{"id":"d69484c8-2057-442b-8be1-de23f91fbafc","arxiv_id":"2509.01153","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A GNN plus anchor-interval framework detects abnormal respiratory sound events, reaching F1 of 22.3 percent on SPRSound, with anchor scales tuned to the observed event durations.","lead":"This paper builds a graph neural network with anchor intervals to detect abnormal respiratory sounds in audio, handling variable-length clips by turning spectrogram chunks into graph nodes. It reports F1 of 22.3 percent on SPRSound versus 6.81 percent for a CNN baseline, but the comparison is not run under identical evaluation settings and the anchor scales are chosen to match the test data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3's claimed improvement over [42] is not interpretable: the two systems are evaluated on different label sets (normal events included vs. excluded), so the F1 gap may be a metric artifact.","rationale":"The paper's central claim is empirical: the proposed framework improves respiratory sound event detection over prior work and natively handles variable-length audio. The strongest evidence is Table 3, but that comparison is invalid as reported. Section 3.3.2 explicitly states that the baseline [42] includes normal sounds in the detection/evaluation process while the proposed method excludes them. This changes the event-based metric in Eq. (47): Nref, deletions, and insertions are computed over different label sets, so the F1 gap is not a fair measure of architectural advantage. For the HF Lung v1 dataset, no baseline is provided, so no improvement is demonstrated there. The reader's rationale mentions this baseline mismatch, but selects the anchor-scale duration distribution as the weakest assumption. I agree the anchor-scale choice (Sec. 3.3.3, d=[0.5,0.8,1.5]) is a genuine generalization concern because the scales are aligned with the observed duration distribution in the evaluation datasets, and this can overfit the benchmark. However, that concern affects robustness to new datasets and does not invalidate the reported comparison as directly as the protocol mismatch does. The released reference implementation is a positive, but code availability does not remedy the evaluation flaw. Thus, the verdict should remain at the reader's REJECT; my concern strengthens that conclusion rather than moving it.","tokens_in":20842,"tokens_out":4796,"duration_ms":55434,"concrete_test":"Re-run CNN+VAD [42] on SPRSound under the identical protocol used for Ours: remove all 'normal' events from the ground-truth set (abnormal-only Nref), apply the same 200 ms onset / 200 ms-or-10% offset collars, and compute F1/ER with the same event matching. If [42]'s F1 rises from 6.81 to a level close to 22.3, the claimed improvement is an evaluation artifact; if it remains below, the comparison is partially rescued.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that EZhouNet improves abnormal respiratory sound event detection rests on Table 3, where Ours achieves F1=22.3 / ER=1.08 on SPRSound vs. CNN+VAD [42] at F1=6.81 / ER=1.26. The paper states in Sec. 3.3.2 that [42] 'also include normal sounds in the detection process, while our experiments exclude them.' This changes the definition of the event-based metrics in Eq. (47): N (number of events in annotation) is different, and a system that never outputs 'normal' is not charged with false positives or deletions for normal events, while [42] is. The comparison is therefore not apples-to-apples; the 15.5-point F1 gap cannot be attributed to the architecture. For HF Lung v1, no baseline is reported at all, so the second row of Table 3 provides no comparative evidence. Consequently, the empirical support for the central claim is absent until the baseline is re-evaluated under the same protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EZhouNet, an end-to-end framework for abnormal respiratory sound event detection that combines multichannel spectrograms (Mel, Gamma, CQT), a graph construction module that converts spectrogram chunks into variable-length graph nodes, two GAT layers for node feature update, and an anchor-interval refinement module with three predefined duration scales. The anchor intervals are refined via GRU-based local feature extraction, soft offset regression over discrete bins, and separate confidence/classification heads. Experiments are reported on SPRSound 2024 and HF Lung V1, including a comparison against a CNN+VAD baseline [42], a node-level ablation for respiratory position information, and a discussion of limitations. The central claim is that the framework improves temporal localization and event-level detection of abnormal