{"id":"618f8be0-e1a6-418a-a274-0f1e71d1ea8e","arxiv_id":"2505.23132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Patient Domain Supervised Contrastive Learning (PD-SCL) improves lung sound classification on mobile-phone recordings by 2.4 points over an AST baseline, but the result relies on a small private dataset with no error bars.","lead":"The paper adds patient and device identity to supervised contrastive learning for lung sound classification and reports a 2.4-point score gain over a plain Audio Spectrogram Transformer. It argues this makes smartphone-based lung sound screening more feasible, but the reported gain lacks statistical validation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2.4-point gain may be inflated unless the validation patients' stethoscope recordings are excluded from training; the split description in Section IV-A is ambiguous on exactly this point.","rationale":"The reader's verdict is already CONDITIONAL, and its weakest assumption identifies both the patient-disjoint split and the ill-defined loss equation. My stress-test focuses on the split, because that is the most load-bearing condition for the central claim: if stethoscope data from validation patients leaked into training, the 2.4-point gain would not demonstrate cross-patient generalization. I agree with the reader that this is a real ambiguity, but I do not see evidence sufficient to move the verdict to REJECT: the paper does contain the sentence 'Leave-subject-out cross-validation was applied to the patients used in the test data so that they were not used in the training process,' which, read generously, asserts the needed exclusion. The problem is that the preceding sentence and the training/validation description do not make the mechanics of the split explicit, and the source of the ambiguity is the exact point on which the entire empirical claim rests. A concrete check on the fold assignments would settle this. The ill-defined Eq. (4) reinforces the need for code or a corrected formula, but it is a reproducibility defect rather than the primary correctness risk. Because the required clarification is obtainable and the current conditional verdict already makes release contingent on it, I would leave the verdict unchanged rather than escalate.","tokens_in":5699,"tokens_out":3419,"duration_ms":37536,"concrete_test":"Obtain the fold assignment or subject IDs (or the released split-generation code) and, for each of the 5 folds, count whether any stethoscope recording from a patient whose mobile recording is in the validation fold appears in the training set. If the count is zero, the split is clean. If it is nonzero, rerun both AST combined baseline and PD-SCL with those stethoscope files removed from training, using the same hyperparameters, and compare the Sc gap. If the gap collapses or changes materially, the headline 2.4% improvement is inflated by same-patient leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Table II — PD-SCL Sc=86.3 vs AST combined Sc=83.9 — is a claim about generalization to new patients. For that claim to hold, every training fold must exclude stethoscope recordings from the patients whose mobile-phone recordings form the validation fold. The paper's split description (Section IV-A) is ambiguous: it says 'using both stethoscope and mobile phone recordings for training, but using only mobile phone recordings for validation,' and then says leave-subject-out was applied 'to the patients used in the test data so that they were not used in the training process.' Because all 63 mobile-phone patients also have corresponding stethoscope recordings, a training set that includes all stethoscope data while validating only mobile data would put same-patient stethoscope audio in the training set for every validation sample. That is exactly the patient-specific leakage that could inflate the reported gain, since PD-SCL explicitly models patient/domain groupings and would benefit more than a plain cross-entropy baseline from seeing the same patient's stethoscope sounds at training time. The comparison would then measure adaptation to recording device rather than generalization to unseen patients. A secondary reproducibility issue is Eq. (4), whose denominator sums over k with k ≠ j while j is bound in the numerator; this makes the loss as written undefined, but it is secondary to the split question because the intended contrastive loss can be inferred and corrected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Patient Domain Supervised Contrastive Learning (PD-SCL), an auxiliary loss for the Audio Spectrogram Transformer (AST) applied to binary lung-sound classification (normal vs. abnormal) from recordings made with a digital stethoscope and an iPhone microphone. The method adds a supervised contrastive term intended to pull together features from recordings that share a class label but come from different patients or recording domains. On a private pediatric dataset of 8,724 stethoscope and 2,701 mobile-phone recordings, the authors report that AST with PD-SCL achieves Sc=86.3 versus Sc=83.9 for AST fine-tuned on combined stethoscope+phone data, a claimed 2.4-point improvement, using leave-subject-out 5-fold cross-validation with mobile-phone recordings as the validation set.","tokens_in":5967,"tokens_out":2851,"duration_ms":31890,"significance":"If the reported result is valid, the paper would provide a practically relevant demonstration that a transformer-based audio model can be adapted from stethoscope