{"id":"06713e64-fd46-4800-b270-15eb2c7c7300","arxiv_id":"2505.20745","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Pre-trained acoustic foundation models estimate heart rate from phonocardiogram recordings with accuracy similar to a conventional feature-based method.","lead":"This paper tests whether pre-trained acoustic AI models can estimate heart rate from recorded heartbeat sounds. It finds the models work about as well as a traditional audio-feature method, with one proprietary model slightly better.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline CLAP advantage is a 0.03 bpm gap obtained by selecting the best of 12 embedding layers on test data, with no multiple-comparison or significance testing; the 'outperforms' claim is therefore not statistically supported.","rationale":"The reader's weakest_assumption flags label reliability from S1-onset annotations, which is a legitimate concern given that the entire benchmark target is generated from those annotations and no independent validation is offered. However, the most load-bearing weakness for the paper's advertised claim is statistical: the 0.03 bpm advantage is smaller than the noise in the measurement and is obtained through test-set layer selection. Table 1's min(MAE) explicitly chooses the best embedding layer, and the paper does not report whether layer 6 was pre-registered or selected on validation data. Without a paired significance test or a correction for the 12 layers examined, the observed gap is well within what chance could produce. This does not invalidate the paper's broader finding that FM representations perform comparably to acoustic features; it only invalidates the stronger 'outperforms' wording. The reader's conditional verdict already captures the need for reanalysis, so I would keep it CONDITIONAL rather than escalate to REJECT. The paper's benchmark is useful, and the overclaim can be fixed by re-analysis, multiple-comparison correction, and softened conclusions.","tokens_in":8369,"tokens_out":4773,"duration_ms":53063,"concrete_test":"Compute per-split MAE differences between in-house CLAP layer 6 and the baseline across the six splits, then run a paired Wilcoxon signed-rank or permutation test on these six differences; if p>0.05, the superiority claim fails. To address layer selection, repeat using a family-wise correction over the 12 layers (e.g., a permutation null taking the maximum layer-wise mean difference, or Holm-Bonferroni on layer-wise p-values), and report the corrected p-value. If the corrected comparison is not significant, revise the abstract to say FM representations are 'comparable' to the baseline and report the in-house CLAP result as a non-significant trend.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the abstract's statement that in-house CLAP 'outperforms' the baseline, supported by MAE 1.88 vs 1.91 in Table 1. This comparison is not load-bearing on a single fixed model: 1.88 is the minimum of the mean MAE over the 12 embedding layers of the CLAP encoder, so the layer (layer 6) was selected by looking at test performance on the same six splits. Under a null where no layer is better than the baseline, the minimum of 12 positively correlated estimates can easily fall below 1.91; a 0.03 bpm gap is an order of magnitude smaller than the reported split-level spread (CLAP sigma=0.37, baseline STD=0.32). No paired test, confidence interval, or multiple-comparison correction is reported for the key comparison. The label-reliability concern in Section 2 is real: S1-onset-derived HR targets inherit any CirCor annotation bias, and the paper itself flags 'annotation bias' in Section 5. But even granting the labels, the current numbers do not establish superiority. The paper's broader 'comparable performance' finding is consistent with the data; the specific 'outperforms' claim is not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper benchmarks hidden representations from six pre-trained acoustic foundation models (HuBERT, wav2vec2, wavLM, Whisper, CLAP, and an in-house CLAP) for heart rate (HR) estimation from phonocardiogram snippets. Using the CirCor dataset, heart sound snippets are generated with 5-second windows, HR targets are derived from S1-onset annotations via the method of Nie et al. [1], and a downstream 2D CNN is trained on embeddings from each layer of each model. The reported result is that FM representations generally achieve MAE comparable to the acoustic-feature baseline (average MAE 1.91 bpm), with the 6th layer of in-house CLAP achieving an average MAE of 1.88 bpm, which the abstract and conclusions describe as outperforming the baseline.","tokens_in":8712,"tokens_out":2553,"duration_ms":29102,"significance":"If the findings are properly supported, the paper would provide a useful benchmark showing that frozen self-supervised audio representations are competitive with handcrafted acoustic features for auscultation-based HR estimation. The layer-wise analysis across six models and six held-out splits is a useful contribution, and the reporting of model sizes and per-split variance is commendable. The strongest parts are the breadth of the layer sweep and the use of a public dataset with multiple splits. However, the headline superiority claim for in-house CLAP rests on a 0.03 bpm difference that is not statistically tested, and the label-generation procedure is not independently validated; these issues are load-bearing for the paper's central claim.","major_comments":[{"comment":"The claim that in-house CLAP 'outperforms' the baseline is not