respiratory sounds while natively handling variable-length audio.","tokens_in":21171,"tokens_out":4118,"duration_ms":51075,"significance":"If the central claim were adequately supported, the paper would make a useful contribution by transferring anchor-based object detection ideas to respiratory sound event detection, avoiding frame-level post-processing, and proposing a graph-based representation that can in principle handle variable-length recordings. The availability of reference code is a positive feature, and the task itself is clinically relevant and relatively understudied. However, the reported empirical evidence is not yet sufficient to establish the claimed improvement: the comparison with [42] is not controlled, the anchor priors appear to be fitted to the evaluation datasets, key hyperparameters are missing, and no uncertainty estimates or significance tests are provided. These issues are fixable, but they are load-bearing for the paper's main conclusion.","major_comments":[{"comment":"The comparison with CNN+VAD [42] on SPRSound is not apples-to-apples. The text states that [42] also includes normal sounds in the detection process, while 'our experiments exclude them.' In the event-based metrics of Eq. (47), N is the number of reference events, and FP/FN are counted with respect to the label set. A system that never predicts 'normal' cannot be charged with deletions or insertions for normal events, whereas a baseline that must detect them is penalized for every normal-event error. The reported F1 gap (22.3 vs 6.81) and ER gap (1.08 vs 1.26) may therefore be a metric artifact, not an architectural improvement. To support the central claim, the baseline must be re-evaluated under the identical label protocol, or both systems must be reported under both protocols (including class-wise matched labels).","section":"3.3.2, Table 3, Eq. (47)"},{"comment":"The anchor interval priors appear to be fitted to the evaluation datasets. The paper observes that event durations in SPRSound and HF Lung V1 fall at 0.5s, 0.75s, 1s, and 1.25s, and states that these 'align with the designed prior anchor intervals of 0.5s, 0.75s, and 1.5s' (Eq. (25)). Since the anchor durations and densities are chosen after inspecting the test duration distribution, the reported detection performance partly reflects this fitting. This weakens the generalization claim and makes the framework's interval-detection behavior dependent on the evaluation datasets. I recommend either evaluating on a held-out dataset with a different event-duration distribution, or performing a sensitivity analysis over anchor scales/configurations and reporting results for a prior chosen independently of the test data.","section":"3.3.3, Eq. (25)-(26)"},{"comment":"Several load-bearing experimental details are missing. The loss weights w_n_conf, w_n_cls, w_i_conf, w_i_cls, w_i_loc in Eq. (46) are never given numerical values. The IoU threshold theta_iou in Eq. (31)-(32) is defined but its value is never reported. The bin counts B_k and their ranges in Eq. (38)-(39) are also unspecified, and the temporal encoding scale gamma in Eq. (24) is only described as 'e.g. 0.05.' In addition, Tables 1-3 report single runs without standard deviations, confidence intervals, or number of seeds. Given the moderate F1 values and class imbalance, the reported differences (e.g., integrated vs separate head) cannot be assessed for statistical significance. These details are needed both for reproducibility and for interpreting the central empirical claim.","section":"3.3.1-3.3.2, Eq. (31)-(32), Eq. (38)-(39), Eq. (46)"},{"comment":"The ablation for respiratory position information is performed at the node level, not at the event level. The text claims that position information improves recall for all four abnormal classes, and indeed Table 5 shows recall increases for rhonchi, wheeze, stridor, and crackle. However, F1 for crackle decreases (0.7905 to 0.7804), and there are no confidence intervals or significance tests. More importantly, the central claim is about event-level detection (Table 3); a node-level ablation does not directly establish that position information improves the event-level detection of abnormal respiratory sounds. An event-level ablation with and without position information is needed.","section":"3.3.2, Tables 4-5"}],"minor_comments":[{"comment":"Typo in section title: 'stduy' should be 'study.'","section":"3.3.2 (title)"},{"comment":"The argmax notation contains stray vertical bars ('|||{...