{"id":"04c97020-3263-4d46-90e8-8c0e1aedecc2","arxiv_id":"2504.17156","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A waveform-CNN plus spectrogram-transformer plus Bi-GRU fusion reports state-of-the-art sensitivity (90.3%) and total score (93.6%) for binary abnormal versus normal respiratory sound classification on SPRSound.","lead":"Breathing-sound AI that reads both the raw sound wave and the usual frequency picture of the audio scored 90.3% sensitivity and 93.6% total score on a pediatric lung sound dataset. The authors argue that combining the two views with a recurrent network improves detection of abnormal breath sounds like wheezes and crackles.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is unverifiable because the paper never states which SPRSound test split produced Table II, and the binary pooled-event metrics in Eqs. (2)-(6) can be dominated by Wheeze/Fine Crackle, so \"distinguishes pathological classes\" is not demonstrated.","rationale":"I agree with the reader that the evaluation protocol is the weakest point of the paper, and my review sharpens the concern to two concrete, internal evidence gaps: the unspecified test split and the micro-averaged binary sensitivity metric. These are load-bearing because the paper's own Section IV-A labels Testing-2 as the more robust generalization split, and the abstract/conclusion explicitly claim the model can distinguish pathological respiratory sound classes in the plural. The reported metrics only support binary abnormal-versus-normal detection, and even that is only as strong as the unspecified split. Because the authors do not report which split produced Table II, do not report per-class results, and do not provide uncertainty estimates, the central SOTA claim does not follow from the presented evidence. The reader's REJECT verdict is therefore appropriate, and no adjustment is needed. I also note a secondary internal inconsistency in Eq. (1), where p and y appear to be swapped relative to standard focal loss, but the split/per-class issue is the primary barrier to accepting the central claim.","tokens_in":6859,"tokens_out":6560,"duration_ms":61406,"concrete_test":"Recompute the evaluation on both official SPRSound splits, reporting per-class sensitivity/specificity for all seven classes and the pooled Eqs. (2)-(6) metrics, with WLANN and the strongest cited baseline (TRespNET) run under identical preprocessing and event pooling. If the 93.6% TS appears only on Testing-1, or if the pooled sensitivity is driven by Wheeze/Fine Crackle while rare-class sensitivity is near zero, the SOTA and class-discrimination claims fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that Table II's 90.3% SE and 93.6% TS are measured on a protocol that supports a fair SOTA comparison and supports the plural \"pathological respiratory sound classes.\" Two unstated protocol choices break this.\n\nFirst, the paper describes two official test splits in Table I (intra-patient Testing-1 and inter-patient Testing-2) but never says which split produced Table II. By the authors' own description, Testing-2 is the more robust generalization split, yet all headline numbers could come from Testing-1, where patients overlap the training set. Without this specification, the comparison with TRespNET and other cited methods is not interpretable.\n\nSecond, even under the stated metrics (Section IV-C), Eq. (2) micro-averages over all abnormal events, and Table I shows Wheeze and Fine Crackle dominate the abnormal pool: in Testing-2 they are 305+80 of 389 abnormal events (99%), and in Testing-1 they are 108+175 of 316 (89.6%). A model that detects only these two frequent classes can report high sensitivity while failing on Rhonchi, Stridor, Coarse Crackle, and Wheeze&Crackle. No per-class sensitivity or confusion matrix is reported, so the claim that the model \"distinguishes pathological respiratory sound classes\" is not supported by the evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WLANN, a neural architecture for respiratory sound classification that fuses features from a waveform-processing 1D-CNN and an Audio Spectrogram Transformer (AST), then applies a Bi-GRU for frame-level context modeling. The model is trained with multi-class focal loss and evaluated on the SPRSound pediatric respiratory dataset. The authors report 90.3% sensitivity and 93.6% total score, claiming state-of-the-art performance and the ability to distinguish pathological respiratory sound classes. An ablation study in Table III shows that each architectural