PaperlayerView exact revision on arXiv

Source-linked AI summary

Clinically Informed Acoustic-Aware Attribute Modeling for Respiratory Sound Classification

Xinhong Li, Li Xiao, Ziyang Chen, Yuhong Yang, Weiping Tu

arXiv:2610.22854v1cs.SD

TL;DR

Flat four-class RSC does not explicitly model the compositional relationship between crackle and wheeze labels, while the sounds have different temporal patterns. AAAF detects the two attributes with separate acoustic-aware branches and composes the final label. Experiments report effectiveness and cross-dataset generalization, with a 65.30% Score for the last-four-layer strategy on ICBHI.

  • Problem

    Flat RSC classifiers treat four compositional labels as independent targets, while shared temporal aggregation may not capture both transient crackles and sustained wheezes.

    Source: §I · pp. 1–2

  • Method

    AAAF uses two attribute-specific detection branches with hierarchical BEATs representations and composes their crackle- and wheeze-presence predictions into a four-class label.

    Source: §I · p. 2

  • Results

    Experiments report AAAF’s effectiveness and cross-dataset generalization; the last-four-layer strategy achieves a Score of 65.30%.

    Source: §I · p. 2, 2) Impact of Layer Selection · p. 6, §V · p. 7

  • Takeaways & Limitations

    AAAF shows that compositional attribute detection and asymmetric temporal modeling can support respiratory sound classification and generalization across ICBHI and SPRSound.

    Source: §I · p. 2, §V · p. 7

Abstract

from arXiv · show

Automated respiratory sound classification (RSC) aims to identify normal and adventitious respiratory sounds, such as crackles and wheezes. Most existing methods formulate RSC as a flat four-class classification problem over norma, crackle, wheeze and both. However, unlike conventional classification tasks with semantically independent and mutually exclusive categories, the commonly used four-class RSC labels have a compositional structure and can be represented by two clinically meaningful sound attributes: crackle presence and wheeze presence. Existing flat classifiers do not explicitly exploit this structure, which may increase the learning burden by treating the four labels as independent targets. Meanwhile, crackles and wheezes exhibit distinct temporal characteristics, with sparse transient crackle events and sustained continuous wheeze patterns, making a single temporal aggregation strategy less suitable for modeling different adventitious sound evidence. To address these limitations, we propose an Acoustic-Aware Asymmetric Framework (AAAF) for RSC. AAAF reformulates four-class recognition as two sound-attribute detection tasks and recovers the final category through logical composition during inference. Built upon a pretrained BEATs encoder, AAAF employs branch-specific multi-layer fusion and asymmetric temporal aggregation to capture localized crackle evidence and sustained wheeze patterns. Auxiliary normal--abnormal supervision and a batch-adaptive dynamic loss are further introduced to improve global discrimination and robustness under imbalanced labels. Experiments on ICBHI and SPRSound demonstrate that AAAF improves overall RSC performance and generalizes across datasets, achieving a state-of-the-art four-class Score of 65.30% on ICBHI.

I. INTRODUCTION

RSC’s four labels encode two sound attributes, but flat classifiers treat them as independent targets; crackles and wheezes also require different temporal evidence. AAAF reformulates prediction around those attributes and pairs attribute-specific modeling with auxiliary supervision and adaptive loss.

  • The four RSC labels compose crackle and wheeze presence, yet flat classifiers treat them as independent targets and impose artificial mutual exclusion.The paper notes this may increase learning burden and obscure the labels’ compositional structure.Source: §I · p. 1
  • Crackles are localized, transient events, whereas wheezes are sustained; a shared temporal pipeline can struggle to capture both patterns.The paper motivates attribute-specific asymmetric temporal modeling to address this mismatch.Source: §I · pp. 1–2
  • AAAF reformulates four-class recognition as crackle- and wheeze-presence detection, then composes the final category from those attributes.Its pretrained BEATs encoder feeds attribute-specific branches with learnable multi-layer fusion and temporal modeling.Source: §I · p. 2
  • Auxiliary normal–abnormal supervision and batch-adaptive dynamic loss target global discrimination and robustness under imbalanced labels.Experiments on ICBHI and SPRSound are reported to demonstrate effectiveness and cross-dataset generalization.Source: §I · p. 2

II. METHOD

AAAF represents each respiratory event with crackle- and wheeze-presence attributes rather than a single flat label. It derives the four-class prediction by composing the two binary decisions.

