Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Tri-MTL: A Triple Multitask Learning Approach for Respiratory Disease Diagnosis

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read Stethoscope metadata lifts both lung-sound and disease diagnosis scores in a multitask model.

desk verdict A plausible but under-supported MTL extension; the headline gains hinge on an ambiguous run-selection rule and unresolved number discrepancies. read the letter →

arxiv 2505.06271 v1 pith:QMIHRDGN submitted 2025-05-06 cs.LG cs.AIcs.SD

classification cs.LGcs.AIcs.SD
keywords multitasklearningrespiratorysoundclassificationdiseasediagnosisstethoscopemetadataAudioSpectrogramTransformerICBHIdatasetsoftparametersharingmetadata-aided
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that multitask learning can mirror a clinician's workflow: instead of classifying lung sounds and diseases separately, one model should do both and also predict patient metadata, with stethoscope type as a third task. On the ICBHI respiratory sound benchmark, the proposed Tri-MTL with soft parameter sharing and stethoscope metadata reports the best results: lung sound score 60.21 and disease diagnosis score 90.16, exceeding single-task and two-task baselines. The reason this matters is that it suggests recording-device information carries signal about how lung sounds are captured, and that a shared representation can use that signal to improve diagnosis. The authors frame the contribution as an extension of earlier findings that metadata aids respiratory sound classification, now tested inside a multitask framework.

What carries the argument

The machinery is a family of three multitask architectures built on a pretrained Audio Spectrogram Transformer audio encoder. Hard parameter sharing uses one shared encoder with separate classification heads for lung sounds, disease, and metadata, summing the three cross-entropy losses. Soft parameter sharing instead gives each task its own encoder and adds an L2 regularization loss that penalizes differences between corresponding layer weights, balancing task-specific and shared features. Tri-MTL is the version that adds the metadata classifier to either sharing scheme; the pair that carries the reported result is soft sharing with stethoscope-type metadata, where the metadata loss is included alongside the lung and disease losses.

What would settle it

Rerun Table II with the same five seeds and report every run, then compare Soft (Stethoscope) against Soft Two-MTL and Single Task using the full distribution: the lung-sound scores overlap within one standard deviation (60.21±1.42 vs 59.55±0.88) and the disease scores also overlap (90.16±3.19 vs 88.37±2.33), so a paired statistical test on all five runs would settle whether the claimed improvement is real.

Watch

Extended reading notes

Core claim

The central claim is that jointly training lung sound classification, disease diagnosis, and metadata classification improves performance over training either task alone or the two tasks together. The paper finds that the two-task MTL already raises disease diagnosis from 81.51 to 88.37, while lung sound classification barely moves; adding metadata as a third task, specifically the recording stethoscope type under soft parameter sharing, lifts lung sound score to 60.21 and disease diagnosis to 90.16. The authors interpret this as evidence that stethoscope metadata helps the model separate acoustic differences introduced by different recording devices, and that the lung-sound task supplies useful features for diagnosis even though the reverse direction gives less. The result is specific to the ICBHI official 60/40 split and the Audio Spectrogram Transformer backbone.

Load-bearing premise

The reported claim of improvement depends on the five-run statistics being honestly and completely reported; the paper selects runs where Sc is highest, so if that selection hides run-to-run variability, the gains over baselines could vanish.

Editorial extensions

If this is right

  • If the result is correct, stethoscope type should be treated as a useful input or auxiliary task in respiratory sound models, not as a nuisance variable to be averaged away.
  • Multitask learning offers a practical route to disease diagnosis from lung sounds: the two-task model already beats the single-task disease baseline by several score points.
  • Soft parameter sharing, which lets each task keep separate weights while encouraging similarity, appears better suited than hard sharing for this pair of tasks.
  • The reported numbers become the new comparison point for the ICBHI benchmark under AST-based encoders with the official 60/40 split.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit: a model that receives stethoscope type as a conditioning input rather than predicting it as a task might capture the same device-effect information with fewer parameters.
  • If device-specific acoustic effects are the true driver, the gains should shrink on datasets recorded with a single stethoscope type and grow where multiple devices are mixed; that is a checkable prediction the paper does not run.
  • The location-metadata results suggest some metadata tasks can hurt diagnosis (Hard Location falls below single-task performance), so a selective metadata-gating mechanism is a natural next step the authors do not explore.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Tri-MTL, a multitask learning framework for respiratory disease diagnosis that jointly optimizes lung sound classification, disease diagnosis, and a metadata classification task on the ICBHI dataset. The authors compare hard and soft parameter-sharing variants against single-task and two-task baselines using a pretrained Audio Spectrogram Transformer. Their headline result is that adding stethoscope metadata under soft parameter sharing improves both lung sound classification (Sc = 60.21 ± 1.42) and disease diagnosis (Sc = 90.16 ± 3.19), relative to two-task soft sharing (Sc = 59.19 ± 0.93 and 88.37 ± 2.33, respectively). The paper also provides t-SNE visualizations to explain why MTL benefits disease diagnosis more than lung sound classification. The central claim is that stethoscope metadata, integrated as a third task, yields significant improvements in both objectives.