statistically supported. The evidence is min(MAE)=1.88 for the 6th layer versus baseline MAE=1.91, a difference of 0.03 bpm, while the reported split-level spread is σ=0.37 for in-house CLAP at that layer and STD=0.32 for the baseline. The 6th layer was selected as the best among 12 embedding layers after examining test performance on the same six splits, so the minimum-of-12 comparison is subject to selection bias. No paired test, confidence interval, or multiple-comparison correction is reported for this key comparison. The authors should report per-split paired differences and a paired significance test, or use a held-out layer-selection procedure; otherwise the conclusion should be limited to 'comparable performance.'","section":"Section 5, Table 1, Abstract"},{"comment":"The ground-truth HR for each snippet is generated from S1-onset annotations in the CirCor dataset using the method of the authors' prior paper [1], and the manuscript itself acknowledges 'biases and errors in segmentation annotations' (Section 2) and 'annotation bias' (Section 5). Because both the baseline method and the FM-based models are evaluated against these derived labels, the reliability of the label-generation procedure is load-bearing for every reported MAE. No independent validation of the derived HR targets is provided (for example, against ECG-derived HR or manual reference measurements). The authors should either provide such validation or explicitly and consistently frame the results as relative performance under this particular label-generation protocol.","section":"Section 2 and Section 5"}],"minor_comments":[{"comment":"The phrase 'train/validation/data splits' should read 'train/validation/test splits.'","section":"Section 1"},{"comment":"The displayed cross-entropy formula has mismatched subscripts (x_{a,c} versus x_{n,i}) and is not written in standard form; please correct the notation so the sum over classes and the softmax argument are unambiguous.","section":"Section 4, Eq. (1)"},{"comment":"Please clarify that min(MAE) is the minimum over embedding layers of the mean MAE across splits, and that σ_{MAE,j*} is the standard deviation across splits at the layer achieving that minimum; the current header is dense and could be misinterpreted as a minimum over individual split results.","section":"Table 1"},{"comment":"The shaded standard-deviation bands in Figure 5 are useful, but the figure legends do not state the number of splits underlying each mean; adding that information would improve interpretability.","section":"Section 5, Figure 5"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the layer-wise comparison of six acoustic foundation models on phonocardiogram heart rate estimation is genuinely new and practically useful—I'm not aware of another published benchmark that runs HuBERT, wav2vec2, WavLM, Whisper, CLAP, and a proprietary CLAP through the same downstream CNN on the CirCor dataset with six subject-disjoint splits. Second, the headline claim that in-house CLAP \"outperforms\" the baseline is not supported by the numbers as presented. The 1.88 vs 1.91 bpm MAE difference is the minimum over twelve embedding layers, chosen after looking at test performance, and the split-level standard deviations are 0.37 and 0.32. No paired test, confidence interval, or multiple-comparison correction is reported. The stress-test note gets this right.\n\nWhat the paper does well: the \"comparable performance\" message is honest and consistent with the data. Some FMs show smaller variance across splits than the acoustic-feature baseline, which is a real, useful observation for practitioners. The authors also acknowledge the annotation-bias issue in Section 2 and again in Section 5 when discussing poor predictions, and they point out that larger models don't necessarily help. That is fair and non-hype.\n\nSoft spots, in proportion: the \"outperforms\" claim is the load-bearing weakness and should be fixed with proper statistics or dropped. The label-reliability concern is real—heart rate targets derived from S1 onsets inherit any CirCor annotation bias—but it is not unique to this paper, and the authors flag it. Lack of code and unreleased in-house model limit reproducibility, but that is common for industry papers and less damaging for a benchmark that mostly uses public FMs. The self-referential baseline from the authors' own prior work is not a problem in itself; the FM embeddings are evaluated independently against it.\n\nWho this is for: anyone choosing between acoustic features and pretrained representations for PCG vital-sign estimation, and researchers interested in layer-wise properties of audio FMs beyond speech. It deserves a serious referee. The paper would be acceptable after the authors either add significance testing with proper multiple-comparison correction or soften the abstract to \"comparable, sometimes more stable\" and leave the CLAP advantage as a suggestive observation.