{"id":"ffa35173-08a2-481c-80ef-fb5bb8895c79","arxiv_id":"2506.10698","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-encoder masked autoencoder with time-shuffle Siamese and vCLUB mutual-information losses reaches an ICBHI average score of 61.50 and the highest sensitivity among compared respiratory sound classifiers.","lead":"This paper introduces DDE-MAE, a masked autoencoder with two encoders that tries to split disease-related from disease-irrelevant features in respiratory sounds without using domain labels. It reports competitive accuracy on the ICBHI benchmark and argues that the split can make such models more robust to different stethoscopes, patients and recording environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Time-shuffle invariance likely preserves disease-discriminative spectral content; without a disease-label probe on E_i, the claimed disentanglement is unverified.","rationale":"The central claim has two parts: (1) the two encoders actually separate disease-related from disease-irrelevant content, and (2) this separation reduces domain mismatch. The paper's direct evidence for part (1) is Section 3.5, but the probe only tests device/patient labels, not the absence of disease information in E_i. Because Section 2.1's shuffle preserves spectral content and respiratory pathology has strong spectral signatures, E_i can plausibly encode disease while still being shuffle-invariant. The vCLUB loss does not fix this: it only enforces independence between E_i and E_R, not absence of disease in E_i. A linear disease probe on frozen E_i embeddings is the minimal experiment that would settle whether the disentanglement is real. The existing ablation (Table 2) shows both losses improve accuracy, but that is compatible with E_i leaking disease information. The missing cross-domain evaluation is a separate concern targeting part (2); part (1) is more foundational and is exactly where the paper's own evidence is incomplete. The reader's weakest_assumption identifies the same point, and the recommended CONDITIONAL verdict remains appropriate pending this test.","tokens_in":7567,"tokens_out":5488,"duration_ms":66056,"concrete_test":"Freeze the trained disease-irrelevant encoder E_i and train a linear probe on its embeddings to classify the four ICBHI labels (normal, crackle, wheeze, both), using the same train/test split as Table 1. If E_i accuracy is substantially above chance (e.g., AS > 40%), then time-shuffling preserves disease-discriminative content and the encoder is not disease-irrelevant, which would refute the central disentanglement claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is Section 2.1: time-shuffling the spectrogram \"disrupts temporal dependencies while preserving spectral content,\" and the Siamese loss (Section 2.3) makes the disease-irrelevant encoder E_i invariant to this shuffle. This is insecure because adventitious respiratory sounds have strong static spectral signatures: wheeze is a near-stationary tonal component and crackles have characteristic spectral envelopes, so shuffled spectrograms still contain disease-discriminative information. The only disentanglement evidence is Section 3.5 / Fig. 2, which linear-probes E_i and E_R on device and patient labels; it shows E_i captures domain labels but never measures whether E_i can classify the respiratory disease labels. If E_i also classifies disease above chance, then it is not disease-irrelevant, the claimed separation is incomplete, and the mutual-information loss is discarding shared signal rather than isolating pure domain content. Patient labels also correlate with disease state, so high patient accuracy on E_i is not by itself evidence of domain-only encoding.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DDE-MAE, a dual-encoder masked autoencoder for respiratory sound classification. One encoder (E_R) is trained on original spectrograms to capture disease-related features; a second encoder (E_I) is trained with a Siamese loss to be invariant to random time-shuffling of the spectrogram, with the intent of capturing disease-irrelevant static content. A vCLUB-based mutual information loss encourages the two encoders' embeddings to be independent. The method is evaluated on the ICBHI dataset under the official random 60/40 split. On the primary metric AS (average of sensitivity and specificity), the full system reaches 61.50, versus 59.93 for reconstruction-only fine-tuning (S2) and 57.59 for direct classification fine-tuning (B2). An ablation shows that removing either the Siamese loss or the MI loss degrades performance. The qualitative analysis in Fig. 2 probes both encoders on device and patient labels, reporting that E_I captures these domain labels more strongly than E_R.","tokens_in":7711,"tokens_out":3299,"duration_ms":40928,"significance":"If the disentanglement claim were fully validated, the method would be a useful self-supervised contribution to respiratory sound classification under domain shift, since it does not require domain labels. The experimental design is generally sound: the comparison S3 versus S2 isolates the effect of the disentanglement losses from additional reconstruction training, and both losses appear to contribute. The reported improvement is modest but consistent with confidence intervals. However, the central claim in the abstract and introduction—that the method reduces domain mismatch and improves domain adaptation—is not actually tested by the experimental protocol, and the disentanglement evidence is incomplete. These are load-bearing gaps rather than presentation issues.","major_comments":[{"comment":"The core