Significance. If the reported results are valid, Tri-MTL is a simple and practical extension of existing MTL ideas, and the finding that stethoscope metadata improves both tasks would be clinically relevant and would strengthen the prior evidence base for metadata-aided respiratory sound classification. The paper builds on a public benchmark (ICBHI), uses a standard pretrained backbone, and reports comparisons across several metadata attributes, which are useful for the community. However, the significance is currently tempered by ambiguous run-selection reporting, internal numerical inconsistencies, and the absence of significance testing: the claimed improvements are not established by the evidence as presented. The authors also do not release code, which limits reproducibility of the exact numbers.

major comments (3)
  1. [Section III.A.3] The sentence 'every reported value of Sp, Se, and Sc reflects the mean and variance across five independent runs where Sc is highest' is ambiguous and potentially selection-biased. If this means that the five retained runs were selected for having the highest Sc, then the reported means are order statistics rather than unbiased estimates, and all comparisons against Two-MTL and single-task baselines are invalid under the current reporting. If it instead means that five seeds were run and the reported Sc corresponds to the run with highest Sc, this is even more problematic. Please state exactly how the five runs were chosen, report results for all runs or fixed seeds without post hoc selection, and provide per-run breakdowns or code so the reader can verify the statistics.
  2. [Section III.B.2 and Table II] The text in Section III.B.2 reports disease diagnosis Sc = 90.77 ± 1.36 for Soft (Stethoscope), while Table II lists 90.16 ± 3.19 for the same configuration. This is a direct internal inconsistency. In addition, the abstract's clinical-relevance specificity of 85.83% for disease diagnosis does not match Table II's 86.23% for Soft (Stethoscope). These discrepancies mean that the reported statistics are not a complete and reproducible record, so the reader cannot determine which numbers support the headline claim.
  3. [Table II and Abstract] The central claim of 'significant improvements' is not supported by the reported variability. For lung sound classification, Soft (Stethoscope) gives 60.21 ± 1.42 versus Two-MTL Soft 59.19 ± 0.93; for disease diagnosis, 90.16 ± 3.19 versus 88.37 ± 2.33. With five runs per condition and no paired significance test, these differences are within plausible sampling noise. Please report paired tests (e.g., paired t-test or bootstrap) across the same seeds, confidence intervals, or effect sizes for the key comparisons in Table II. Without this, the word 'significant' in the abstract is unsupported.
minor comments (5)
  1. [Equation (3)] Equation (3) contains a typographical error: 'LHard = = Σ' should be 'LHard = Σ'.
  2. [Equation (4)] The cross-entropy loss in Equation (4) is written without division by N; the text mentions this omission, but it would be clearer to write the normalized form or explicitly define the notation as an unnormalized sum over the batch.
  3. [Figure 1 caption] Figure 1(c) includes '× α', but α is not defined anywhere in the text or equations; please either define it or remove it from the figure.
  4. [Section III.A.3] The text says 'mean and variance' but the tables report mean ± standard deviation; please align the wording with the reported quantities.
  5. [Section IV] The conclusion contains a typo: 'ICHBI respiratory sound dataset' should be 'ICBHI respiratory sound dataset'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical MTL benchmark with self-contained test-set comparisons.

full rationale

The paper is an empirical comparison on the fixed ICBHI 60/40 split, not a derivation chain. The central claim that adding a stethoscope-metadata task improves lung sound and disease classification rests on the test-set metrics in Table II and the corresponding experimental protocol in Section III.A. No parameter is fitted to the target metric and then reported as a prediction; the metadata auxiliary task is defined by dataset labels, not by the target outputs. Self-citations [20] and [21] are used only as motivation for considering metadata, and the paper's own independent experiments constitute the evidence, so those citations are not load-bearing. The ambiguous wording in Section III.A.3 ('five independent runs where Sc is highest') is a statistical reporting and reproducibility concern rather than a circularity step, because the reported improvements do not reduce by construction to any fitted input. Similarly, the internal inconsistency between Section III.B.2 (disease Sc = 90.77±1.36) and Table II (90.16±3.19) is a consistency issue, not evidence of circularity. The derivation and evaluation are therefore self-contained; no circular step is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim is an empirical comparison, not a derivation. It relies on the ICBHI dataset, the pretrained AST backbone, and the choices of auxiliary tasks and hyperparameters. No new entities are introduced. The main free parameters are the soft-sharing regularization weight, the class weights for the weighted loss, and the fixed training hyperparameters.

