pith. sign in

arxiv: 2605.29862 · v1 · pith:725RQU5Bnew · submitted 2026-05-28 · 📡 eess.AS · cs.AI· cs.SD

Mitigating Stethoscope-Induced Shortcuts in Respiratory Sound Classification under Federated Domain Generalization with Causality-Inspired Interventions

Heejoon Koo , Yoon Tae Kim , Miika Toikkanen , June-Woo Kim This is my paper

Pith reviewed 2026-06-29 05:30 UTC · model grok-4.3

classification 📡 eess.AS cs.AIcs.SD
keywords respiratory sound classificationfederated domain generalizationcausality-inspired interventionsstethoscope variabilitydevice shiftscounterfactual augmentationmultimodal pretraining
0
0 comments X

The pith

A causality-inspired federated framework with style interventions and gradient alignment outperforms baselines for respiratory sound classification on unseen stethoscopes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to solve the problem of AI models for classifying respiratory sounds failing when tested on stethoscopes different from those used in training. Its empirical analysis finds that device style and disease signals are tightly mixed, so simply stripping style fails. The proposed solution is a federated domain generalization method that perturbs style while preserving content, augments text metadata to block shortcuts, and aligns gradients across clients. These steps run on top of a multimodal language-audio model and are evaluated by leaving out entire devices during testing on two public datasets. If the claim holds, medical AI for lung sounds could deploy across clinics without collecting data from every possible device in advance.

Core claim

The authors claim that a causality-inspired multimodal FedDG framework, which combines a device style intervention network performing content-preserving style perturbations, counterfactual text augmentation that neutralizes metadata shortcuts, and gradient alignment that produces device-invariant representations, outperforms standard data augmentation and federated learning baselines under leave-one-device-out validation on the ICBHI and SPRSound datasets.

What carries the argument

The causality-inspired multimodal FedDG framework that performs content-preserving style perturbations via a device style intervention network, neutralizes metadata shortcuts with counterfactual text augmentation, and aligns gradients across clients to learn device-invariant features.

If this is right

  • Models trained this way generalize to stethoscopes never seen during federated training.
  • The three components together avoid the failure mode of entangled style and content.
  • Multimodal pretraining supplies the base representations that the interventions refine.
  • Gradient alignment produces representations that remain stable across heterogeneous client devices.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same intervention pattern could be tested on other audio medical signals such as heart sounds or cough recordings where device type varies.
  • If the entanglement premise holds more broadly, similar causality tools might reduce site-specific bias in non-audio medical imaging tasks.
  • Real-world clinics could adopt the method without sharing raw patient recordings, provided the federated setup scales to dozens of sites.
  • Adding patient demographics as an extra text modality might further block unintended shortcuts beyond device metadata.

Load-bearing premise

Stethoscope-induced style and disease-specific content are tightly entangled, making deterministic style removal unreliable.

What would settle it

Demonstrating that a simple deterministic style removal method achieves equal or better leave-one-device-out performance on the same datasets would falsify the need for the proposed intervention network.

read the original abstract

AI-driven respiratory sound classification (RSC) is promising for automated pulmonary disease detection, yet multi-site deployment is hindered by inter-stethoscope variability. We introduce a federated domain generalization (FedDG) formulation for RSC under stethoscope-induced device shifts, where clients use heterogeneous devices and the model is evaluated on unseen devices. Our empirical analysis shows that stethoscope-induced style and disease-specific content are tightly entangled, making deterministic style removal unreliable. In response, we propose a causality-inspired multimodal FedDG framework that combines: (i) a causality-inspired device style intervention network that performs content-preserving style perturbations, (ii) counterfactual text augmentation that neutralizes metadata shortcuts, and (iii) gradient alignment that facilitates device-invariant representations across clients. Built on a multimodal language-audio pretraining model, it outperforms conventional data augmentation and federated learning baselines in leave-one-device-out validation on ICBHI and SPRSound datasets. Code will be released upon publication.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces a causality-inspired multimodal federated domain generalization (FedDG) framework for respiratory sound classification (RSC) to address stethoscope-induced device shifts. It combines (i) a causality-inspired device style intervention network performing content-preserving style perturbations, (ii) counterfactual text augmentation to neutralize metadata shortcuts, and (iii) gradient alignment to promote device-invariant representations. Built on a multimodal language-audio pretraining model, the approach is claimed to outperform conventional data augmentation and federated learning baselines under leave-one-device-out validation on the ICBHI and SPRSound datasets. The motivation stems from an empirical observation that stethoscope style and disease content are tightly entangled.