component contributes to the reported performance.","tokens_in":7195,"tokens_out":4056,"duration_ms":38486,"significance":"If the reported results are reproducible and the comparison is fair, WLANN would be a meaningful contribution to respiratory sound classification: the fusion of waveform and spectrogram features with recurrent context modeling is a sensible design, and the ablation in Table III credibly attributes gains to each component. The use of the public SPRSound dataset and the comparison with several published methods are also strengths. However, the significance is currently limited by unverifiable protocol details: it is not stated which test split produced the headline numbers, the metrics pool all abnormal classes into a single binary positive class, and no per-class results are reported, so the central claims of state-of-the-art performance and multiclass discrimination are not yet established.","major_comments":[{"comment":"The manuscript does not state which SPRSound test split produced the results in Table II. Table I defines two official splits, Testing-1 (intra-patient) and Testing-2 (inter-patient), and the authors themselves describe Testing-2 as the more robust generalization split. Because the two splits have very different patient overlap and class distributions, the headline values of 90.3% sensitivity and 93.6% total score are not interpretable without specifying the split, and the comparison with TRespNET and other cited methods is not a controlled comparison unless all methods were evaluated on the same split with the same preprocessing. Please state the split for every table and report results on both splits.","section":"Section IV-D, Table II"},{"comment":"The evaluation metrics micro-average over pooled abnormal events, so a model that detects only the frequent classes could achieve high sensitivity while failing on rare pathological classes. From Table I, Wheeze and Fine Crackle together constitute 99% of abnormal events in Testing-2 (305+80 of 389) and 89.6% in Testing-1 (108+175 of 316). The abstract and conclusion claim that the model distinguishes pathological respiratory sound classes, but the paper reports no per-class sensitivity, specificity, or confusion matrix. The multiclass discrimination claim is therefore not supported by the provided evidence. Please add per-class results and, if appropriate, macro-averaged metrics.","section":"Section IV-C, Eqs. (2)-(6)"},{"comment":"The comparison with previous methods lacks essential protocol details. It is not stated whether the cited numbers in Table II are taken from the original publications, re-implemented by the authors, or evaluated on the same test split with the same event-level pooling and the same preprocessing chain. Without this information, the state-of-the-art claim is not verifiable. Please specify the origin of each compared result and provide a reproducible evaluation protocol, ideally with code or a detailed description of how each baseline was run.","section":"Section IV-D, Table II"},{"comment":"Equation (1) does not appear to be a correct formulation of multiclass focal loss. With y_i denoting the predicted probability and p_i denoting the true distribution, the term (1-y_i)^gamma * p_i * log(y_i) would down-weight confident predictions and multiply by the target probability, which is not the standard focal loss that down-weights the contribution of easy normal-class examples. If this is a typesetting artifact, please provide the correct formula and define each variable precisely; if the loss is intentionally different, explain the design and its effect.","section":"Section III-D, Eq. (1)"},{"comment":"The manuscript omits standard training details needed to reproduce the results, including the optimizer, learning rate schedule, batch size, number of epochs, validation strategy, random seeds, and the number of model parameters. Since the paper makes a state-of-the-art claim on a public benchmark, these details are necessary for reproducibility and for assessing whether the comparison with baselines is controlled.","section":"Section IV-B and IV-D"}],"minor_comments":[{"comment":"The affiliation text contains a typo: 'Mateials' should be 'Materials'.","section":"Author affiliation"},{"comment":"The heading 'D. Waveform-Logmel Audio Neural Networks' duplicates the title of Section III-C; this section appears to be about the loss function and should be renamed accordingly.","section":"Section III-D heading"},{"comment":"The caption reads 'combined alone channel axis' and should