{"id":"a0241a63-55ff-49c3-b498-526b7d708c12","arxiv_id":"2506.01148","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A bandit-weighted cross-attention fusion of codec and spectral audio features is reported to improve heart murmur classification in the paper's own experiments.","lead":"This paper introduces BAOMI, a fusion method that combines neural audio codec representations with spectral features for heart murmur classification, using a bandit mechanism to weight attention heads. The authors report improved accuracy on the CirCor dataset, though the evaluation lacks error bars and a direct state-of-the-art comparison.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central empirical claim depends on the unstated unit of five-fold splitting in §4.1: if folds are recording-level rather than patient-level, repeated recordings from the same 963 patients leak across train/test, inflating the reported BAOMI gains and the SOTA claim.","rationale":"The reader identified the same load-bearing concern: the evaluation section does not state whether the five-fold cross-validation is stratified by patient. This is the most direct threat to the central empirical claim because all headline results are built on those folds. If the same patient's multiple recordings appear in both training and test, the model can exploit patient identity, and the reported improvements from BAOMI over individual features and cross-attention would not reflect generalization to new patients. The paper also has secondary weaknesses: the bandit mechanism in Section 3.1 is under-specified (Eq. 5 requires a per-head loss but no per-head classifier is described, and no exploration strategy is given), and the SOTA claim is argued indirectly by comparing to published MFCC results rather than running prior methods on the same split. However, those issues affect reproducibility and the strength of the SOTA statement, whereas the split-unit problem can invalidate the entire experimental comparison. The fix is straightforward: re-run with patient-stratified folds and report per-fold results. For that reason, the appropriate disposition remains CONDITIONAL rather than outright rejection, and my read does not change the reader's verdict.","tokens_in":8531,"tokens_out":3246,"duration_ms":39603,"concrete_test":"Download the released code and reproduce Table 2 with patient-level GroupKFold (all recordings of a patient in the same fold) or the official CirCor patient-disjoint split, using identical hyperparameters. Inspect the per-fold DAC+MFCC scores: if accuracy or macro-F1 drops materially relative to MFCC-only and cross-attention baselines under patient-disjoint folds, the claimed fusion benefit is inflated by leakage; if performance is preserved, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 reports five-fold cross-validation over 3,163 recordings from 963 patients, but never states whether folds are split by patient or by recording. Because each patient contributes multiple PCG recordings and murmur status is a patient-level attribute, recording-level folds place same-patient recordings in both training and test sets. The model can then memorize patient-specific recording characteristics rather than learn murmur-relevant generalization. Every headline number in Table 2, including the best BAOMI DAC+MFCC result (Acc 89.93%, MA-F1 79.37%), is an average over such folds. The paper also reports no per-fold scores or error bars, so it is impossible to tell whether the reported advantage over MFCC alone (80.90/73.28) and over the cross-attention baseline (89.81/75.06) is stable or due to a single favorable fold. Because the central claim is an empirical superiority claim, this missing split specification is the load-bearing weakness: if the folds are not patient-stratified, the comparison is not about classification of new patients and the SOTA statement is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BAOMI, a bandit-based cross-attention mechanism for fusing neural audio codec representations (NACRs such as EnCodec, DAC, SNAC, SpeechTokenizer) with spectral features (MFCC, LFCC) for heart murmur classification on the CirCor DigiScope dataset. The mechanism learns per-head weights via Q-values updated from loss-reduction rewards, and the fused representation is used for classification. The authors report that BAOMI outperforms individual features and a standard multi-head cross-attention baseline across many combinations, with the best result from DAC+MFCC (accuracy 89.93%, macro F1 79.37%, weighted F1 89.67%). They claim to establish a new state of the art for heart murmur classification.","tokens_in":1542,"tokens_out":2071,"duration_ms":73351,"significance":"If the reported results are reliable, the paper offers a useful fusion strategy that combines complementary acoustic representations, and it makes code and checkpoints publicly available. The idea of using bandit-based head weighting in cross-attention is interesting and could transfer to other multimodal fusion tasks. However, the significance is currently limited by evaluation weaknesses: the cross-validation split unit is unspecified, no variance measures are reported, and the state-of-the-art claim is argued only indirectly. The contribution is therefore promising but not yet convincingly established.","major_comments":[{"comment":"The unit of five-fold cross-validation is unspecified. The dataset is described as 3,163 recordings from 963 patients, and the task is heart murmur classification, where the murmur label is a patient-level attribute. The text only says 'four folds are used for training and one fold for testing' without stating whether folds are stratified by patient or by recording. If