\n\nRecommendation: send it to peer review, but tell the authors the current \"outperforms\" claim needs statistical backing before publication.","headline":"A useful layer-wise benchmark of audio FMs for PCG heart rate estimation, but the headline CLAP advantage is a post-hoc 0.03 bpm gap with no significance testing; the 'comparable performance' finding is solid, the 'outperforms' claim is not.","tokens_in":9120,"tokens_out":1526,"would_cite":true,"duration_ms":18018,"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":"Heart rate can be estimated from heart-sound recordings using representations from pre-trained audio foundation models, and one contrastively trained encoder beats the acoustic-feature baseline at 1.88 vs 1.91 bpm mean absolute error.","keywords":["heart rate estimation","phonocardiogram","auscultation","acoustic foundation models","self-supervised learning","layer-wise analysis","CLAP","transfer learning"],"falsifier":"Rerun the same layer-wise comparison on the same snippets but with heart-rate labels obtained from an independent reference, such as a synchronized ECG recording or expert re-annotation of S1/S2 boundaries; the central claim fails if the in-house CLAP's 1.88 bpm result no longer beats the baseline's 1.91 bpm under independently verified labels.","tokens_in":8176,"feed_emoji":"🫀","tokens_out":9104,"duration_ms":82364,"temperature":0.7,"pith_summary":"This paper asks whether hidden representations of self-supervised audio foundation models, originally trained on speech and general audio events, carry enough information about heart sounds to estimate heart rate from phonocardiogram recordings. Using a public PCG dataset split six ways, the authors compare representations from every encoder layer of six foundation models against a handcrafted acoustic-feature baseline. They find the foundation-model representations are broadly comparable to the baseline, and the audio encoder of an in-house CLAP model trained on general audio events achieves a lower mean absolute error (1.88 bpm) than the baseline (1.91 bpm) across all splits, despite the domain mismatch. A layer-wise analysis shows that mid-level layers of this CLAP encoder carry the most useful cardiac information, and that deeper layers of speech-oriented models lose cardiorespiratory content. The result matters because it suggests general audio foundation models can be reused for vital-sign monitoring without retraining the encoder.","feed_headline":"Foundation model audio features estimate heart rate from heart sounds","feed_subtitle":"One CLAP encoder reaches 1.88 bpm mean error on phonocardiograms, under the 1.91 bpm acoustic baseline.","key_machinery":"The machinery is the layer-wise extraction of hidden representations: for each foundation model, the output of the $n$-th embedding layer of its audio encoder is taken from 5-second PCG snippets and fed into a downstream 2D CNN classifier that predicts heart rate as one of 141 classes ($HR \\in [40,180]$ bpm) with cross-entropy loss. The in-house CLAP model, which carries the main positive result, is a 12-layer ViT audio encoder with 768 hidden dimensions, pre-trained in two stages: first by masked spectrogram reconstruction in the style of AudioMAE on roughly three million audio samples, then by contrastive alignment with a text encoder on about five million captions. This two-stage, general-audio-event pretraining is what the paper credits for the encoder's ability to preserve cardiac cues in mid-level layers.","core_discovery":"The central claim is that pre-trained acoustic foundation-model representations are a viable feature source for heart-rate estimation from auscultation, and that a contrastively trained audio encoder can beat a purpose-built acoustic-feature method. In the paper's layer-wise evaluation, each FM's embedding-layer output feeds a small 2D CNN that classifies heart rate into 141 bins from 40 to 180 bpm; the best single result is the sixth embedding layer of the in-house CLAP model, which reaches an average MAE of 1.88 bpm across six train/validation/test splits, versus 1.91 bpm for the baseline. The paper further reports that the shallow layers of ASR-oriented models (HuBERT, wav2vec2, WavLM, Whisper) generally perform better than deep layers, that larger model size does not guarantee better heart-rate estimates, and that averaging the baseline's predictions with FM-based predictions improves accuracy on every split. The authors attribute the in-house CLAP advantage to its training on a wide range of audio events rather than speech-dominated corpora.","pith_inferences":["A testable extension: probe the same six encoders on other cardiorespiratory sounds (breathing, cough, murmurs); the layer-wise pattern predicts that mid-level layers of contrastive audio encoders will again carry the most signal.","A concrete next experiment: concatenate the baseline acoustic features with the in-house CLAP layer-6 vector before the downstream CNN; the paper's averaging result suggests fusion would lower the MAE below 1.88 bpm.","Because the reported difference is only 0.03 bpm, an independent-label rerun (ECG-synchronized) is needed to know whether the true effect is bigger, smaller, or absent; the paper does not provide that check.","The paper's attribution of the advantage to audio-event pretraining breadth could be isolated by ablating the two pretraining stages of the in-house CLAP encoder, but that ablation is not reported here."],"forward_implications":["General audio foundation models can be used as frozen feature extractors for heart-rate estimation, avoiding task-specific pretraining and achieving accuracy comparable to handcrafted acoustic features.","The in-house CLAP's advantage over the baseline (1.88 vs 1.91 bpm MAE) identifies contrastive audio-event pretraining, rather than speech-oriented pretraining, as the more transferable recipe for cardiac acoustics.","The layer-wise curves provide a direct fine-tuning target: for ASR-based models, shallow and middle layers retain cardiorespiratory information while deep linguistic layers lose it, so adaptation should focus on the earlier layers.","Averaging FM-based and acoustic-feature