read 'combined along the channel axis'.","section":"Figure 3 caption"},{"comment":"The metric equations (2)-(6) are badly garbled in the current typeset version (e.g., the numerator and denominator symbols appear corrupted). Please ensure that the formulas for sensitivity, specificity, average score, harmonic score, and total score are typeset correctly.","section":"Section IV-C"},{"comment":"The paper uses inconsistent notation for sensitivity: 'SE' in the abstract, 'Sen' in the tables, and 'SN' in the equations. Please standardize.","section":"Throughout"},{"comment":"References [12] and [27] refer to the same paper by Chen et al. (BioCAS 2022) and should be merged or distinguished.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core architecture and ablation are reasonable, and the paper may become a solid contribution if the authors disclose the test split, add per-class and split-level results, and clarify the baseline comparison protocol. In its current state, the state-of-the-art claim is not verifiable, which is the main reason I recommend major revision rather than acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a competent engineering paper with a reasonable architecture, a consistent ablation, and a headline claim that the text does not actually support. The likely useful parts are reproducible; the SOTA claim needs the missing evaluation details before it means anything.\n\nThe genuinely new piece is narrow. Dual-route waveform plus log-mel input is already in TRespNET [30], and CRNN-style recurrent context is standard. What WLANN adds is a specific fusion: a 1D-CNN waveform stream joined along the channel axis with an AST spectrogram stream, then fed to a Bi-GRU. That is a plausible combination, and the ablation in Table III is the paper's strongest asset—it shows each component earns its keep, from waveform-CNN alone at 43.1% sensitivity up to the full model at 90.3%, with the fusion and the GRU both contributing. The write-up is clear enough to reproduce: resampling to 16 kHz, 40-850 Hz Butterworth, 16x16 patches with overlap 8, focal loss gamma 2.0. That counts for something.\n\nThe soft spots are real and mostly in the evaluation. First, the paper describes two official SPRSound test splits in Table I but never says which one produced Table II. Testing-1 shares patients with the training set; Testing-2 does not. Without that sentence, the comparison against TRespNET and the rest is not interpretable. Second, the headline metrics are binary. Equations (2)-(6) micro-average over pooled abnormal events, and the class distribution means Wheeze plus Fine Crackle are 89.6% of abnormal events in Testing-1 and 99% in Testing-2. A model that nails those two can look strong while missing Rhonchi, Stridor, and Coarse Crackle entirely. No per-class sensitivity, no confusion matrix, so the abstract's wording about 'distinguishing pathological respiratory sound classes' is not supported. Third, the comparison table omits the very AST-based SOTA methods the introduction cites as the best current systems; Table II stops at TRespNET. That weakens 'outperforming previous studies' to 'outperforming a selected subset.'\n\nNothing in my reading suggests the method is broken. The problems are reporting problems, and each is fixable: state the split, give per-class numbers, add the missing baselines. As-is, the central claim fails on the evidence presented.\n\nThe paper is for people building medical audio systems who want a concrete recipe for fusing waveform and spectrogram features with a recurrent context layer, and it is a good teaching example of why evaluation protocol details matter. It deserves peer review rather than desk rejection, because it is reproducible on a public dataset and the gaps are addressable. I would want those revisions before I would take the SOTA claim at face value.","headline":"A competent architecture-plus-ablation paper whose SOTA claim is not verifiable as written: the test split is unnamed and the binary metrics can be carried by just two frequent classes.","tokens_in":7676,"tokens_out":5631,"would_cite":false,"duration_ms":48741,"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":"This paper claims that a dual-input network reading both raw waveform and log-mel spectrogram, fused and context-modeled by a bidirectional GRU, achieves top published results on the SPRSound pediatric respiratory dataset.","keywords":["respiratory sound classification","waveform-logmel audio neural networks","Audio Spectrogram Transformer","Bi-GRU","multi-class