Significance. If the results hold with proper validation, the work could contribute to practical deployment of RSC models across heterogeneous devices in federated settings by providing a principled way to intervene on entangled style-content factors. The use of causality-inspired interventions, multimodal pretraining, and code release upon publication are positive elements that would support reproducibility and extension in medical audio domain generalization.

major comments (2)
  1. [Abstract] Abstract: The central claim of outperformance over baselines in leave-one-device-out validation is stated without any quantitative metrics, tables of results, statistical tests, or specific baseline names, which is load-bearing for assessing whether the combined interventions actually deliver device-invariant representations.
  2. [Abstract] Abstract: The motivation rests on an 'empirical analysis' showing tight entanglement between style and content, yet no description of the analysis method, dataset used for the analysis, or quantitative measure of entanglement is provided, leaving the justification for the three interventions unverified.
minor comments (1)
  1. [Abstract] The abstract is dense with technical terms (FedDG, counterfactual text augmentation, gradient alignment) without brief definitions or pointers to where they are formalized in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive feedback on the abstract. We address each major comment below and agree that the abstract can be strengthened for better self-containment while preserving conciseness. The full manuscript provides supporting details in dedicated sections.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of outperformance over baselines in leave-one-device-out validation is stated without any quantitative metrics, tables of results, statistical tests, or specific baseline names, which is load-bearing for assessing whether the combined interventions actually deliver device-invariant representations.

    Authors: We agree the abstract would be more informative with key quantitative support. In the revision, we will add specific metrics (e.g., accuracy and AUC improvements on ICBHI and SPRSound under leave-one-device-out), name the primary baselines (FedAvg, Mixup, and standard FL methods), and note that results include statistical significance testing via paired t-tests across multiple runs. Full tables and analysis appear in Section 4; this addition will directly address the load-bearing claim without exceeding abstract length limits. revision: yes

  2. Referee: [Abstract] Abstract: The motivation rests on an 'empirical analysis' showing tight entanglement between style and content, yet no description of the analysis method, dataset used for the analysis, or quantitative measure of entanglement is provided, leaving the justification for the three interventions unverified.

    Authors: The empirical analysis is fully described in Section 3.1, including the method (mutual information between device-style embeddings from a pretrained encoder and disease labels), dataset (ICBHI), and quantitative results (elevated MI scores confirming entanglement, rendering deterministic removal unreliable). To make the abstract self-contained, we will insert a brief clause such as 'via mutual information analysis on ICBHI showing high style-content entanglement' to justify the causality-inspired interventions. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper is an empirical ML methods contribution. Its central claim is that a proposed multimodal FedDG framework (style intervention network + counterfactual text augmentation + gradient alignment) outperforms baselines on leave-one-device-out splits of ICBHI and SPRSound. The motivation rests on an empirical observation of style-content entanglement rather than any derivation chain. No equations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text; the evaluation protocol is external to the method itself and the result is not forced by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption of tight entanglement between style and content, plus the unverified effectiveness of the three proposed interventions; no free parameters or invented entities are mentioned in the abstract.

axioms (1)
  • domain assumption Stethoscope-induced style and disease-specific content are tightly entangled, making deterministic style removal unreliable
    Presented as result of empirical analysis in the abstract; underpins the need for the causality-inspired interventions.

pith-pipeline@v0.9.1-grok · 5717 in / 1243 out tokens · 29855 ms · 2026-06-29T05:30:41.579507+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

    cs.SD 2026-06 unverdicted novelty 5.0

    QLung introduces quality-adaptive angular margins derived from spectral entropy and RMS energy to improve generalization in respiratory sound classification, reporting 2.46% gain on ICBHI and top OOD results on SPRSound.

Reference graph

Works this paper leans on

31 extracted references · 4 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Auscultation of the respiratory system,

    Malay Sarkar, Irappa Madabhavi, Narasimhalu Niranjan, and Megha Dogra, “Auscultation of the respiratory system,”Annals of thoracic medicine, vol. 10, no. 3, 2015

    2015

  2. [2]

    Patch-mix contrastive learning with audio spec- trogram transformer on respiratory sound classification,

    Sangmin Bae, June-Woo Kim, Won-Yang Cho, Hyerim Baek, Soy- oun Son, Byungjo Lee, Changwan Ha, Kyongpil Tae, Sungnyun Kim, and Se-Young Yun, “Patch-mix contrastive learning with audio spec- trogram transformer on respiratory sound classification,” inProc. Interspeech 2023, 2023

