{"id":"35c0a9c3-b538-4457-ba5c-407735640561","arxiv_id":"2507.20052","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN with temporal self-attention and automatic frequency band selection improves respiratory sound classification accuracy and cuts computation by about half on public benchmark datasets.","lead":"This paper proposes a compact neural network for classifying respiratory sounds that combines a CNN with temporal self-attention and an automatic frequency band selection module that removes noisy spectral regions. The method reports state-of-the-art accuracy on two public lung-sound benchmarks while roughly halving computational cost, which matters for building real-time and low-resource diagnostic tools.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA comparisons are not yet like-for-like: SPRSound 2023 training data is unspecified, the age-specific ICBHI AS is an unweighted subgroup average, and the Patch-Mix baseline is not reproduced.","rationale":"The reader's conditional verdict is the right one. The internal ablations (TSA placement, band elimination, lambda sensitivity) are coherent and suggest the method has real merit, and the FLOP reduction at 50% frequency retention is a concrete, checkable benefit. The load-bearing weakness is not the method itself but the benchmark comparisons: the paper relies on cross-paper numbers without reproducing baselines, leaves the SPRSound 2023 training protocol undefined, and reports age-specific ICBHI results as an unweighted average of adult and child models rather than a pooled test-set metric. These are protocol gaps, not fraud or internal contradiction. One focused audit of the training set, the age-specific metric, and the Patch-Mix baseline would settle whether the claimed state-of-the-art results are real or an artifact of comparison choices. Therefore the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":19303,"tokens_out":10071,"duration_ms":121571,"concrete_test":"Conduct one protocol audit: rerun the SPRSound 2023 rows of Table IV using the exact training set prescribed by the BioCAS 2023 challenge rules, recompute the age-specific ICBHI rows of Table III as a single pooled AS over the full test set, and reproduce the Patch-Mix CL baseline in the authors' own preprocessing pipeline before adding FBS. If the SPRSound 2023 training protocol differs from the official one, or the pooled age-specific AS drops materially, or the reproduced baseline is not 62.37 within noise, the headline SOTA and benchmark claims are not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline 'new benchmarks/SOTA' requires that every row in Tables III and IV is a like-for-like comparison on the same test set under the same training protocol. That condition is not met for at least three reported entries. First, Section IV.A.2 describes SPRSound 2023 as 'test-only recordings' with 871 recordings, while Section IV.D says the authors 'used the official train-test splits provided with each dataset'; no training set for the 2023 rows is actually specified. If the 2023 models were trained on SPRSound 2022 splits, or on any other set, the Table IV improvements (66.66 to 71.20 TS on Task 1; 80.97 to 83.20 on Task 2) are not controlled and the 'new benchmarks' claim is unverified. Second, Section IV.B states that for age-specific models 'the reported performance metric is the average of both models.' The Table III age-specific Task-1 AS of 62.68 is exactly the mean of the adult 58.03 and child 67.32 scores from Table II, not a score computed on the full ICBHI test set. All comparator rows are pooled test-set AS values, so this row is not comparable even if the adult/child split were balanced. Third, the transformer 'new record' 64.01 on ICBHI is compared against the published 62.37 for Patch-Mix CL; the paper gives no in-house reproduction of that baseline. The FBS-specific gain could therefore be partly a preprocessing or evaluation artifact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a compact CNN-Temporal Self-Attention (CNN-TSA) architecture for respiratory sound classification, which combines a CNN6 backbone with temporal self-attention applied after frequency aggregation, and an Importance-Based Frequency Band Selection (FBS) module that uses Grad-CAM (or Integrated Gradients for transformer backbones) to iteratively remove low-importance mel bands. It also introduces age-specific adult/child models for the ICBHI dataset. The system is evaluated on ICBHI 2017, SPRSound 2022, and