art performance on these tasks.","tokens_in":17151,"tokens_out":5110,"duration_ms":53104,"significance":"If the reported results are valid, this is a practically significant contribution: it demonstrates a scalable way to apply transformer-based raw-signal models to small, imbalanced medical audio datasets. The internal augmented-versus-original comparisons are consistent and large, which is a clear strength: the training-a multimodal baseline collapses (TNR ≈ 2%) while the augmented model reaches TNR ≈ 90%; the vest MCC rises from 0.28 to 0.51; and the CinC subject-level MCC rises from 0.70 to 0.83. The interpretability analyses are also a useful addition. However, the subject-level SOTA claims rest on an evaluation protocol whose patient-level separation and synthetic-data boundary are not fully specified. Confirming those details is essential before the headline numbers can be accepted.","major_comments":[{"comment":"The evaluation protocol does not state that the 60/20/20 split is patient-disjoint. CinC 2016 contains 3153 recordings from 764 patients, so multiple recordings per patient exist. Since subject-level results are obtained by averaging fragment predictions per subject, a recording-level split would allow the model to see fragments from the same patient in both training and test sets, inflating accuracy, UAR, and MCC. Please state explicitly that all recordings of each patient are confined to a single split, for all five shuffled splits, and confirm the same holds for the training-a subset. If this was not done, the experiments should be rerun with patient-disjoint splits.","section":"Section 2.1.1, Tables 12 and 15"},{"comment":"The boundary of the synthetic-data pipeline is underspecified. Section 2.1.1 says synthetic data is generated from training-a, but Section 3.2.1 says single-channel synthetic PCG was generated using ECG from the Icentia dataset, and Section 3.2.2 says synthetic mPCG was generated from training-a and training-b. It is not stated whether the diffusion generators were trained only on the training partition, or whether any test-subject recordings were used as conditioning inputs for generation. If a test subject's ECG/PCG was used to synthesize training signals, test information indirectly enters training. Please specify, for each experiment, exactly which data were used to train the generative models and which data were used as conditioning inputs for the generated training set.","section":"Section 2.1.1 vs Sections 3.2.1 and 3.2.2"},{"comment":"The literature-comparison tables mix metrics from different configurations. In Table 12, the 16kHz All subject-level configuration gives Acc=92.48, UAR=93.05, TPR=93.63, TNR=92.48, F1=94.93, MCC=0.8283, while the 4.125kHz All configuration gives Acc=92.98, UAR=92.48, TPR=91.58, TNR=93.35, F1=95.42, MCC=0.8064. Table 13's 'This study' row lists Acc=92.98, UAR=92.48, TPR=93.63, TNR=92.48, F1=94.93, which is not obtainable from any single configuration in Table 12. Similarly, the text after Table 16 claims the proposed model has the highest metric 'other than TPR', but the RNN row reports TNR=91.10% compared with the proposed model's 90.10%, so TNR is not highest. The SOTA claim should be tied to a single, clearly identified configuration and accurate comparison rows.","section":"Table 13 and accompanying text; Table 16 and accompanying text"},{"comment":"The headline numbers are inconsistent across the abstract, conclusion, and tables. The abstract and conclusion report subject-level accuracy 92.48% and UAR 93.05% for CinC, matching the 16kHz All row in Table 12. However, Table 13 reports Acc 92.98% for 'This study' while also reporting UAR 92.48%, which is the 4.125kHz configuration's UAR, not the 16kHz configuration's. This inconsistency makes it difficult to determine which configuration is being claimed as state of the art. Please harmonize the reported values across the abstract, body, and comparison tables.","section":"Section 4.1, Table 12 vs Table 13; abstract vs Section 4"}],"minor_comments":[{"comment":"The abstract lists six numerical values but names only five metrics ('accuracy, UAR, sensitivity, specificity and MCC'). The sixth value (94.93%) is F1; please name F1 explicitly in the metric list.","section":"Abstract and Conclusion"},{"comment":"The text says 'out of 409 recordings, 405 contain both signal types (288 abnormal, 117 normal)'. The counts sum to 405, but 288/405 = 71.1% abnormal, which is inconsistent with Table 1's training-a abnormal percentage of 67.5%. Please clarify which numbers refer to the subset of 405 versus the full 409.","section":"Section 2.1.1"},{"comment":"Section 3.3 states that for training-a and CinC, synthetic signals are limited to two segments per recording, while Section 3.4.1 states that only three segments are extracted from each synthetic subject. Please make the segment counts consistent.","section":"Sections 3.3 and 3.4.1"},{"comment":"The text says WaveGrad and DiffWave each generated '3,200 patients’ data' using Icentia ECG as conditioners. This is useful detail, but it would help to state whether these synthetic subjects overlap with the CinC training subjects in any way, particularly for the training-a experiments.","section":"Section 3.2.1"}],"recommendation":"major_revision","confidential_remarks":"The central question for the editor is whether the authors can confirm patient-disjoint splitting and a clean synthetic-data boundary. If yes, the paper likely merits publication after revision; if not, the SOTA claims would need to be recomputed. The manuscript would also benefit from a code/data release to verify the split and the generative pipeline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful applied paper with a load-bearing evaluation risk. The headline numbers should not be taken at face value until the authors clarify whether their splits are patient-disjoint and whether the diffusion generators ever saw test subjects.