premise that time-shuffling the spectrogram preserves only disease-irrelevant content is not verified. Section 2.1 states that shuffling 'disrupts temporal dependencies while preserving spectral content,' but adventitious respiratory sounds have strong static spectral signatures: wheezes are near-stationary tonal components and crackles have characteristic spectral envelopes. A shuffled spectrogram can therefore still contain disease-discriminative information. Making E_I invariant to shuffling via the Siamese loss does not guarantee that E_I is disease-irrelevant. The only disentanglement evidence is Fig. 2, which linear-probes E_I and E_R on device and patient labels. It never probes E_I on the respiratory disease labels. If E_I classifies disease above chance, the claimed separation is incomplete, and the vCLUB MI loss may be discarding shared signal rather than removing domain-only content. In addition, patient labels correlate with disease state, so high patient accuracy on E_I is not by itself evidence of domain-only encoding. A disease-label linear probe on E_I (and on the shuffled-input representation) is needed to support the disentanglement claim.","section":"Section 2.1 and Section 3.5"},{"comment":"The paper's central claim of reducing domain mismatch and enabling generalization to unseen domains is not tested. All experiments use the official random 60/40 split of ICBHI described in Section 3.1, so train and test recordings likely share stethoscope types, patients, and recording environments. The abstract and introduction motivate the work by domain mismatch from different stethoscopes, patient demographics, and environments, and Section 3.6 argues the method does not need domain labels; however, no experiment evaluates the method on unseen stethoscopes, unseen patients, or unseen recording conditions. A cross-domain evaluation, such as leave-one-stethoscope-out or a patient-disjoint split, is required to substantiate the domain-adaptation claim. Without it, the reported improvements could reflect better in-distribution representation learning rather than domain-mismatch reduction.","section":"Section 3.1 and Section 3.6"},{"comment":"The state-of-the-art comparison is reported without confidence intervals or significance tests for the baseline methods, and the interpretation leans heavily on sensitivity. DDE-MAE has the highest SE (53.69) but the lowest or near-lowest SP (69.32) among the top-tier methods, with AS (61.50) within the range of AST + patch-mix CL (62.37) and SG-SCL (61.71). Since the primary metric is the average of SE and SP, a high-SE/low-SP operating point is not necessarily better; a sensitivity-focused conclusion needs a stated clinical or evaluation rationale and ideally a paired significance test or confidence intervals for the baselines.","section":"Section 3.6 and Table 3"}],"minor_comments":[{"comment":"Table 2 contains two rows both labeled '- MI Loss' (69.11 and 67.61). The second row likely denotes removal of both losses or a different configuration; the label should be corrected so the ablation is unambiguous.","section":"Table 2"},{"comment":"The text says 'the disease-related representation Z_R and the disease-relevant representation Z_I'; for consistency with the rest of the paper, Z_I should be called disease-irrelevant.","section":"Section 2.2, after Eq. (3)"},{"comment":"The heading 'Traning details' contains a typo; it should be 'Training details'.","section":"Section 3.3"},{"comment":"The paper does not specify the spectrogram parameters (FFT size, hop length, window function, number of mel bands) or the MAE masking ratio and patch size used in the experiments. These details are needed for reproducibility.","section":"Section 3.3 and Section 2.3"},{"comment":"The sentence 'the comparisons between B1 and B2, as well as between S1 and S3 reveal that pretraining plays a pivotal role' is supported by the table, but the magnitude of improvement is stated for S1 over B1 and S3 over B2; the corresponding confidence intervals overlap for some metrics, so the language should be softened to avoid overstating significance.","section":"Section 3.4, after Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically reasonable as a short empirical paper, but the abstract and introduction overclaim domain-adaptation performance that the experimental protocol does not measure. The main technical risk is the unvalidated assumption that time-shuffle invariance implies disease-irrelevance; without a disease-label probe on E_I, the disentanglement story is not established. I would recommend major revision with a request for either a cross-domain evaluation or a substantial reframing of the claims, plus the disease-label probe on E_I."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this paper is a well-assembled combination of known self-supervised parts, and the ablation work is honest, but the claimed payoff—reducing domain mismatch—is not measured by the experiments. All numbers come from the standard random 60/40 split of ICBHI, so there is no evidence about unseen stethoscopes or patients.\n\nWhat is genuinely new: a dual-encoder MAE for respiratory sounds, where one encoder is trained to be invariant to time-shuffling via a Siamese loss and the two encoders are decorrelated with vCLUB. That exact combination hasn't appeared before. The ablations are the strongest part: Table 1 compares against reconstruction-only fine-tuning (S2) and shows a real but modest gain (AS 61.50 vs 59.93), and Table 2 shows both losses contribute, with the MI loss doing more work. The SOTA comparison is fair—they note that AST+patch-mix CL and SG-SCL beat them on AS, and their 'best sensitivity' claim holds. No circularity in the fitting.