focal loss","SPRSound dataset","computer-assisted auscultation"],"falsifier":"Run WLANN under the paper's training recipe on the SPRSound inter-patient split and report per-class sensitivity for Rhonchi, Stridor, Coarse Crackle, Fine Crackle, and Wheeze; the central claim would fail if the inter-patient total score falls below the 90.9% DenseNet169 baseline in the paper's Table II, or if rare classes show near-zero sensitivity.","tokens_in":6653,"feed_emoji":"🩺","tokens_out":8280,"duration_ms":69644,"temperature":0.7,"pith_summary":"Respiratory sounds are hard to classify automatically because abnormal events are scarce and because a single input representation loses information: waveforms lack frequency detail, and log-mel spectrograms lose fine structure at high frequencies and have low temporal resolution. This paper proposes WLANN, a network that reads both the raw waveform and the log-mel spectrogram, fuses the two feature streams, and uses a bidirectional GRU to model how frames evolve over time. On the SPRSound pediatric dataset the model reaches 90.3% sensitivity and a 93.6% total score, which the authors report as the best among the published methods they compare with. If the result holds, automated auscultation moves closer to being a usable clinical aid for detecting abnormal breath sounds like crackles and wheezes.","feed_headline":"Respiratory-sound model hits 90.3% sensitivity, 93.6% total score","feed_subtitle":"Fusing waveform and log-mel with Bi-GRU context beats prior respiratory sound classifiers.","key_machinery":"The load-bearing object is the fusion tensor $C \\in \\mathbb{R}^{F \\times T \\times (C+C/F)}$: the waveform branch's 1D-CNN output is reshaped into $F$ frequency bins and concatenated with the AST branch's $C$-channel spectrogram output along the channel axis. This fused frequency–time representation is mean-pooled over frequency, then modeled by a bidirectional gated recurrent unit (Bi-GRU), a network that reads the frame sequence forward and backward; a linear layer with sigmoid activation maps to class labels. Multi-class focal loss with $\\gamma=2$ down-weights the abundant normal class so training focuses on abnormal sounds. The argument is carried by this particular fusion: waveform features supply time-domain detail the spectrogram loses, the spectrogram supplies frequency structure the waveform lacks, and Bi-GRU supplies the frame-to-frame context that makes abnormal events recognizable.","core_discovery":"The central claim is that a dual-branch architecture — a 1D-CNN reading the raw waveform and an Audio Spectrogram Transformer reading the log-mel spectrogram — whose outputs are concatenated along the channel axis and passed through a Bi-GRU, classifies SPRSound respiratory events better than any previously published system the authors compare against. The reported numbers are 90.3% sensitivity, 96.9% specificity, and 93.6% total score, versus 90.9% total score for the strongest prior entry in the comparison table. Ablation results in the paper show that the fused input outperforms either branch alone, and that adding Bi-GRU raises sensitivity from 86.1% to 90.3%, supporting the paper's conclusion that frame-level temporal context is what makes abnormal sounds detectable.","pith_inferences":["The paper reports only pooled abnormal-event sensitivity; a per-class breakdown would clarify whether the 90.3% reflects genuine detection of rare classes such as Stridor and Rhonchi or is dominated by the abundant Fine Crackle and Wheeze events.","Because the 40–850 Hz Butterworth filter removes content above 850 Hz, the waveform branch may be blind to high-frequency wheeze harmonics; testing a wider passband could improve or clarify the fusion's contribution.","The same waveform-plus-spectrogram-plus-Bi-GRU recipe is domain-agnostic and could transfer to other auscultation signals such as heart or bowel sounds, where abnormal events are also rare and class-imbalanced.","If the reported numbers come from the intra-patient split, the clinically relevant next result is inter-patient performance; that number would decide whether the gain generalizes to patients not seen in training."],"forward_implications":["A model with the reported sensitivity and specificity balance — 90.3% sensitivity at 96.9% specificity — would detect abnormal respiratory events in about nine out of ten cases while keeping false alarms low.","The ablation table shows that the full WLANN beats both single-branch models: fused waveform plus spectrogram without Bi-GRU scores 91.2% total