SPRSound 2023 for binary and 4-class tasks, reporting state-of-the-art or benchmark results and up to 50% FLOPs reduction.","tokens_in":19569,"tokens_out":5680,"duration_ms":65074,"significance":"If the claims hold, FBS is a practical, model-agnostic input-regularization technique that improves both accuracy and efficiency, and the compact CNN-TSA would be attractive for deployment in resource-constrained settings. The paper has several strengths: held-out test evaluation, 5-fold cross-validation for band selection, means over five runs with standard deviations, and detailed ablations of TSA placement, number of retained bands, and the lambda hyperparameter. However, the headline claims of new benchmarks and state-of-the-art performance depend on comparisons that are not yet like-for-like, which weakens the central contribution until those comparisons are controlled.","major_comments":[{"comment":"The SPRSound 2023 training set is never specified. Section IV.A.2 describes SPRSound 2023 as '871 test-only recordings' with no training split, while Section IV.D states that the authors 'used the official train-test splits provided with each dataset.' Since SPRSound 2023 has no official training split, it is unclear what data the SPRSound 2023 rows in Table IV were trained on. If the models were trained on SPRSound 2022 splits or any other set, the reported improvements from 66.66 to 71.20 TS (Task 1) and 80.97 to 83.20 TS (Task 2) are not controlled against the cited baselines, which may have used different training data. The authors must specify the exact training set used for SPRSound 2023 and, if necessary, re-run the cited baselines under the same training protocol to support the 'new benchmark' claim.","section":"Section IV.A.2 and IV.D"},{"comment":"For age-specific models, Section IV.B states that 'the reported performance metric is the average of both models.' The age-specific Task-1 AS of 62.68% in Table III is exactly the mean of the adult 58.03% and child 67.32% scores from Table II, not a score computed on the full ICBHI test set. All comparator rows in Table III report pooled test-set AS values, so this row is not directly comparable, and the statement that the age-specific model 'matches' Patch-Mix CL (62.37%) is an artifact of the unweighted subgroup average. The authors should report a pooled AS computed on the full ICBHI test set for the age-specific models, or perform per-subgroup comparisons against comparators split by the same age threshold.","section":"Section IV.B and Table III"},{"comment":"The 'new record' of 64.01% for Patch-Mix CL + FBS(IS) is compared against the published Patch-Mix CL score of 62.37% from Bae et al. [17] with no in-house reproduction under the authors' preprocessing (16 kHz resampling, 64 mel bands, 50–2000 Hz range, SpecAugment). Because FBS modifies the input representation and the baseline score is taken from the original paper, the improvement could be due to preprocessing or evaluation-protocol differences rather than to FBS. Provide a same-protocol Patch-Mix CL baseline without FBS and report the FBS gain relative to that reproduced baseline.","section":"Section V.B and Table III"},{"comment":"The hyperparameter λ in Equation (7) is central to the FBS importance score, but the paper does not state how λ was selected for the final models in Tables III and IV, nor whether the values shown in Figure 3 were computed on the test sets. The ablation shows that multiclass performance peaks at λ = 0.5–0.6 and binary performance improves up to λ = 1.0. If λ was chosen based on test-set performance, the reported gains over the full-spectrum baseline are inflated by selection on the test set. Specify the λ values used in each final model and describe the selection procedure (e.g., inner cross-validation on the training folds).","section":"Section III.B.1 and Figure 3"}],"minor_comments":[{"comment":"The age threshold used to split the ICBHI dataset into adult and child models is not reported anywhere; without this threshold the age-specific experiments cannot be reproduced.","section":"Section IV.B"},{"comment":"In the SPRSound 2023 Task 2 block, there are two rows labeled 'CNN-TSA +FBS(BS) 50%f req (Ours)'; the second row almost certainly denotes 75% frequency retention and should be relabeled.","section":"Table IV"},{"comment":"The Grad-CAM formula omits the ReLU activation, and the text says this is deliberate, but the authors should justify why the omission is appropriate and whether the negative attributions