II. METHOD

  • AAAF comprises task reformulation and inference composition, branch-specific multi-layer fusion, asymmetric temporal aggregation, auxiliary global supervision, and dynamic batch-adaptive loss.Source: §II · p. 2

A. Task Reformulation and Inference Composition

B. Acoustic-Aware Asymmetric Architecture

AAAF uses independently fused BEATs representations and different temporal aggregation strategies for crackle and wheeze branches. The resulting branch representations support attribute predictions and auxiliary global supervision.

1) Branch-Specific Multi-Layer Fusion:

  • Each attribute branch independently learns a weighted fusion of selected BEATs layers before its temporal aggregation stage.Both branches use the last four BEATs layers as candidates, with independent fusion weights.Source: 1) Branch-Specific Multi-Layer Fusion · p. 2

2) Acoustic-Aware Asymmetric Temporal Aggregation:

1) Auxiliary Global Supervision:

Auxiliary normal–abnormal supervision complements crackle- and wheeze-specific learning by preserving global discrimination between normal sounds and abnormal events.

1) Auxiliary Global Supervision:

  • The auxiliary task adds a global normal-versus-abnormal constraint alongside the separate crackle- and wheeze-presence branches.It is introduced because reliable classification also requires distinguishing normal respiratory sounds from abnormal acoustic events.Source: 1) Auxiliary Global Supervision · pp. 3–4
  • The auxiliary target treats normal as y = 0, while the branch-specific representations are concatenated and passed to a binary classifier.The concatenated representation is formed from zc and zw.Source: 1) Auxiliary Global Supervision · p. 4
  • The auxiliary objective encourages the attribute branches to retain global normal–abnormal discriminability while learning crackle- and wheeze-oriented representations.Source: 1) Auxiliary Global Supervision · p. 4

2) Dynamic Batch-Adaptive Loss:

The dynamic batch-adaptive loss adjusts positive-class weighting to minibatch prevalence, uses weighted cross-entropy across tasks, and incorporates the auxiliary loss into the final objective.

2) Dynamic Batch-Adaptive Loss:

  • Positive-class weights are adapted to each task’s positive ratio in the current minibatch, addressing prevalence variation across batches.The strategy targets imbalance in crackle, wheeze, and auxiliary normal–abnormal tasks.Source: 2) Dynamic Batch-Adaptive Loss · p. 4
  • A maximum limits positive weights under extreme imbalance, while the negative-class weight is fixed to 1 for each task.Source: 2) Dynamic Batch-Adaptive Loss · p. 4
  • The method applies weighted cross-entropy to each binary task and combines the task losses in a final objective whose auxiliary-supervision contribution is controlled by λ.Source: 2) Dynamic Batch-Adaptive Loss · p. 4

III. EXPERIMENTAL SETUP

Experiments use official dataset splits, a pretrained BEATs encoder, and the standard ICBHI metrics for four-class and binary evaluation.

A. Datasets

  • ICBHI evaluation uses the official subject-wise split: 4,142 respiratory cycles for training and 2,756 for testing, from 920 recordings and 126 subjects.The dataset contains 6,898 annotated cycles across Normal, Crackle, Wheeze, and Both.Source: A. Datasets · p. 4
  • SPRSound external validation uses its official split of 6,656 training and 2,433 testing events, mapping seven event categories into the same four-class space.Coarse/Fine Crackle map to Crackle; Rhonchi/Wheeze/Stridor to Wheeze; Wheeze+Crackle to Both; Normal remains unchanged.Source: A. Datasets · p. 4

B. Implementation Details

  • Audio is resampled to 16 kHz and cropped or padded to 5-second clips before the AudioSet-pretrained BEATs encoder; experiments use five seeds and report mean ± standard deviation.Adam, batch size 32, weight decay 1 × 10−6, cosine scheduling, differential learning rates, and λ = 0.2 are specified; thresholds are selected without test labels and fixed across seeds.Source: B. Implementation Details · p. 4

C. Evaluation Metrics

  • The ICBHI protocol reports Specificity, Sensitivity, and Score = (Sp + Se)/2; binary evaluation merges adventitious classes into Abnormal.For four-class evaluation, Normal is negative and the three adventitious classes are positive.Source: C. Evaluation Metrics · p. 4

IV. RESULTS

AAAF achieves the best reported ICBHI four-class Score and improves on a reproduced BEATs baseline in SPRSound validation; ablations report gains from its components.