recordings to smartphone recordings while controlling for patient-specific variability, which is useful for telemedicine and remote screening. The manuscript's strengths include a clinically collected paired dataset, a clear motivation for combining abundant stethoscope data with scarce mobile-phone data, and a direct comparison against a domain-adversarial baseline. However, the central empirical claim currently rests on an ambiguous data-split description and an ill-defined loss equation, and the quantitative comparison lacks error bars or significance testing. These issues are fixable but must be resolved before the claimed improvement can be taken as established.","major_comments":[{"comment":"The data-split description is ambiguous on the load-bearing point of patient-disjointness. The text states that 'using both stethoscope and mobile phone recordings for training, but using only mobile phone recordings for validation' and that leave-subject-out cross-validation was applied 'to the patients used in the test data so that they were not used in the training process.' Because every mobile-phone patient also has corresponding stethoscope recordings, this wording could mean that stethoscope recordings from the very patients whose mobile recordings form a validation fold are included in the training set of that fold. If so, the comparison measures adaptation to the recording device, not generalization to unseen patients, and PD-SCL, which explicitly uses patient/domain groupings, would benefit more than the plain cross-entropy baseline. The paper must state unambiguously that, for each fold, all stethoscope recordings from validation patients are excluded from training, or provide the exact patient-level fold construction.","section":"Section IV-A"},{"comment":"Equation (4) is ill-defined as written. The numerator sums over j ∈ P(i), while the denominator sums over k with k ≠ j, but j is a bound variable in the numerator, so the denominator cannot reference it. The intended contrastive loss is inferable, but the actual implementation could differ from any reader's reconstruction. The authors should replace Eq. (4) with a standard supervised-contrastive form, for example, using a denominator that sums over all anchors k ≠ i (or over all non-positive samples k ∉ P(i)), and should state explicitly whether positive pairs require different patient IDs, different domain IDs, or both, and whether same-patient same-class pairs are treated as negatives or excluded.","section":"Equation (4), Section III-C"},{"comment":"The central comparison is reported without any measure of variability. Only a single Sc value is given for each method across the 5-fold cross-validation, with no per-fold scores, standard deviation, confidence interval, or significance test. The claimed 2.4-point improvement (86.3 vs. 83.9) may be within fold-to-fold noise, especially given the small size of the mobile-phone validation set (2,701 recordings from 63 patients). The authors should report per-fold results and either confidence intervals or a paired significance test across folds.","section":"Table II and Section IV-D"}],"minor_comments":[{"comment":"The word 'significantly' is used to describe the 2.4% improvement, but no statistical significance test is reported; please either add a test or rephrase to 'numerically improved.'","section":"Abstract and Section I"},{"comment":"The paragraph ends with 'so that they were not used in the training process' and is missing a terminal period; the sentence should also be completed with a direct statement that this exclusion applies to stethoscope as well as mobile-phone recordings.","section":"Section IV-A, last sentence"},{"comment":"The description 'The sum of negative similarities is calculated and normalized by the number of pairs' is vague; please specify whether the normalization is by the number of negative pairs or by the total number of samples, and align it with the corrected equation.","section":"Section III-C, step 5"},{"comment":"The ROC curve figure is not described in the text beyond a single mention; the figure should include axis labels, the AUC for each model, and a legend identifying each curve, and the text should state what the figure demonstrates.","section":"Figure 1"},{"comment":"Reference [9] describes 'Stethoscope-guided supervised contrastive learning,' which appears very closely related to the proposed method; the authors should explicitly state the differences in the positive-pair definition and in the use of patient identities, rather than only listing the reference.","section":"Section II-C and Reference [9]"},{"comment":"The evaluation metric section uses ICBHI terminology but only presents a binary normal/abnormal task; please clarify whether crackle, wheeze, and both are treated as a single abnormal class, and define the exact mapping from the four ICBHI labels to the two classes.","section":"Section IV-C"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue and addresses a practical problem, but the ambiguity in the data-split description is critical because it directly affects whether the central 2.4% improvement is a valid patient-generalization result. The ill-defined loss equation is also a reproducibility barrier. Both issues are fixable, so major revision is appropriate rather than rejection. The authors should also be asked to clarify the novelty relative to their own closely related prior work in reference [9], since the current text does not distinguish the two methods."