\n\nWhat's actually new: they take the Wav2Vec2 + diffusion augmentation recipe from their prior work [32] and extend it in two directions: a multichannel PCG diffusion model that uses channel-conditioning labels, and a Wav2Vec2 classifier that fuses PCG+ECG or multiple PCG channels. That combination is not in the cited literature. The internal comparisons are consistent and large -- the unaugmented multimodal baseline collapses (TNR around 2%) while the augmented version reaches TNR around 90%, and vest MCC rises from 0.28 to 0.51. That gives real support for the claim that augmentation is doing the work. The architecture is clearly described and the training schedules are detailed enough to reproduce with effort.\n\nThe soft spots are real but mostly addressable. The main one: the CinC 2016 dataset has 3153 recordings from 764 patients, and the paper says only 'A 60-20-20 split is used.' It never states that all recordings from one patient stay in one split. Since subject-level metrics are computed by averaging fragment predictions per subject, recording-level splits would let the model exploit patient identity and inflate the reported 92.48% accuracy and 0.8283 MCC. The same ambiguity applies to the synthetic data: Section 2.1.1 says synthetic data is generated from training-a, but Section 3.2.1 says single-channel synthetic PCG was conditioned on Icentia ECG, and Section 3.2.2 says mPCG synthetic data used training-a and training-b. The paper doesn't say whether the diffusion generators were trained only on the training split. If they saw test subjects, the test set is indirectly observed. No code or data is released, so this can't be checked. The SOTA comparisons in Table 19 compare against a different vest dataset (fewer subjects, breath-held, hand-annotated), so the 'near-SOTA' claim there is softer than the text implies.\n\nMinor issue: the sampling-rate grid search may have been done on validation, but it's not stated explicitly; not a major problem.\n\nWho should read this: people working on scarce physiological audio who want to see how far diffusion augmentation can push transformer fine-tuning. It's a useful data point even if the exact numbers are conditional. It deserves a serious referee, but the referee needs to push on the split and generation boundary. If the authors clarify that and release code, this becomes a solid contribution.\n\nMy recommendation: send it to peer review. Don't desk-reject. The core method and internal comparisons are valuable; the missing split details are fixable, and the paper should not be rejected for them without giving the authors a chance to respond.","headline":"Solid applied ML with a load-bearing evaluation risk: the SOTA numbers hinge on a patient-level split that is never stated.","tokens_in":17601,"tokens_out":3257,"would_cite":false,"duration_ms":36561,"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":"A Wav2Vec2 classifier fine-tuned on synthetic and augmented heart sounds reaches state-of-the-art accuracy on abnormal heart sound detection across single-channel, multimodal, and multichannel inputs.","keywords":["heart sound classification","phonocardiogram","Wav2Vec2","diffusion models","data augmentation","multimodal learning","multichannel PCG","electrocardiogram"],"falsifier":"Inspect the CinC 2016 split by listing patients in each train and test partition; if any patient contributes recordings to both, recompute subject-level accuracy, UAR, and MCC after grouping all recordings by patient. Also check whether the conditioning ECG or PCG inputs used to train WaveGrad and DiffWave include any recordings from test-set subjects.","tokens_in":16697,"feed_emoji":"🫀","tokens_out":3257,"duration_ms":40651,"temperature":0.7,"pith_summary":"The paper tries to show that data scarcity, not architecture, is what holds transformer-based heart-sound classifiers back. By combining traditional signal augmentations with synthetic heart-sound recordings generated by diffusion models, and fine-tuning a Wav2Vec2 encoder, the authors report state-of-the-art results on three setups: single-channel heart sounds, synchronised heart-sound and ECG signals, and six-channel heart sounds from a wearable vest. The practical payoff would be a single scalable pipeline that can take any number of heart-sound channels plus ECG as input and support automated pre-screening for cardiovascular disease. The method is aimed at realistic, noisy, free-breathing conditions, where prior work often required breath-holding and manual heart-cycle labels.","feed_headline":"Synthetic heart sounds lift classifier to state-of-the-art accuracy","feed_subtitle":"WaveGrad