\n\nThe soft spot is the load-bearing assumption in Section 2.1. Time-shuffling a spectrogram preserves spectral content, and adventitious sounds like wheezes have strong static spectral signatures, so an encoder made invariant to shuffling is not necessarily disease-irrelevant. The paper's own disentanglement evidence (Fig. 2) only probes device and patient classification accuracy; it never asks whether E_i can classify disease. Since patient labels correlate with disease state, high patient accuracy on E_i is not proof of domain-only encoding. Section 3.5 honestly admits E_R is still somewhat influenced by device/patient, which is good but undercuts the 'disentangled' claim. Also missing: code and a few training details.\n\nWho should read it: anyone working on self-supervised learning for medical audio. It deserves a serious referee—the comparison and ablations are useful—but a referee should require a cross-domain experiment (unseen stethoscope or patient split) and a disease-label probe of E_i before accepting the domain-adaptation narrative. I would engage with it: send to review, expect major revision.","headline":"A plausible small win on ICBHI from combining Audio-MAE with Siamese and vCLUB losses, but the paper's central domain-mismatch claim is not tested by its random-split experiments.","tokens_in":8275,"tokens_out":2779,"would_cite":false,"duration_ms":30934,"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 dual-encoder masked autoencoder that separates disease-related from disease-irrelevant sound features improves respiratory sound classification and reduces the need for domain labels.","keywords":["respiratory sound classification","feature disentanglement","masked autoencoder","self-supervised learning","domain adaptation","ICBHI dataset","mutual information minimization","Siamese loss"],"falsifier":"A direct test is to train only the disease-irrelevant encoder on the same ICBHI classes and measure its classification accuracy after disentanglement: if it stays far above chance, then time shuffling did not strip out disease information. A quantitative version is to estimate the mutual information between the time-shuffled spectrogram and the disease label; if that estimate is not near zero, the assumed separation fails.","tokens_in":7320,"feed_emoji":"🫁","tokens_out":7593,"duration_ms":82231,"temperature":0.7,"pith_summary":"This paper proposes a modified masked autoencoder, DDE-MAE, for classifying respiratory sounds, and argues that forcing the model to split its representation into two independent streams improves accuracy when recordings come from many stethoscopes, patients, and environments. The disease-related encoder works on the original spectrogram, while the disease-irrelevant encoder is trained to be invariant to random shuffling of the time axis, so it must rely on static, background-like content. A mutual-information penalty keeps the two streams independent. On the ICBHI respiratory sound benchmark, the full model reaches an average score of 61.50%, beating reconstruction-only fine-tuning by 1.57 points and reaching the highest sensitivity (53.69%) among the compared methods. This matters because the method promises domain adaptation without any labels for device, patient, or environment.","feed_headline":"Dual-encoder lung sound model tops sensitivity without domain labels","feed_subtitle":"Its disentangled features reach the highest sensitivity among compared methods on the ICBHI benchmark, with no domain labels.","key_machinery":"The central mechanism is a dual-encoder masked autoencoder in which one encoder ($E_r$) learns disease-related content from the intact spectrogram and the other ($E_i$) learns disease-irrelevant content by seeing both the original and a time-shuffled version of the spectrogram. Time shuffling randomly reorders the time axis, destroying temporal dependencies while keeping spectral content, and a Siamese loss forces $E_i$ to produce similar embeddings for the original and shuffled inputs. Independence between the two streams is enforced by minimizing the vCLUB upper bound on their mutual information. Both encoders feed one decoder that reconstructs masked spectrogram patches, so the representation pair must jointly retain enough information to rebuild the input. After this disentanglement stage, only the disease-related encoder is fine-tuned for the classification task.","core_discovery":"The paper's discovery claim is that feature disentanglement—rather than extra reconstruction training or domain supervision—is what drives the improvement. In their ablation, the full dual-encoder setup with both Siamese and mutual-information losses reaches an average score of 61.50%, against 59.93% for the same model fine-tuned only to reconstruct before classifying. The ablation also shows both losses contribute: removing the mutual-information loss drops the average score to 61.30%, and removing the Siamese loss drops it to 60.14%. The authors additionally report that the disease-irrelevant encoder, when probed with a linear classifier, predicts device and patient labels far better than the disease-related encoder does, while the disease-related encoder remains relatively focused on disease cues. This is presented as evidence that the two streams have indeed separated confounding factors from diagnostic content.","pith_inferences":["The same time-shuffling invariance could be transferred to other temporal audio diagnostics, such as heart sounds, bowel sounds, or sleep sounds, where the distinction between transient