have a clear interpretation in the frequency-selection context.","section":"Equation (3)"},{"comment":"The stopping criterion 'no degradation in CV performance' is vague; specify which metric is monitored, what tolerance is allowed, and how many bands were ultimately retained for each dataset in the final models.","section":"Algorithm 1"},{"comment":"Preprocessing details for the transformer adaptation are not given; the paper only says the pipeline from [17] was adopted. Provide the number of mel bands, normalization, and any augmentation used for the Patch-Mix CL + FBS experiments so that they can be reproduced.","section":"Section IV.C"},{"comment":"The description of the three-layer CNN6 variant for SPRSound lacks the channel sizes, kernel sizes, and pooling configurations; include a layer-by-layer table or a precise architectural specification.","section":"Section III.A"}],"recommendation":"major_revision","confidential_remarks":"The empirical methodology is largely sound, but the headline contribution--setting new benchmarks and achieving SOTA--rests on at least three uncontrolled comparisons: the SPRSound 2023 training data ambiguity, the non-comparable age-specific AS aggregation, and the absence of an in-house Patch-Mix CL baseline. These are fixable with additional experiments or by softening the claims. No code or public reproducibility artifact is mentioned; requesting one would strengthen a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, useful empirical paper with a genuinely new frequency-band-selection trick, but the headline SOTA claims are slightly ahead of the evidence because three of the comparison rows aren't like-for-like.\n\nThe FBS module is the real contribution. Using Grad-CAM (or Integrated Gradients for transformers) to score each mel band, then iteratively pruning the least informative bands, is a simple idea that seems to work: it improves accuracy while cutting FLOPs roughly in half. The authors validate it on their own compact CNN-TSA and on a published AST-based baseline, which makes the method feel portable. The ablation on TSA placement is well done, and the evaluation is mostly honest: held-out test sets, 5-fold CV for the selection step, means over five runs.\n\nThe soft spots are all in the SOTA comparison, not in the core method. First, SPRSound 2023 is described as 'test-only recordings,' but the paper never says what training data was used for those rows. That makes the 71.20 and 83.20 figures impossible to verify. Second, the age-specific ICBHI AS is computed as the unweighted average of the adult and child models, not as a pooled test-set metric. Table III compares that number directly against pooled AS values from other papers, which is apples to oranges. Third, the 'new record' on ICBHI via FBS + Patch-Mix CL is a comparison against the published 62.37 baseline with no in-house reproduction, so part of the gain could be preprocessing or evaluation protocol. The lambda and r hyperparameters are also under-specified, but that's minor.\n\nNone of these flaws sink the paper. The central claim—that attribution-guided band pruning helps and saves compute—is plausible and probably robust. But the authors should either run controlled baselines or soften the benchmark language. A serious referee should ask for the SPRSound 2023 training protocol, a pooled age-specific metric, and ideally one reproduced baseline.\n\nFor anyone working on respiratory sound classification or efficient audio models, this is worth reading. It deserves peer review; with those clarifications it could be a solid accept.","headline":"Useful FBS trick with honest held-out evaluation, but the SOTA claims overreach on three comparison rows.","tokens_in":20142,"tokens_out":2771,"would_cite":true,"duration_ms":28598,"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 frequency-selection module for lung-sound AI boosts accuracy on two benchmark datasets while halving computation.","keywords":["respiratory sound classification","frequency band selection","temporal self-attention","CNN","audio spectrogram transformer","Grad-CAM","SPRSound","ICBHI"],"falsifier":"Retrain the strongest compared baselines (e.g., Patch-Mix CL and the proposed CNN-TSA with all 64 bands) on the exact training splits and preprocessing used in this paper and check whether the FBS-masked models still outperform them. If the baselines recover their original scores but the masked models do not, the benchmark claim fails. A complementary test would be to apply