and DiffWave audio plus Wav2Vec2 fine-tuning hit 92–93% accuracy across single, paired, and multichannel heart signals.","key_machinery":"The argument rests on coupling a Wav2Vec2 BASE encoder with an enlarged training set produced by two diffusion models and a multichannel-aware augmentation pipeline. WaveGrad and DiffWave synthesize PCG waveforms conditioned on ECG mel-spectrograms and disease labels; for multichannel data, the conditioning label specifies the source channel and the target channel being generated. The classifier concatenates features extracted from each input channel and passes them through an MLP, with LoRA fine-tuning and an SVM added for the small vest dataset. The model is trained in stages that alternate original, synthetic, and augmented data to prevent overfitting.","core_discovery":"The central discovery is that a raw-signal transformer encoder, normally too data-hungry for small medical audio sets, becomes state of the art once its training set is enlarged with a mix of traditional augmentations and synthetic PCG waveforms generated by WaveGrad and DiffWave. On the CinC 2016 single-channel PCG dataset, the paper reports subject-level accuracy of 92.48%, UAR of 93.05%, sensitivity of 93.63%, specificity of 92.48%, and MCC of 0.8283. On the synchronised PCG+ECG training-a subset, it reports 93.14% accuracy, 92.21% UAR, 94.35% sensitivity, 90.10% specificity, and MCC 0.8380. On a six-channel wearable vest mPCG dataset with 157 subjects, it reports 77.13% accuracy, 74.25%","pith_inferences":["If patient-level separation in the reported splits holds, a natural testable projection is that the augmentation benefit should grow as dataset size and signal quality shrink; ablating synthetic-data volume on the vest dataset could confirm this ordering.","The channel-to-channel conditioning scheme for the multichannel diffusion model could transfer to other multi-sensor physiological arrays, such as multi-lead ECG, where the model learns inter-channel dependencies rather than relying on hand-specified relationships.","A focused experiment generating synthetic-only training sets and probing transfer could reveal whether the gain comes from added morphological variability or from regularisation of the Wav2Vec2 encoder; the paper does not disentangle these two effects.","The reported near-parity with breath-held prior work under free-breathing, automatically segmented conditions suggests that the method's practical value may be higher than raw metric comparisons imply, since acquisition burden is much lower."],"forward_implications":["Transformer-based raw-signal classifiers can be made practical for small medical audio datasets when supported by synthetic and augmented training data.","The same architecture scales from single-channel PCG to PCG+ECG to multichannel PCG with only feature-concatenation changes, supporting a unified pre-screening model.","Diffusion-based synthetic data makes the multimodal model trainable: without augmentation, the PCG+ECG model collapsed to roughly 51% UAR with nearly zero specificity.","The strong results on noisy, free-breathing vest data suggest the method tolerates realistic acquisition conditions better than prior work that required breath-holding and manual heart-cycle segmentation.","Sampling-rate tuning matters: 4.125 kHz matched to relevant heart-sound frequency bands outperformed 16 kHz on the multimodal and vest datasets, while 16 kHz worked best on the full CinC dataset."],"fun_headline_variants":["Synthetic heart sounds push classifier to 93% accuracy","Diffusion models synthesize heart sounds, boosting AI to SOTA","Augmented biosignals enable SOTA heart sound classification","WaveGrad and DiffWave jolt Wav2Vec2 to top heart-sound scores","Synthetic PCG+ECG signals unlock SOTA heart-sound classification"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The reported subject-level numbers assume the 60/20/20 split separates patients, not just recordings, and that the diffusion generators never see test-subject data; the text never explicitly states either condition.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic heart sounds push classifier to 93% accuracy","Diffusion models synthesize heart sounds, boosting AI to SOTA","Augmented biosignals enable SOTA heart sound classification","WaveGrad and DiffWave jolt Wav2Vec2 to top heart-sound scores","Synthetic PCG+ECG signals unlock SOTA heart-sound classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000795,"raw_usage":{"total_tokens":3465,"prompt_tokens":999,"completion_tokens":2466,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":743,"completion_tokens_details":{"reasoning_tokens":2372}},"tokens_in":743,"tokens_out":2466,"duration_ms":20482,"temperature":1.0,"reasoning_tokens":2372,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:45:44.251963+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the CinC 2016 split by listing patients in each train and test partition; if any patient contributes recordings to both, recompute subject-level accuracy, UAR, and MCC after grouping all recordings by patient. Also check whether the conditioning ECG or PCG inputs used to train WaveGrad and DiffWave include any recordings from test-set subjects.","supporting_citations":[],"review_version":1}