1) Effectiveness of Proposed Components:

  • Branch-specific multi-layer fusion raises Sensitivity from 48.21% to 50.55%, and adding asymmetric temporal aggregation yields a 64.29% Score.The reported aggregation benefit concerns distinct strategies for transient crackles and sustained wheezes.Source: 1) Effectiveness of Proposed Components · pp. 4–5

A. Main Comparison

  • 65.30% four-class Score and 50.35% Sensitivity are the best reported ICBHI results; two-class evaluation reaches a 71.45% Score and 68.89% Sensitivity.The comparison is reported in Table I.Source: A. Main Comparison · p. 4, Table I · p. 5
  • 88.54% versus 88.38% four-class Score and 92.21% versus 92.02% two-class Score are reported for AAAF and reproduced BEATs, respectively, on SPRSound.The external-validation comparison is reported in Table II.Source: A. Main Comparison · p. 4, Table II · p. 5

1) Effectiveness of Proposed Components:

  • Auxiliary supervision raises Score to 64.49%, dynamic loss reaches 51.78% Sensitivity, and the combined components achieve the best Score of 65.30%.The authors describe the components as complementary, associating fusion with enriched representations, auxiliary supervision with global abnormality discrimination, and dynamic loss with robustness under imbalanced labels.Source: 1) Effectiveness of Proposed Components · p. 5

2) Impact of Layer Selection:

Layer-selection and Top-K ablations identify last-four-layer fusion and K = 5 as the best tested settings, each achieving a Score of 65.30%.

2) Impact of Layer Selection:

  • 65.30% Score is achieved by fusing the last four BEATs layers, exceeding final-layer-only (64.08%), interval layers (64.40%), and all-layer fusion (63.30%).Deeper layers provide a balance between representation diversity and feature consistency; all-layer fusion may add noisy or less relevant features.Source: 2) Impact of Layer Selection · pp. 5–6, Table IV · p. 6

3) Impact of Top-K Pooling:

  • 65.30% overall Score is achieved with K = 5, the best tested Top-K pooling setting for the crackle branch.Small K may miss contextual cues, while larger K may add irrelevant background and dilute localized crackle evidence; K = 5 is used in all experiments.Source: Table V · p. 6, 3) Impact of Top-K Pooling · p. 6

4) Frame-wise Branch Responses:

On a representative Both sample, the branches show complementary temporal selectivity; across ICBHI backbones, AAAF raises Score, reaching 65.30% with BEATs. The conclusion reports competitive generalization on SPRSound under distribution shifts and identifies robustness variations for future work.

4) Frame-wise Branch Responses:

  • On a representative Both sample, crackle responses peak at localized transients while wheeze responses remain broader and sustained, with the branches relatively inactive on the other's dominant intervals.Selected Top-K frames concentrate on high-activation crackle segments, providing qualitative evidence of complementary attribute-specific temporal cues.Source: Figure 3 · p. 6, 4) Frame-wise Branch Responses · p. 6

C. Backbone Generalization

  • 65.30% Score with BEATs improves over standard CE's 63.49% on ICBHI four-class evaluation; AAAF raises Score across all evaluated backbones.ResNet-50 and EfficientNet-B0 Score gains are 54.24% to 55.29% and 57.68% to 58.98%, respectively; AST gains 1.63 percentage points, improving both Specificity and Sensitivity.Source: Table VI · p. 6, C. Backbone Generalization · p. 6

V. CONCLUSION

  • The conclusion reports competitive SPRSound generalization under distribution shifts and identifies demographic, acquisition, and class-imbalance variations as targets for future robustness work.It also reports strong ICBHI performance in four-class classification and two-class abnormality detection.Source: §V · pp. 6–7
Loading 2610.22854v1…