    2023

  3. [3]

    Bts: Bridging text and sound modalities for metadata-aided respiratory sound classification,

    June Woo Kim, Miika Toikkanen, Yera Choi, Seoung Eun Moon, and Ho Young Jung, “Bts: Bridging text and sound modalities for metadata-aided respiratory sound classification,” inProceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, 2024

    2024

  4. [4]

    Adaptive metadata-guided supervised contrastive learning for domain adaptation on respiratory sound classification,

    June-Woo Kim, Miika Toikkanen, Amin Jalali, Minseok Kim, Hye-Ji Han, Hyunwoo Kim, Wonwoo Shin, Ho-Young Jung, and Kyunghoon Kim, “Adaptive metadata-guided supervised contrastive learning for domain adaptation on respiratory sound classification,” IEEE Journal of Biomedical and Health Informatics, 2025

    2025

  5. [5]

    Lungmix: A mixup-based strategy for generalization in respiratory sound classification,

    Shijia Ge, Weixiang Zhang, Shuzhao Xie, Baixu Yan, and Zhi Wang, “Lungmix: A mixup-based strategy for generalization in respiratory sound classification,” inICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025

    2025

  6. [6]

    Empowering multimodal respiratory sound classification with counterfactual adversarial debiasing for out-of- distribution robustness,

    Heejoon Koo, Miika Toikkanen, Yoon Tae Kim, Soo Yong Kim, and June-Woo Kim, “Empowering multimodal respiratory sound classification with counterfactual adversarial debiasing for out-of- distribution robustness,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026

    2026

  7. [7]

    A survey on federated learning,

    Chen Zhang, Yu Xie, Hang Bai, Bin Yu, Weihong Li, and Yuan Gao, “A survey on federated learning,”Knowledge-Based Systems, vol. 216, 2021

    2021

  8. [8]

    Communication-efficient learning of deep networks from decentralized data,

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inArtificial intelligence and statistics. Pmlr, 2017

    2017

  9. [9]

    Empowering remote healthcare with federated learning for early di- agnosis of pulmonary disease,

    Santosh Kumar, Alexey V Shvetsov, and Saeed Hamood Alsamhi, “Empowering remote healthcare with federated learning for early di- agnosis of pulmonary disease,”IEEE Internet of Things Journal, 2025

    2025

  10. [10]

    Semi-supervised open-set federated learning for heterogeneous respiratory sound clas- sification,

    Won-Yang Cho, HyeSun Chang, and Sangjun Lee, “Semi-supervised open-set federated learning for heterogeneous respiratory sound clas- sification,” in2026 International Conference on AI x Data and Knowledge Engineering (AIxDKE). IEEE, 2026

    2026

  11. [11]

    A respiratory sound database for the development of automated classification,

    BM Rocha, Dimitris Filos, Lea Mendes, Ioannis V ogiatzis, Eleni Perantoni, Evangelos Kaimakamis, P Natsiavas, Ana Oliveira, C J´acome, A Marques, et al., “A respiratory sound database for the development of automated classification,” inInternational confer- ence on biomedical and health informatics. Springer, 2017

    2017

  12. [12]

    Electronic stethoscope filtering mimics the perceived sound characteristics of acoustic stethoscope,

    Valerie Rennoll, Ian McLane, Dimitra Emmanouilidou, James West, and Mounya Elhilali, “Electronic stethoscope filtering mimics the perceived sound characteristics of acoustic stethoscope,”IEEE jour- nal of biomedical and health informatics, vol. 25, no. 5, 2020

    2020

  13. [13]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Jun- suk Choe, and Youngjoon Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” inProceedings of the IEEE/CVF international conference on computer vision, 2019

    2019

  14. [14]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le, “Specaugment: A simple data augmentation method for automatic speech recognition,”arXiv preprint arXiv:1904.08779, 2019

    work page arXiv 1904

  15. [15]

    Repaugment: Input-agnostic representation-level augmentation for respiratory sound classification,

    June-Woo Kim, Miika Toikkanen, Sangmin Bae, Minseok Kim, and Ho-Young Jung, “Repaugment: Input-agnostic representation-level augmentation for respiratory sound classification,” in2024 46th An- nual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2024

    2024

  16. [16]

    Selecting data augmentation for simulating interventions,

    Maximilian Ilse, Jakub M Tomczak, and Patrick Forr ´e, “Selecting data augmentation for simulating interventions,” inInternational conference on machine learning. PMLR, 2021