the split is by recording, then multiple recordings from the same patient can appear in both training and test folds, allowing the model to memorize patient-specific recording characteristics and inflate the reported scores in Tables 1 and 2 (e.g., BAOMI DAC+MFCC accuracy 89.93%). Please specify the split unit and, if it is not patient-stratified, repeat the experiments with patient-level folds and report per-fold results.","section":"Section 4.1"},{"comment":"No measure of variance or statistical significance is reported. The tables show only the average over five folds for each metric. For the best combination (DAC+MFCC), the BAOMI accuracy advantage over the cross-attention baseline is only 0.12 percentage points (89.93 vs. 89.81), although the macro-F1 improvement is larger (79.37 vs. 75.06). Without per-fold scores, standard deviations, or a paired test, it is impossible to judge whether the proposed mechanism provides a stable improvement or whether the difference arises from a single favorable fold. Please report per-fold values and error bars or a significance test.","section":"Section 4.2, Tables 1 and 2"},{"comment":"The claim of setting a new state of the art is not directly supported. The authors argue that BAOMI improves over individual spectral features that were reported as state of the art in prior works [23,24,25,9], but they do not compare BAOMI against those specific methods under the same evaluation protocol. Prior works may use different data splits, feature sets, or preprocessing, so the indirect comparison does not establish a new state of the art. Please either implement and evaluate the most relevant prior methods on the same folds, or temper the claim to 'competitive with' or 'better than the individual-feature baselines considered here.'","section":"Section 4.2, final paragraph"},{"comment":"The bandit reward R_h is computed from the reduction of the cross-entropy loss, which is the same objective minimized by the whole network for classification. This makes the head weighting directly optimized toward the training loss, and the paper does not clarify whether ΔL_h is measured on the training split or a held-out split. The risk is that the bandit selects heads that overfit the training loss rather than learn a generalizable fusion. Please clarify the computation of ΔL_h and discuss this self-referential tuning risk. Also, the hyperparameters γ and ε in Eq. (5) are not given in Section 3.1 or in the training details; please state their values.","section":"Equation (5), Section 3.1"}],"minor_comments":[{"comment":"There are several typos: 'a agent' should be 'an agent', 'intially' should be 'initially', and 'calclulated' in Section 3.1 should be 'calculated'. Please proofread throughout.","section":"Abstract and Section 1"},{"comment":"The class distribution '179, 695, and 68 samples' sums to 942, not 963 as stated for the number of patients. Please clarify whether these counts refer to patients or recordings, and reconcile the totals.","section":"Section 4.1"},{"comment":"The phrase 'we keep the modeling details as same as used in the downstream modeling' should be 'we keep the modeling details the same as those used in downstream modeling.' Also, for SpeechTokenizer, 'we add the codes and average it' should be 'average them.'","section":"Section 3.1"},{"comment":"The caption 't-SNE Plots- (a) CNN (MFCC) ...' should use a colon or period instead of a dash, and the figure labels should be formatted consistently.","section":"Figure 2"},{"comment":"The abbreviations NAC and NACR are used interchangeably; please define both at first use and use them consistently thereafter.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The key issue is the unspecified cross-validation split unit; if the current splits are recording-level, the headline results may be invalid. The authors should re-run with patient-stratified folds and report variance. The bandit mechanism itself is interesting, but the state-of-the-art claim needs direct comparison or a more modest phrasing. The paper would be strengthened by providing code runnable with proper splits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the BAOMI heart murmur paper. Short version: it is a reasonable fusion experiment with a clean writeup and a welcome code release, but the headline result is untrustworthy until the cross-validation split is specified. Section 4.1 says five-fold CV on 3,163 recordings from 963 patients and doesn't say whether splits are by patient or recording. Since murmur label is per patient and each patient has multiple PCGs, recording-level splits put the same patient in train and test. That is exactly the setup that inflates accuracy by letting the model memorize patient-specific recording artifacts. The stress-test note gets this right. This is the load-bearing flaw.\n\nWhat's genuinely new: the paper is first in its cited set to combine neural codec representations with spectral features for HMC, and the bandit-style reweighting of attention heads—softmax over Q-values updated by loss reduction—is a small but sensible twist on cross-attention. The experimental scope is a plus: six codecs, two spectral feature sets, many pairs, and the code is public. The BAOMI results are better than the cross-attention baseline in nearly every row of Table 2, which is evidence the weighting does something beyond random variation.