the frequency masks learned on one dataset to an unseen respiratory sound dataset and see whether the accuracy gain transfers.","tokens_in":19055,"feed_emoji":"🫁","tokens_out":5892,"duration_ms":62550,"temperature":0.7,"pith_summary":"This paper tries to show that respiratory sound classification can be made both more accurate and far cheaper by teaching the model to ignore useless frequency bands. It proposes a compact CNN with a temporal self-attention module (CNN-TSA) and an importance-based frequency band selection (FBS) routine that removes noisy Mel-spectrogram bands before classification. The authors claim this combination sets new state-of-the-art results on the SPRSound-2022/2023 pediatric datasets and competitive or best results on ICBHI-2017, while reducing FLOPs by up to 50%. If true, the FBS module is a practical, architecture-agnostic drop-in that helps bring automated lung-sound analysis to real-time, resource-limited settings.","feed_headline":"Frequency selection boosts lung-sound AI while halving compute","feed_subtitle":"A compact CNN with attention beats larger models on respiratory benchmarks by keeping only the informative frequency bands.","key_machinery":"The load-bearing mechanism is the importance-based frequency band selection (FBS) module, which scores each of the 64 Mel-frequency bands with $I_f = \\mathrm{Mean}[f] - \\lambda\\,\\mathrm{MaxDiff}[f]$, where $\\mathrm{Mean}[f]$ is the band's average Grad-CAM attribution across classes and $\\mathrm{MaxDiff}[f]$ is the largest inter-class attribution gap. The module iteratively removes the four lowest-scoring bands, retrains with 5-fold cross-validation, and stops when performance degrades, yielding a binary frequency mask. This mask is applied to the input spectrogram, so the downstream CNN-TSA network processes only the informative bands; the temporal self-attention operates on a frequency-aggregated representation to capture long-range temporal dependencies at low cost.","core_discovery":"The paper's central claim is that a model-driven importance score for each frequency band, computed from Grad-CAM attributions and penalized by inter-class inconsistency, can identify a smaller subset of Mel-spectrogram bands that improves classification accuracy while cutting computation roughly in half. Applied to the authors' CNN-TSA network—CNN6 backbone with frequency aggregation and temporal self-attention—this frequency band selection yields new benchmarks on SPRSound-2022 and SPRSound-2023 and state-of-the-art results on ICBHI-2017, using only 1.1M–4.6M parameters. The same FBS procedure, using Integrated Gradients instead of Grad-CAM, is shown to improve a transformer baseline (Patch-Mix with Audio Spectrogram Transformer) to a new ICBHI record. The authors also report that training separate models for adult and pediatric groups improves accuracy on ICBHI.","pith_inferences":["The paper does not specify which training data was used for the SPRSound-2023 evaluation (the dataset is test-only in the challenge), so the reported 'new benchmark' may depend on an unstated training-set choice; a reader should treat that comparison as provisional.","FBS's dependence on Grad-CAM (or Integrated Gradients) means the selected mask inherits the attribution method's biases; comparing masks from different attribution methods on the same data could reveal how sensitive the gains are to that choice.","The same pruning idea could be tested on other physiological sounds, such as heart or bowel sounds, which also have noisy spectra, though the paper does not evaluate this.","The iterative elimination procedure retrains the model many times; a cheaper variant that predicts band importance from a single pass could make the method more practical for large datasets."],"forward_implications":["Respiratory sound classification can run in real time on embedded hardware: the best models use 1.1M–4.6M parameters and about half the FLOPs of their unfiltered counterparts.","Frequency band selection acts as a form of input regularization, so other audio-classification tasks with noisy spectra may benefit from attribution-guided band pruning.","The FBS module improves both CNN and transformer backbones, making it a drop-in enhancement for existing respiratory sound classifiers.","Age-specific training offers a simple accuracy gain on datasets spanning children and adults, without changing the architecture.","Retaining only 50–75% of Mel bands is consistently enough for top performance, indicating that