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a legitimate incremental extension of stethoscope-guided supervised contrastive learning: it adds patient ID to positive pair selection, which is a small but real variation on Ref. [9]. The paired pediatric dataset of stethoscope and iPhone recordings is useful, and the comparison against a mobile-only AST baseline and domain-adversarial training is reasonable. Credit where due: the authors identify a real problem (device and patient variability) and propose a sensible fix.\n\nBut the central claim does not hold up as stated. The stress-test concern is correct. Section IV-A says both stethoscope and mobile recordings are used for training but only mobile recordings for validation, with leave-subject-out applied to test patients. Because every mobile patient also has stethoscope recordings, that split description permits same-patient stethoscope audio in training for every validation sample. If that is what was done, the 2.4-point gain reflects adaptation to the recording device, not generalization to unseen patients. This is load-bearing, not a nitpick. The sentence \"Leave-subject-out cross-validation was applied to the patients used in the test data\" is ambiguous and must be clarified.\n\nThere are secondary problems. Equation (4) is ill-defined: the denominator sums over k with k ≠ j while j is bound in the numerator. The intended contrastive loss is inferable, but as written it is not a valid loss. There are no error bars, no per-fold results, and no significance test, so \"significantly improved\" is unsupported. The dataset is private, and no code is released, so reproducibility is limited. The novelty relative to Ref. [9] is modest; the paper does not overclaim that, but it does claim a robustness and generalizability that the evidence does not yet establish.\n\nThe method is coherent and the question matters for remote and low-resource settings, so this deserves a serious referee. But the referee should ask for the corrected loss, an explicit description of exactly which patients' stethoscope recordings enter each training fold, per-fold numbers, and ideally code and data.\n\nMy recommendation: send to peer review, but expect heavy revision. I would not cite it in its current form.","headline":"A plausible but incremental contrastive-learning variant whose 2.4-point gain hinges on an ambiguous patient split that could easily be leakage; the loss equation is also ill-defined as written.","tokens_in":6450,"tokens_out":1670,"would_cite":false,"duration_ms":19595,"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":"The paper claims that adding a patient-domain supervised contrastive loss to the Audio Spectrogram Transformer improves smartphone lung-sound classification by 2.4 points over the same model trained without it.","keywords":["lung sound classification","supervised contrastive learning","patient domain adaptation","audio spectrogram transformer","smartphone microphone","pediatric respiratory sounds","domain shift","mobile health"],"falsifier":"Implement equation (4) literally as printed: it cannot be evaluated because the denominator sums over $k$ with $k\\neq j$ while $j$ is bound in the numerator, which would falsify the claim that the loss is fully specified. If the intended denominator is instead $k\\neq i$, then the decisive check is whether the same leave-subject-out protocol reproduces Sc=86.3 and a 2.4-point gain over the 83.9 combined-data baseline; a smaller or vanished gap would show the printed definition does not support the claimed improvement.","tokens_in":5510,"feed_emoji":"📱","tokens_out":12506,"duration_ms":111384,"temperature":0.7,"pith_summary":"The paper asks whether ordinary smartphone microphones can record lung sounds accurately enough for automated classification, and it proposes a training method to bridge two gaps: the audio-style gap between electronic stethoscopes and phone microphones, and the variability between individual patients. Its central claim is that a contrastive loss organized by patient identity, called Patient Domain Supervised Contrastive Learning (PD-SCL), closes both gaps when the Audio Spectrogram Transformer (AST) is fine-tuned on combined stethoscope and mobile data. On a private pediatric dataset with leave-subject-out 5-fold cross-validation, the method scores Sc=86.3, a 2.4-point improvement over the same AST model fine-tuned on the combined data without PD-SCL (83.9). If that result holds, smartphone-based lung sound screening becomes a realistic complement to in-clinic auscultation, especially where stethoscopes and specialists are scarce.","feed_headline":"Smartphone lung-sound AI gains 2.4 points via patient-aware loss","feed_subtitle":"A patient-aware loss closes the device gap, so phone recordings can be classified with no stethoscope data from the same patient.","key_machinery":"The load-bearing object is the PD-SCL loss, a supervised contrastive objective computed on L2-normalized feature vectors extracted from the AST model. The similarity between samples $i$ and $j$ is a temperature-scaled dot product $\\mathrm{Sim}(i,j)=f_i^T f_j/\\tau$, and positive pairs are defined as same-label pairs from different patients or domains while negative pairs have different labels. The loss contrasts the sum of positive similarities against the sum of negative similarities in a log-sum-exp form, and it is combined with cross-entropy as $L_{\\mathrm{total}}=L_{\\mathrm{CE}}+\\lambda L_{\\mathrm{PD\\text{-}SCL}}$ with $\\lambda=0.5$. The