predictions improves accuracy on every split, suggesting ensemble schemes that combine both feature families are a practical route to more robust HR estimation.","Evaluation across multiple data splits is necessary; the paper finds that data-split variation can affect model comparisons more than model size, and larger FMs do not systematically help."],"supporting_citations":[{"why":"Supplies the acoustic-feature baseline, the S1-onset-based heart-rate labeling procedure, and the downstream CNN architecture the paper compares against and adapts.","marker":"[1]"},{"why":"Provides the public phonocardiogram dataset with S1/S2 segmentation annotations that define the snippets and labels used in every experiment.","marker":"[23]"},{"why":"Defines HuBERT, one of the six compared foundation models, and its masked-prediction self-supervised pretraining.","marker":"[24]"},{"why":"Defines wav2vec2, one of the compared speech-oriented foundation models.","marker":"[25]"},{"why":"Defines WavLM, one of the compared foundation models.","marker":"[26]"},{"why":"Defines Whisper, one of the compared foundation models.","marker":"[27]"},{"why":"Defines CLAP's contrastive language-audio pretraining, the objective the in-house CLAP model adapts.","marker":"[28]"},{"why":"Defines the AudioMAE masked-reconstruction pretraining that the in-house CLAP audio encoder's first stage follows.","marker":"[29]"},{"why":"Supplies the layer-wise analysis methodology that motivates probing every encoder layer for cardiorespiratory content.","marker":"[19]"}],"fun_headline_variants":["General audio model beats heart-rate baseline on phonocardiogram","CLAP encoder edges past acoustic baseline for heart-rate estimation","1.88 bpm error: general audio model reads heart rate from sounds","Heart rate from heart sounds: AI audio model beats acoustic baseline","Pre-trained acoustic FMs encode heart rate in auscultation signals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ground-truth heart rate of each snippet is derived from the S1 onset annotations in the public PCG dataset using the labeling method of the baseline paper, so if those onset annotations carry systematic bias or noise, every reported MAE is measured against an unreliable target.","fun_headline_variants_meta":{"raw":{"variants":["General audio model beats heart-rate baseline on phonocardiogram","CLAP encoder edges past acoustic baseline for heart-rate estimation","1.88 bpm error: general audio model reads heart rate from sounds","Heart rate from heart sounds: AI audio model beats acoustic baseline","Pre-trained acoustic FMs encode heart rate in auscultation signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001237,"raw_usage":{"total_tokens":5095,"prompt_tokens":979,"completion_tokens":4116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":4027}},"tokens_in":595,"tokens_out":4116,"duration_ms":32027,"temperature":1.0,"reasoning_tokens":4027,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:46:39.146578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same layer-wise comparison on the same snippets but with heart-rate labels obtained from an independent reference, such as a synchronized ECG recording or expert re-annotation of S1/S2 boundaries; the central claim fails if the in-house CLAP's 1.88 bpm result no longer beats the baseline's 1.91 bpm under independently verified labels.","supporting_citations":[{"cited_title":"Exploring wav2vec 2.0 model for heart murmur detection,","cited_arxiv_id":null,"evidence_quote":"Defines Whisper, one of the compared foundation models."},{"cited_title":"Foundation Model Hidden Representations for Heart Rate Estimation from Auscultation","cited_arxiv_id":"2505.20745","evidence_quote":"Supplies the acoustic-feature baseline, the S1-onset-based heart-rate labeling procedure, and the downstream CNN architecture the paper compares against and adapts."},{"cited_title":"Speech emotion recognition using self- supervised features,","cited_arxiv_id":null,"evidence_quote":"Provides the public phonocardiogram dataset with S1/S2 segmentation annotations that define the snippets and labels used in every experiment."},{"cited_title":"Multi-modal dataset across exertion levels: Capturing post-exercise speech, breathing, and phonocardio- gram,","cited_arxiv_id":null,"evidence_quote":"Defines HuBERT, one of the six compared foundation models, and its masked-prediction self-supervised pretraining."},{"cited_title":"Layer-wise anal- ysis of a self-supervised speech representation model,","cited_arxiv_id":null,"evidence_quote":"Defines wav2vec2, one of the compared speech-oriented foundation models."},{"cited_title":"Pre-Trained Foundation Model representations to uncover Breathing patterns in Speech","cited_arxiv_id":"2407.13035","evidence_quote":"Defines WavLM, one of the compared foundation models."},{"cited_title":"Foundation models for cardiovascular disease detection via biosignals from digital stethoscopes,","cited_arxiv_id":null,"evidence_quote":"Defines CLAP's contrastive language-audio pretraining, the objective the in-house CLAP model adapts."},{"cited_title":"The circor digiscope phonocardiogram dataset,","cited_arxiv_id":null,"evidence_quote":"Defines the AudioMAE masked-reconstruction pretraining that the in-house CLAP audio encoder's first stage follows."},{"cited_title":"Learning image-based representations for heart sound classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the layer-wise analysis methodology that motivates probing every encoder layer for cardiorespiratory content."}],"review_version":1}