much of the spectrogram is redundant or harmful."],"supporting_citations":[{"why":"Provides the ICBHI 2017 dataset and official train/test split used for all ICBHI experiments.","marker":"[28]"},{"why":"Provides the SPRSound-2022 dataset and defines the tasks and metrics for the pediatric benchmarks.","marker":"[29]"},{"why":"Supplies the CNN6 backbone architecture and AudioSet pretrained weights that the CNN-TSA network is built on.","marker":"[47]"},{"why":"Grad-CAM is the attribution method used to compute per-band importance scores for the proposed FBS.","marker":"[51]"},{"why":"The Patch-Mix CL transformer baseline that FBS is adapted to, and the state-of-the-art comparison that FBS improves.","marker":"[17]"},{"why":"Integrated Gradients is used instead of Grad-CAM when FBS is applied to the transformer model.","marker":"[52]"},{"why":"Audio Spectrogram Transformer is the backbone of the transformer baseline used for the FBS drop-in experiment.","marker":"[19]"}],"fun_headline_variants":["Frequency selection halves compute and improves lung-sound AI","Lung-sound AI: frequency band selection cuts FLOPs in half","Compact CNN with attention and frequency selection beats larger models","Frequency filter cuts lung-sound AI compute in half","Frequency selection: half the FLOPs, better lung-sound AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of new benchmarks assumes that the compared prior results were obtained under the same training data, preprocessing, and evaluation protocol; if protocol differences drive the gains, the headline improvement would not be attributable to the method.","fun_headline_variants_meta":{"raw":{"variants":["Frequency selection halves compute and improves lung-sound AI","Lung-sound AI: frequency band selection cuts FLOPs in half","Compact CNN with attention and frequency selection beats larger models","Frequency filter cuts lung-sound AI compute in half","Frequency selection: half the FLOPs, better lung-sound AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001479,"raw_usage":{"total_tokens":5946,"prompt_tokens":952,"completion_tokens":4994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":4909}},"tokens_in":568,"tokens_out":4994,"duration_ms":36418,"temperature":1.0,"reasoning_tokens":4909,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:50:45.127055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the strongest compared baselines (e.g., Patch-Mix CL and the proposed CNN-TSA with all 64 bands) on the exact training splits and preprocessing used in this paper and check whether the FBS-masked models still outperform them. If the baselines recover their original scores but the masked models do not, the benchmark claim fails. A complementary test would be to apply the frequency masks learned on one dataset to an unseen respiratory sound dataset and see whether the accuracy gain transfers.","supporting_citations":[{"cited_title":"An open access database for the evaluation of respiratory sound classification algorithms,","cited_arxiv_id":null,"evidence_quote":"Provides the ICBHI 2017 dataset and official train/test split used for all ICBHI experiments."},{"cited_title":"Sprsound: Open-source sjtu paediatric respiratory sound database,","cited_arxiv_id":null,"evidence_quote":"Provides the SPRSound-2022 dataset and defines the tasks and metrics for the pediatric benchmarks."},{"cited_title":"Panns: Large-scale pretrained audio neural networks for audio pattern recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN6 backbone architecture and AudioSet pretrained weights that the CNN-TSA network is built on."},{"cited_title":"Grad-cam: Visual explanations from deep networks via gradient-based localization,","cited_arxiv_id":null,"evidence_quote":"Grad-CAM is the attribution method used to compute per-band importance scores for the proposed FBS."},{"cited_title":"Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classification","cited_arxiv_id":null,"evidence_quote":"The Patch-Mix CL transformer baseline that FBS is adapted to, and the state-of-the-art comparison that FBS improves."},{"cited_title":"Axiomatic attribution for deep networks,","cited_arxiv_id":null,"evidence_quote":"Integrated Gradients is used instead of Grad-CAM when FBS is applied to the transformer model."},{"cited_title":"Ast: Audio spectrogram transformer,","cited_arxiv_id":null,"evidence_quote":"Audio Spectrogram Transformer is the backbone of the transformer baseline used for the FBS drop-in experiment."}],"review_version":1}