}|||') and the tie-breaking rule could be stated more cleanly. Please clarify whether ties are resolved by the first-appearing class in chronological order.","section":"Eq. (9)"},{"comment":"The text refers to 'Fig. 2' when describing the network architecture, but the first mentioned figure is 'Fig. 1'; please check figure numbering. Also, Fig. 4 in the text corresponds to 'The convergence of interval classification losses' but is placed after the evaluation metrics section; renumbering may improve readability.","section":"Figures 1-3"},{"comment":"The reference list appears to be duplicated: the same entries [1]-[9] and then a second full 'References' list with the same numbered papers are printed. Please remove the duplicate block.","section":"References"},{"comment":"For the HF Lung V1 row, no baseline value is reported. The table should clearly state 'not comparable / not reported' rather than a dash, to avoid implying an empty or zero baseline.","section":"Table 3"},{"comment":"The discussion admits that variable-length audio could not be batched in practice due to the PyTorch Lightning collate_fn, and training used sequential loading. This is an honest limitation, but it should also be noted in Section 1 and the abstract contribution statement, since 'handling variable-length audio' is presented as a native capability.","section":"3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a preprint submitted to Elsevier, and the reference list is accidentally duplicated in the arXiv version; this should be corrected before any archival publication. The more substantive concern is that the anchor priors appear to be explicitly matched to the test duration distribution, and the only external comparison uses a different label protocol. These issues can be addressed by re-running the baseline under a matched protocol, reporting all hyperparameters and uncertainty estimates, and adding a robustness experiment on an unseen duration distribution. If those additions are made, the paper could become a solid contribution; in the current form, the central claim is not yet empirically established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real engineering contribution—GNN over chunked spectrograms plus anchor-interval refinement is new for respiratory sound event detection, and they ship code—but the empirical case as written doesn't carry the headline. The main comparison in Table 3 is confounded, and the anchor design is fitted to the evaluation distribution. I'd still send it to a serious referee, because the architecture is worth engaging with and the flaws are fixable.\n\nWhat's good: The framework is coherent: nodes from time-chunked multi-spectrogram features, GAT updates, then YOLO-style anchors with soft IoU labels and offset bin regression. The paper does real ablations (integrated vs separate heads, edge-attention variants, offset resolution) and reports class-wise results on two datasets. It is also unusually honest: Sec 3.3.2 openly states the baseline [42] includes normal events while their experiments exclude them, and Sec 3.4 admits variable-length inputs could not be batched in the implementation. That transparency helps a reviewer.\n\nThe soft spots are exactly where the reader put them. The Table 3 claim is not interpretable: with different normal-event handling, the event-based F1 and ER are computed over different N and different FP/FN definitions. The 15.5-point F1 gap cannot be attributed to the architecture. HF Lung v1 has no baseline row at all. Second, Sec 3.3.3 says the observed event durations 'align with the designed prior anchor intervals'—the anchors are effectively tuned to the test distribution, so calling them priors is generous. Third, the abstract's variable-length claim is undermined by the implementation note; the graph structure is variable-length in principle, but the actual training loads sequentially. Add to that the missing error bars and the unreported loss weights/IoU threshold, and the reported numbers are single-run and hard to gauge. The absolute F1 values (22.3 and 16.8) are low, though the authors acknowledge this.\n\nNone of this is fatal to the underlying idea. The combination is novel and could be a decent baseline for future event-level respiratory sound detection. But the paper as submitted does not establish its claimed improvement. It needs a re-run against [42] under an identical protocol, a fixed anchor design justified without test-set peeking, and full hyperparameter reporting.