mechanism is meant to pull same-class representations together across devices and patients while pushing different classes apart, which is how it aims to reduce both the device gap and patient variability.","core_discovery":"On the authors' terms, the discovery is that a supervised contrastive loss defined over patient and device domains transfers knowledge from abundant electronic-stethoscope recordings to scarce mobile-phone recordings without sacrificing sensitivity. In PD-SCL, two samples form a positive pair when they share a label but come from different patients or different recording devices, and different-label samples form negative pairs; the loss is added to the cross-entropy objective with weight $\\lambda=0.5$ and temperature $\\tau=0.5$. Under leave-subject-out 5-fold cross-validation, with both stethoscope and mobile recordings used for training but only mobile recordings used for validation, PD-SCL reaches $S_p=89.9$, $S_e=82.7$, $S_c=86.3$, outperforming AST fine-tuned on mobile data alone (84.1), AST fine-tuned on combined data (83.9), and a domain-adversarial training baseline (83.7).","pith_inferences":["The paper leaves open how much of the 2.4-point gain comes from the contrastive loss itself versus the way positives are sampled by patient and device; an ablation swapping PD-SCL for standard supervised contrastive learning would separate these.","A natural extension is to test PD-SCL on a public respiratory sound benchmark where patient identifiers are available, since the reported gain is demonstrated on one private pediatric dataset.","Because the printed loss formula is not well-defined as written, a reader who reproduces the method will have to infer the intended denominator; if that denominator is the standard one over all other samples, then PD-SCL is best understood as a positive-pair sampling rule rather than a fundamentally new loss."],"forward_implications":["If PD-SCL performs as reported, a model can be trained on stethoscope recordings and still classify smartphone recordings from patients whose stethoscope data was never seen.","The evaluation protocol, with validation restricted to mobile recordings, means the reported Sc=86.3 estimates generalization to new patients rather than fit to known subjects.","The loss recipe (cross-entropy plus a weighted PD-SCL term) gives other researchers a concrete starting point for combining multi-device medical audio.","Treating crackle, wheeze, and both as a single abnormal class keeps the screening task simple enough for a phone-based tool."],"supporting_citations":[{"why":"Supplies the Audio Spectrogram Transformer backbone used in all experiments.","marker":"[17]"},{"why":"Defines the supervised contrastive loss family that PD-SCL adapts to patient and device domains.","marker":"[14]"},{"why":"Provides the domain-adversarial training baseline that PD-SCL is compared against.","marker":"[18]"},{"why":"Introduces the stethoscope-guided supervised contrastive approach whose cross-domain idea PD-SCL extends.","marker":"[9]"},{"why":"Shows contrastive learning combined with the AST for respiratory sounds, the setup PD-SCL builds on.","marker":"[8]"},{"why":"Supplies the contrastive learning formulation used for representation alignment.","marker":"[15]"}],"fun_headline_variants":["Patient-aware loss boosts smartphone lung-sound AI by 2.4%","Phone lung-sound AI up 2.4% with patient-aware contrastive loss","Contrastive learning closes stethoscope-to-phone gap in lung sounds","Phone mic lung diagnosis gains 2.4% via patient-aware AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 2.4-point gain depends on keeping every stethoscope recording of validation patients out of training and on implementing the PD-SCL loss in the intended form, because equation (4) as printed has a denominator whose summation variable is not defined.","fun_headline_variants_meta":{"raw":{"variants":["Patient-aware loss boosts smartphone lung-sound AI by 2.4%","Phone lung-sound AI up 2.4% with patient-aware contrastive loss","Contrastive learning closes stethoscope-to-phone gap in lung sounds","Phone mic lung diagnosis gains 2.4% via patient-aware AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00099,"raw_usage":{"total_tokens":4187,"prompt_tokens":925,"completion_tokens":3262,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":3178}},"tokens_in":541,"tokens_out":3262,"duration_ms":22915,"temperature":1.0,"reasoning_tokens":3178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:52:00.665913+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement equation (4) literally as printed: it cannot be evaluated because the denominator sums over $k$ with $k\\neq j$ while $j$ is bound in the numerator, which would falsify the claim that the loss is fully specified. If the intended denominator is instead $k\\neq i$, then the decisive check is whether the same leave-subject-out protocol reproduces Sc=86.3 and a 2.4-point gain over the 83.9 combined-data baseline; a smaller or vanished gap would show the printed definition does not support the claimed improvement.","supporting_citations":[{"cited_title":"Supervised contrastive learn- ing,","cited_arxiv_id":null,"evidence_quote":"Defines the supervised contrastive loss family that PD-SCL adapts to patient and device domains."},{"cited_title":"A simple framework for contrastive learning of visual representations,","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive learning formulation used for representation alignment."}],"review_version":1}