events and steady background also separates disease from recording conditions, though that transfer is untested here.","The premise that temporal shuffling preserves non-disease content is likely to be only partially true: crackles are transient, so some disease information is temporal, and a natural alternative design might pair time shuffling with frequency masking to define the irrelevant stream.","The device and patient probe result implies a practical diagnostic use the paper does not develop: the disease-irrelevant encoder could serve as a domain fingerprint for flagging out-of-distribution recordings.","Because the gap between direct training and AudioSet pretraining is large, the marginal value of disentanglement may depend on the pretrained initialization, and could shrink or grow under other pretraining recipes."],"forward_implications":["If the claim holds, respiratory sound classifiers can be made more robust to new stethoscopes and recording environments without collecting any domain labels, because the confounding factors are absorbed by the disease-irrelevant encoder.","The disentanglement stage adds a real gain beyond reconstruction: the full setup beats reconstruction-only fine-tuning by 1.57 points in average score, so the separation mechanism, not just extra data exposure, is responsible for the improvement.","The method reaches the highest reported sensitivity among compared approaches (53.69%), which matters for screening because missing abnormal breath sounds is costlier than over-flagging.","The disease-related encoder's residual sensitivity to device and patient factors, acknowledged by the authors, suggests that further refinement of the separation could yield still larger gains.","Since the approach needs no domain labels and no extra modalities, it can be applied directly to existing respiratory sound collections, including small ones, starting from a pretrained Audio-MAE."],"supporting_citations":[{"why":"Supplies the masked autoencoder encoder-decoder structure that DDE-MAE modifies by duplicating the encoder.","marker":"[11]"},{"why":"Provides Audio-MAE, the pretrained model whose encoder is duplicated and fine-tuned in both training stages.","marker":"[15]"},{"why":"AudioSet is the pretraining corpus for Audio-MAE; the benefit of this pretraining is quantified in the B2 versus B1 comparison.","marker":"[16]"},{"why":"Provides the vCLUB upper bound used to estimate and minimize mutual information between the two encoder embeddings.","marker":"[12]"},{"why":"Supplies the Siamese loss that enforces invariance of the disease-irrelevant encoder to time shuffling.","marker":"[13]"},{"why":"The ICBHI dataset and its official 60/40 split are the benchmark on which all reported metrics are measured.","marker":"[10]"}],"fun_headline_variants":["Disentangling dual encoders lifts lung sound accuracy","Separating disease from device noise aids lung audio AI","Dual-encoder disentanglement beats single-loss training","No domain labels needed as disentanglement wins lung benchmark","Lung audio model gains from splitting disease and domain cues"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that randomly shuffling the order of time frames in the spectrogram wipes out the information a doctor would use to hear disease—temporal patterns like crackles and wheezes—while leaving the static spectral content of devices and patients intact, so a network forced to ignore the shuffle learns only non-disease features.","fun_headline_variants_meta":{"raw":{"variants":["Disentangling dual encoders lifts lung sound accuracy","Separating disease from device noise aids lung audio AI","Dual-encoder disentanglement beats single-loss training","No domain labels needed as disentanglement wins lung benchmark","Lung audio model gains from splitting disease and domain cues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000846,"raw_usage":{"total_tokens":3627,"prompt_tokens":835,"completion_tokens":2792,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":2715}},"tokens_in":451,"tokens_out":2792,"duration_ms":24500,"temperature":1.0,"reasoning_tokens":2715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:20:04.388342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to train only the disease-irrelevant encoder on the same ICBHI classes and measure its classification accuracy after disentanglement: if it stays far above chance, then time shuffling did not strip out disease information. A quantitative version is to estimate the mutual information between the time-shuffled spectrogram and the disease label; if that estimate is not near zero, the assumed separation fails.","supporting_citations":[{"cited_title":"Multi-view spectro- gram transformer for respiratory sound classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the masked autoencoder encoder-decoder structure that DDE-MAE modifies by duplicating the encoder."},{"cited_title":"A respiratory sound database for the development of auto- mated classification,","cited_arxiv_id":null,"evidence_quote":"Provides Audio-MAE, the pretrained model whose encoder is duplicated and fine-tuned in both training stages."},{"cited_title":"Masked autoencoders are scalable vision learners,","cited_arxiv_id":null,"evidence_quote":"AudioSet is the pretraining corpus for Audio-MAE; the benefit of this pretraining is quantified in the B2 versus B1 comparison."},{"cited_title":"Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classifi- cation,","cited_arxiv_id":null,"evidence_quote":"The ICBHI dataset and its official 60/40 split are the benchmark on which all reported metrics are measured."}],"review_version":1}