\n\nFor you: worth a reading-group discussion on evaluation pitfalls in medical audio event detection, but I wouldn't cite it yet. For the editor: send to peer review; a good reviewer will ask for exactly the re-evaluation above.","headline":"Novel GNN+anchor framework for respiratory event detection, but the evaluation as written doesn't prove the headline improvement—the main baseline comparison is confounded and the anchor priors are fitted to the test set.","tokens_in":21608,"tokens_out":2893,"would_cite":false,"duration_ms":33648,"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":"Anchor intervals give lung-sound detection direct event timing","keywords":["respiratory sound event detection","graph neural network","anchor interval","variable-length audio","SPRSound","HF Lung V1","auscultation","sound event detection"],"falsifier":"Run the same framework on a dataset of respiratory recordings whose abnormal events are predominantly shorter than 0.3 s or longer than 2.5 s, with all other training details unchanged. If event-level F1 drops sharply relative to a version with anchor scales matched to those durations, the reliance on fixed anchor priors is confirmed.","tokens_in":20764,"feed_emoji":"🩺","tokens_out":6801,"duration_ms":72715,"temperature":0.7,"pith_summary":"This paper claims that respiratory sound event detection can be cast as an end-to-end interval-refinement problem instead of frame-level classification plus post-processing. It builds a graph from a multi-channel spectrogram, where each node is a short chunk of audio and edges connect neighboring chunks in time, then refines predefined anchor intervals to output event boundaries directly. Because the graph can contain any number of nodes, the framework accepts variable-length recordings without resampling or padding. On the SPRSound 2024 and HF Lung V1 datasets, the method reports event-level F1 scores of 22.3% and 16.8%, outperforming the compared CNN+VAD baseline on SPRSound, and the authors show that adding respiratory position information raises recall across abnormal sound classes. A sympathetic reader would care because direct boundary learning and variable-length input address two known limitations in respiratory sound event detection.","feed_headline":"Anchor intervals give lung-sound detection direct event timing","feed_subtitle":"An end-to-end graph network localizes wheeze, crackle, stridor, and rhonchi in variable-length audio.","key_machinery":"Anchor interval refinement is the load-bearing mechanism: a fixed set of candidate intervals at three scales, with soft IoU-based labels for confidence, class, and target interval, lets the network regress start/end offsets instead of predicting event boundaries from scratch. The supporting machinery is the spectrogram-to-graph construction, in which each group of five spectrogram frames becomes a node, edges form a directed temporal chain, and GAT layers update node features while preserving variable-length structure.","core_discovery":"The paper's central claim is that separating sound-event detection into interval confidence, interval category, and interval-boundary regression over predefined anchor intervals lets a network learn event timing directly, rather than inferring boundaries from frame-level outputs. The anchor intervals are temporal windows at three fixed durations (0.5, 0.8, and 1.5 seconds normalized to clip length), each labeled by its best IoU with a ground-truth event; the network predicts an offset, a confidence, and a class per anchor and decodes refined intervals. The graph representation is the enabler of variable-length handling: every five spectrogram frames form one node, nodes are linked in a forwa","pith_inferences":["Because the anchor scales are fixed to the duration distribution observed in SPRSound and HF Lung v1, the framework should transfer best to settings with similar event durations; new populations with systematically longer or shorter abnormal events would likely require re-tuning the scales.","The same anchor-interval idea could be applied to other periodic or event-based bioacoustic tasks, such as cough or snore detection, where event durations are variable and frame-level post-processing is the norm.","A testable extension is to replace fixed anchor scales with learned or adaptive anchors, such as a small set of learnable durations per clip, which would remove the manual prior without abandoning the regression formulation.","The chain-graph connectivity limits long-range context; using higher-order edges or a second graph over longer spans might improve detection of sustained wheeze and rhonchi."],"forward_implications":["Event-level respiratory sound detection can be trained end-to-end, since interval confidence, category, and localization are learned jointly from anchor labels.","Variable-length recordings can be processed without fixed-length constraints, because the input graph naturally grows with the number of chunks.","Adding respiratory position information improves recall of abnormal events, so location cues are worth including in