score, and adding Bi-GRU raises it to 93.6%.","Multi-class focal loss with $\\gamma=2$ is part of the recipe; it down-weights normal sounds so that training concentrates on the scarce abnormal classes.","The architecture is assembled from publicly available components (AST, 1D-CNN, Bi-GRU) and evaluated on an open dataset, so the comparison can be replicated and extended.","If the result holds, automated stethoscope analysis can move from simply identifying normal breathing to reliably flagging crackles and wheezes, the step needed for clinical use."],"supporting_citations":[{"why":"Supplies the Audio Spectrogram Transformer branch that processes log-mel spectrogram patches, the model's second input stream.","marker":"[7]"},{"why":"Defines the vision-transformer patch embedding and positional-encoding design that AST inherits for spectrogram patches.","marker":"[16]"},{"why":"Supplies the multi-class focal loss used to counter the dataset's class imbalance.","marker":"[19]"},{"why":"Provides the SPRSound dataset, its seven event classes, and the official intra-patient and inter-patient train-test splits used in the evaluation.","marker":"[20]"},{"why":"Supplies SpecAugment time-warping and frequency-masking augmentation used to train the AST branch.","marker":"[24]"},{"why":"Is the closest prior dual-input baseline (spectrogram plus raw series with attention) that the paper's state-of-the-art claim must beat.","marker":"[30]"}],"fun_headline_variants":["Dual-input network beats prior respiratory sound classifiers","Fusing waveform and log-mel lifts respiratory sound scores","Bi-GRU context pushes respiratory sound sensitivity to 90.3%","Waveform + log-mel fusion tops respiratory classification","New audio net outperforms on respiratory sound detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline numbers assume that all methods in the comparison table were evaluated on the same SPRSound test split with the same preprocessing, but the paper does not state which split produced its 90.3% sensitivity and 93.6% total score.","fun_headline_variants_meta":{"raw":{"variants":["Dual-input network beats prior respiratory sound classifiers","Fusing waveform and log-mel lifts respiratory sound scores","Bi-GRU context pushes respiratory sound sensitivity to 90.3%","Waveform + log-mel fusion tops respiratory classification","New audio net outperforms on respiratory sound detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001101,"raw_usage":{"total_tokens":4552,"prompt_tokens":862,"completion_tokens":3690,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":3611}},"tokens_in":478,"tokens_out":3690,"duration_ms":22443,"temperature":1.0,"reasoning_tokens":3611,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:47:52.993968+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run WLANN under the paper's training recipe on the SPRSound inter-patient split and report per-class sensitivity for Rhonchi, Stridor, Coarse Crackle, Fine Crackle, and Wheeze; the central claim would fail if the inter-patient total score falls below the 90.9% DenseNet169 baseline in the paper's Table II, or if rare classes show near-zero sensitivity.","supporting_citations":[{"cited_title":"Automatic Speech and Voice Disorder Detection using Deep Learning-A Systematic Literature Review","cited_arxiv_id":null,"evidence_quote":"Supplies the Audio Spectrogram Transformer branch that processes log-mel spectrogram patches, the model's second input stream."},{"cited_title":"Transformer -based approach to pathology diagnosis using audio spectrogram","cited_arxiv_id":null,"evidence_quote":"Defines the vision-transformer patch embedding and positional-encoding design that AST inherits for spectrogram patches."},{"cited_title":"HCRNNIDS: Hybrid convolu tional recurrent neural network -based network intrusion detection system","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-class focal loss used to counter the dataset's class imbalance."},{"cited_title":"Automatic adventitious respiratory sound analysis: A systematic review","cited_arxiv_id":null,"evidence_quote":"Supplies SpecAugment time-warping and frequency-masking augmentation used to train the AST branch."},{"cited_title":"An effective lung sound classification system for respiratory disease diagnosis using densenet cnn model with sound pre - processing engine","cited_arxiv_id":null,"evidence_quote":"Is the closest prior dual-input baseline (spectrogram plus raw series with attention) that the paper's state-of-the-art claim must beat."}],"review_version":1}