\n\nThe soft spots, in order of severity. First, the split problem above. Second, no error bars or per-fold numbers anywhere. Some gains are small (e.g., DAC+MFCC accuracy 89.93 vs 89.81 for cross-attention), and without variance we can't judge whether they are real or a lucky fold. Third, the SOTA claim is indirect: the paper argues that since MFCC alone beats prior reported scores, improving on MFCC means SOTA, but prior works used different protocols, classes, and possibly data splits. There is no direct table comparing BAOMI to published methods under the same evaluation. Fourth, the 'bandit' mechanism is under-specified. Eq. 5 computes reward from per-head loss reduction, but the paper never explains how a single head's loss contribution is computed before head outputs are concatenated. The Q-update is an exponential moving average with no exploration, so calling it a multi-armed bandit is generous. These last points are fixable with clarification and a modest rewrite.\n\nMy recommendation: send it to review. The split issue is potentially fatal, but it is also checkable—ask for patient-stratified five-fold CV with per-fold results and confidence intervals. If the gains survive, the paper is a solid incremental contribution for the audio fusion and heart-sound communities. If they don't, it should be withdrawn or repositioned as a negative result. It deserves referee time either way.","headline":"Plausible fusion recipe for heart murmur classification, but the headline results rest on an unspecified CV split that may leak patients; needs patient-stratified re-evaluation before the SOTA claim can be taken seriously.","tokens_in":9295,"tokens_out":3194,"would_cite":false,"duration_ms":36222,"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":"Fusing neural codec embeddings with spectral features via bandit-weighted cross-attention improves heart murmur classification, with the DAC+MFCC pair reaching 89.93% accuracy and 79.37% macro F1 on the CirCor subset.","keywords":["heart murmur classification","neural audio codec","spectral features","cross-attention","multi-armed bandit","feature fusion","EnCodec","MFCC"],"falsifier":"Re-run the full BAOMI protocol on the CirCor dataset with folds split strictly by patient ID, ensuring no patient appears in both training and test, and check whether the DAC+MFCC result (89.93% accuracy, 79.37% macro F1) and its margin over the cross-attention baseline persist; if the margin shrinks or vanishes, the state-of-the-art claim rests on leakage from non-patient-disjoint splits.","tokens_in":8364,"feed_emoji":"🫀","tokens_out":2795,"duration_ms":27539,"temperature":0.7,"pith_summary":"The paper claims that combining neural audio codec representations (NACRs), such as EnCodec and DAC, with spectral features (MFCC and LFCC) improves heart murmur classification beyond either feature family used alone. It introduces BAOMI, a fusion mechanism in which a multi-armed bandit assigns more weight to the attention heads that reduce classification loss most, so that noisy heads contribute less to the fused representation. On the CirCor DigiScope subset, the DAC+MFCC combination under BAOMI reaches 89.93% accuracy and 79.37% macro F1, the best reported in the paper. The authors argue this sets a new state of the art for heart murmur classification and that it demonstrates complementary behavior between codec-level acoustic detail and spectral energy content.","feed_headline":"Bandit-weighted fusion hits 89.9% heart murmur accuracy","feed_subtitle":"Mixing neural codec and spectral features with head-weighting beats either representation alone.","key_machinery":"The central object is BAOMI, a fusion framework built on a multi-head cross-attention mechanism augmented with a multi-armed bandit. The NACR and spectral features are each passed through a 1D CNN, flattened, and projected into query, key, and value spaces for two cross-attention flows (NACR attending to spectral and spectral attending to NACR). Each attention head maintains a Q-value updated by a reward proportional to how much that head reduces the cross-entropy loss; these Q-values are converted into soft weights that scale the heads before concatenation into the fused representation. This head-weighting step is what distinguishes BAOMI from the standard cross-attention baseline, and it is the mechanism the paper credits for mitigating noise from less informative heads.","core_discovery":"The central claim is that neural audio codec representations and spectral features are complementary for heart murmur classification, and that their fusion is best achieved by a bandit-based cross-attention mechanism that dynamically weights attention heads according to their contribution to loss reduction. On the CirCor subset, BAOMI with DAC and MFCC achieves 89.93% accuracy, 79.37% macro F1, and 89.67% weighted F1, outperforming individual NACRs, individual spectral features, and a standard cross-attention fusion baseline across all tested combinations. The paper further claims that this constitutes a new state of the art for heart murmur classification on the dataset.","pith_inferences":["Editorial inference: The reported gains may depend on the unstated split design; if the five-fold splits are recording-level rather than patient-level, the accuracy advantage could shrink under patient-disjoint evaluation, so the state-of-the-art claim should be re-tested on patient-stratified folds.","Editorial inference: The same bandit-weighted cross-attention recipe could be applied to other codec-spectral pairs (e.g., LFCC with SNAC24) and to other audio-classification domains such as respiratory sound or speech emotion classification, where similar complementary structure exists.","Editorial inference: The paper does not ablate the number of attention heads; a natural testable extension is whether the bandit weighting still helps with 2, 8, or 16 heads, or whether the benefit is specific to the 4-head configuration used here.","Editorial inference: Since NACRs alone underperform spectral