clinical audio models.","Five-frame grouping favors discontinuous sounds such as crackle; continuous sounds such as wheeze, rhonchi, and stridor need larger groups or longer context.","The reported event F1 remains modest in absolute terms, indicating that anchor priors and frame grouping are not yet sufficient for clinical use alone."],"supporting_citations":[{"why":"Supplies the SPRSound pediatric dataset and gold-standard event annotations used for the main experiments.","marker":"[43]"},{"why":"Supplies the HF Lung v1 dataset and the CNN+BiLSTM approach that serves as the generalization comparison.","marker":"[40]"},{"why":"Provides the CNN+VAD baseline compared on SPRSound and the two-stage detection pipeline the paper argues against.","marker":"[42]"},{"why":"Introduces a YOLO-like anchor design for audio segmentation, the direct source of the anchor-interval idea.","marker":"[31]"},{"why":"Presents region-based convolutional recurrent networks for audio event detection, linking region or anchor proposals to SED.","marker":"[32]"},{"why":"Defines the event-based F1 and error-rate metrics used to evaluate detection.","marker":"[34]"},{"why":"Introduces graph-based learning for abnormal lung sound classification, motivating the GNN representation.","marker":"[37]"}],"fun_headline_variants":["Anchor intervals deliver direct timing for lung-sound events","Graph-based network pinpoints wheeze and crackle events","Variable-length audio handled with graph and anchor intervals","Position clues sharpen abnormal lung sound detection","End-to-end event localization for respiratory sounds"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The framework depends on the three fixed anchor durations (0.5, 0.8, and 1.5 seconds) matching the durations of abnormal events in the target data; if a new dataset's events fall mostly outside these ranges, the offset regressor must compensate on its own.","fun_headline_variants_meta":{"raw":{"variants":["Anchor intervals deliver direct timing for lung-sound events","Graph-based network pinpoints wheeze and crackle events","Variable-length audio handled with graph and anchor intervals","Position clues sharpen abnormal lung sound detection","End-to-end event localization for respiratory sounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00035,"raw_usage":{"total_tokens":1747,"prompt_tokens":743,"completion_tokens":1004,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":932}},"tokens_in":487,"tokens_out":1004,"duration_ms":11433,"temperature":1.0,"reasoning_tokens":932,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:49:22.123836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same framework on a dataset of respiratory recordings whose abnormal events are predominantly shorter than 0.3 s or longer than 2.5 s, with all other training details unchanged. If event-level F1 drops sharply relative to a version with anchor scales matched to those durations, the reliance on fixed anchor priors is confirmed.","supporting_citations":[{"cited_title":"Meta: Data compression and event detection grand challenge 2024 with sprsound dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the SPRSound pediatric dataset and gold-standard event annotations used for the main experiments."},{"cited_title":"Benchmarking of eight recurrent neural network variants for breath phase and adventitious sound detection on a self-developed open- access lung sound database—hf_lung_v1,","cited_arxiv_id":null,"evidence_quote":"Supplies the HF Lung v1 dataset and the CNN+BiLSTM approach that serves as the generalization comparison."},{"cited_title":"A novel ai- inspired method and system implementation for detecting and clas- sifyingpediatricrespiratorysoundevents,","cited_arxiv_id":null,"evidence_quote":"Provides the CNN+VAD baseline compared on SPRSound and the two-stage detection pipeline the paper argues against."},{"cited_title":"You only hear once: a yolo-like algorithm for audio segmentation and sound event detec- tion,","cited_arxiv_id":null,"evidence_quote":"Introduces a YOLO-like anchor design for audio segmentation, the direct source of the anchor-interval idea."},{"cited_title":"R-crnn: Region-based convolutional recurrent neural network for audio event detection,","cited_arxiv_id":null,"evidence_quote":"Presents region-based convolutional recurrent networks for audio event detection, linking region or anchor proposals to SED."},{"cited_title":"Graph-based semi- supervised one class support vector machine for detecting abnor- mal lung sounds,","cited_arxiv_id":null,"evidence_quote":"Introduces graph-based learning for abnormal lung sound classification, motivating the GNN representation."}],"review_version":1}