    2021

  17. [17]

    Causality-inspired single-source domain generalization for medical image segmentation,

    Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert, “Causality-inspired single-source domain generalization for medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 42, no. 4, 2022

    2022

  18. [18]

    Understanding shortcut learning through the lens of causality & robustness,

    Yonghan Jung, “Understanding shortcut learning through the lens of causality & robustness,” 2022

    2022

  19. [19]

    A comprehensive survey on generative diffusion models for structured data,

    Heejoon Koo and To Eun Kim, “A comprehensive survey on generative diffusion models for structured data,”arXiv preprint arXiv:2306.04139, 2023

    work page arXiv 2023

  20. [20]

    Next visit diagnosis prediction via medical code- centric multimodal contrastive ehr modelling with hierarchical reg- ularisation,

    Heejoon Koo, “Next visit diagnosis prediction via medical code- centric multimodal contrastive ehr modelling with hierarchical reg- ularisation,” inFindings of the Association for Computational Lin- guistics: EACL 2024, 2024

    2024

  21. [21]

    Fedsr: A simple and effective domain generalization method for federated learning,

    A Tuan Nguyen, Philip Torr, and Ser Nam Lim, “Fedsr: A simple and effective domain generalization method for federated learning,” Advances in Neural Information Processing Systems, vol. 35, 2022

    2022

  22. [22]

    Out-of-distribution generalization of federated learning via implicit invariant relationships,

    Yaming Guo, Kai Guo, Xiaofeng Cao, Tieru Wu, and Yi Chang, “Out-of-distribution generalization of federated learning via implicit invariant relationships,” inInternational Conference on Machine Learning. PMLR, 2023

    2023

  23. [23]

    Federated out-of- distribution generalization: A causal augmentation view,

    Runhui Zhang, Sijin Zhou, and Zhuang Qi, “Federated out-of- distribution generalization: A causal augmentation view,”arXiv preprint arXiv:2504.19882, 2025

    work page arXiv 2025

  24. [24]

    Causal inference in statistics: An overview,

    Judea Pearl, “Causal inference in statistics: An overview,”Statistics Surveys, vol. 3, 01 2009

    2009

  25. [25]

    Large-scale contrastive language- audio pretraining with feature fusion and keyword-to-caption aug- mentation,

    Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg- Kirkpatrick, and Shlomo Dubnov, “Large-scale contrastive language- audio pretraining with feature fusion and keyword-to-caption aug- mentation,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023

    2023

  26. [26]

    Sprsound: Open-source sjtu paediatric respiratory sound database,

    Qing Zhang, Jing Zhang, Jiajun Yuan, Huajie Huang, Yuhang Zhang, Baoqin Zhang, Gaomei Lv, Shuzhu Lin, Na Wang, Xin Liu, et al., “Sprsound: Open-source sjtu paediatric respiratory sound database,” IEEE Transactions on Biomedical Circuits and Systems, 2022

    2022

  27. [27]

    Overcoming uncertain incompleteness for robust multimodal sequential diagnosis prediction via curriculum data eras- ing guided knowledge distillation,

    Heejoon Koo, “Overcoming uncertain incompleteness for robust multimodal sequential diagnosis prediction via curriculum data eras- ing guided knowledge distillation,” inICASSP 2025-2025 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025

    2025

  28. [28]

    Promptfl: Let federated participants cooperatively learn prompts instead of models–federated learning in age of foundation model,

    Tao Guo, Song Guo, Junxiao Wang, Xueyang Tang, and Wen- chao Xu, “Promptfl: Let federated participants cooperatively learn prompts instead of models–federated learning in age of foundation model,”IEEE Transactions on Mobile Computing, vol. 23, no. 5, 2023

    2023

  29. [29]

    mixup: Beyond Empirical Risk Minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz, “mixup: Beyond empirical risk minimization,”arXiv preprint arXiv:1710.09412, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    Ast: Audio spectro- gram transformer,

    Yuan Gong, Yu-An Chung, and James Glass, “Ast: Audio spectro- gram transformer,” inProc. Interspeech 2021, 2021

    2021

  31. [31]

    Stethoscope-guided supervised contrastive learn- ing for cross-domain adaptation on respiratory sound classification,

    June-Woo Kim, Sangmin Bae, Won-Yang Cho, Byungjo Lee, and Ho-Young Jung, “Stethoscope-guided supervised contrastive learn- ing for cross-domain adaptation on respiratory sound classification,” inICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024

    2024