features on this dataset, the practical message for practitioners is that codec embeddings add value primarily through fusion, not as a replacement for classical spectral descriptors."],"forward_implications":["Fusing NACRs with spectral features becomes a viable default recipe for heart murmur classification, since the best fused combination outperforms the best individual feature set by roughly 9 percentage points in accuracy and 6 points in macro F1.","The bandit-based head-weighting mechanism should generalize to other feature-fusion and multimodal tasks where some attention heads are expected to be noisy or uninformative.","The performance ordering across codecs (DAC and SNAC24 leading) suggests that codec choice matters and that pre-training data distribution influences transferability to heart sounds.","If the result holds under patient-disjoint validation, the approach provides a practical pipeline for classifying murmur presence from raw PCG audio without hand-crafted segmentation.","The improved t-SNE clustering and confusion matrices indicate the fused representation separates the Present and Absent classes more cleanly than either feature family alone."],"supporting_citations":[{"why":"Supplies the EnCodec architecture used to extract one of the neural audio codec representations.","marker":"[14]"},{"why":"Supplies the DAC model, whose representation yields the best fusion result when paired with MFCC.","marker":"[16]"},{"why":"Provides the CirCor DigiScope dataset used for all experiments.","marker":"[18]"},{"why":"Serves as a prior state-of-the-art method using MFCC features for heart murmur classification.","marker":"[9]"},{"why":"One of the improved-MFCC baselines that BAOMI is compared against for the state-of-the-art claim.","marker":"[23]"},{"why":"Another improved-MFCC baseline used to position the proposed method's results.","marker":"[24]"},{"why":"Deep-residual learning baseline with improved MFCC features used in the SOTA comparison.","marker":"[25]"},{"why":"Establishes the prior use of neural audio codec representations for heart sound classification.","marker":"[11]"},{"why":"Provides a vector cross-attention fusion approach used as a strong baseline for comparison.","marker":"[22]"}],"fun_headline_variants":["Bandit cross-attention fuses codec and spectral features to 89.9% accuracy","89.9% murmur accuracy with bandit-weighted codec-spectral fusion","Adaptive bandit attention fuses audio codecs and MFCC for 89.9% accuracy","Bandit-weighted cross-attention achieves 89.9% heart murmur accuracy","Bandit-based attention fusion yields 89.9% murmur accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The five-fold cross-validation is stated as five folds over 3,163 recordings from 963 patients, but the paper does not say whether all recordings of a single patient are kept within one fold, so if a patient's recordings appear in both training and test sets the reported accuracy gains could be inflated by patient memorization.","fun_headline_variants_meta":{"raw":{"variants":["Bandit cross-attention fuses codec and spectral features to 89.9% accuracy","89.9% murmur accuracy with bandit-weighted codec-spectral fusion","Adaptive bandit attention fuses audio codecs and MFCC for 89.9% accuracy","Bandit-weighted cross-attention achieves 89.9% heart murmur accuracy","Bandit-based attention fusion yields 89.9% murmur accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002055,"raw_usage":{"total_tokens":7946,"prompt_tokens":839,"completion_tokens":7107,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":6999}},"tokens_in":455,"tokens_out":7107,"duration_ms":45910,"temperature":1.0,"reasoning_tokens":6999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:50:23.666834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the full BAOMI protocol on the CirCor dataset with folds split strictly by patient ID, ensuring no patient appears in both training and test, and check whether the DAC+MFCC result (89.93% accuracy, 79.37% macro F1) and its margin over the cross-attention baseline persist; if the margin shrinks or vanishes, the state-of-the-art claim rests on leakage from non-patient-disjoint splits.","supporting_citations":[{"cited_title":"Heart murmur classification using a capsule neural network,","cited_arxiv_id":null,"evidence_quote":"Supplies the EnCodec architecture used to extract one of the neural audio codec representations."},{"cited_title":"Time-domain heart sound classification using neural audio codecs,","cited_arxiv_id":null,"evidence_quote":"Supplies the DAC model, whose representation yields the best fusion result when paired with MFCC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as a prior state-of-the-art method using MFCC features for heart murmur classification."},{"cited_title":"Snac: Multi- scale neural audio codec,","cited_arxiv_id":null,"evidence_quote":"One of the improved-MFCC baselines that BAOMI is compared against for the state-of-the-art claim."},{"cited_title":"The circor digiscope dataset: from murmur detection to murmur classifica- tion,","cited_arxiv_id":null,"evidence_quote":"Another improved-MFCC baseline used to position the proposed method's results."},{"cited_title":"Physiobank, physiotoolkit, and physionet: Compo- nents of a new research resource for complex physiologic signals,","cited_arxiv_id":null,"evidence_quote":"Deep-residual learning baseline with improved MFCC features used in the SOTA comparison."},{"cited_title":"Model-driven heart rate estimation and heart murmur detection based on phonocardiogram,","cited_arxiv_id":null,"evidence_quote":"Establishes the prior use of neural audio codec representations for heart sound classification."}],"review_version":1}