free parameters (3)
  • Regularization weight lambda = 0.1
    Set manually for LReg in Eq. (7) during soft parameter sharing training. The best result depends on this value, and no sensitivity analysis is reported.
  • Disease diagnosis class weights = Inversely proportional to class frequency
    Applied to the cross-entropy loss for disease diagnosis (Sec. III.A.2). These weights are derived from the training set and influence the reported specificity/sensitivity balance.
  • Training hyperparameters (learning rate, epochs, batch size) = 5e-5, 50, 8
    Fixed for all runs without tuning. The comparison assumes these settings are adequate for all configurations, and no hyperparameter search is reported.
assumptions (3)
  • domain assumption AST pretrained on AudioSet transfers to respiratory sound classification
    The paper uses a pretrained Audio Spectrogram Transformer and fine-tunes it on ICBHI (Sec. III.A.2). If the pretrained representations are inappropriate for respiratory acoustics, the comparison is moot.
  • domain assumption The stethoscope metadata in ICBHI is recorded accurately and consistently across recordings
    The central claim depends on stethoscope type being a meaningful, reliable auxiliary label. If the metadata is noisy or inconsistently recorded, the auxiliary task could be uninformative.
  • domain assumption The official 60/40 split of ICBHI is patient-disjoint
    The analysis trusts the official pre-split described in Sec. III.A.1. A leak between training and test patients would invalidate the performance comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tri-MTL: A Triple Multitask Learning Approach for Respiratory Disease Diagnosis." pith.science (2026). https://pith.science/paper/QMIHRDGN

@misc{pith2026250506271,
  author       = {Pith},
  title        = {Pith review of: Tri-MTL: A Triple Multitask Learning Approach for Respiratory Disease Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMIHRDGN}},
  note         = {Machine review of arXiv:2505.06271}
}
read the original abstract

Auscultation remains a cornerstone of clinical practice, essential for both initial evaluation and continuous monitoring. Clinicians listen to the lung sounds and make a diagnosis by combining the patient's medical history and test results. Given this strong association, multitask learning (MTL) can offer a compelling framework to simultaneously model these relationships, integrating respiratory sound patterns with disease manifestations. While MTL has shown considerable promise in medical applications, a significant research gap remains in understanding the complex interplay between respiratory sounds, disease manifestations, and patient metadata attributes. This study investigates how integrating MTL with cutting-edge deep learning architectures can enhance both respiratory sound classification and disease diagnosis. Specifically, we extend recent findings regarding the beneficial impact of metadata on respiratory sound classification by evaluating its effectiveness within an MTL framework. Our comprehensive experiments reveal significant improvements in both lung sound classification and diagnostic performance when the stethoscope information is incorporated into the MTL architecture.

Figures

Figures reproduced from arXiv: 2505.06271 by the authors.

Figure 1
Figure 1. Illustration of the proposed MTL approaches for joint lung sound classification and disease diagnosis. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. t-SNE visualizations illustrating how models trained [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

    Laennec versus forbes: tied for the score! how technology helps us interpret auscultation,

    P. Coucke, “Laennec versus forbes: tied for the score! how technology helps us interpret auscultation,” Revue Medicale de Liege , vol. 74, no. 10, pp. 543–551, 2019

  2. [2]

    Auscultation of the respiratory system,

    M. Sarkar, I. Madabhavi, N. Niranjan, and M. Dogra, “Auscultation of the respiratory system,” Annals of thoracic medicine, vol. 10, no. 3, pp. 158–168, 2015

  3. [3]

    The coming era of a new auscultation system for analyzing respiratory sounds,

    Y . Kim, Y . Hyon, S. Lee, S.-D. Woo, T. Ha, and C. Chung, “The coming era of a new auscultation system for analyzing respiratory sounds,” BMC Pulmonary Medicine , vol. 22, no. 1, p. 119, 2022

  4. [4]

    Approach to the respiratory patient,

    C. J. Miller, “Approach to the respiratory patient,” Veterinary Clinics of North America: Small Animal Practice, vol. 37, no. 5, pp. 861–878, 2007

  5. [5]

    Multi-task convolu- tional neural network for patient detection and skin segmentation in continuous non-contact vital sign monitoring,

    S. Chaichulee, M. Villarroel, J. Jorge, C. Arteta, G. Green, K. Mc- Cormick, A. Zisserman, and L. Tarassenko, “Multi-task convolu- tional neural network for patient detection and skin segmentation in continuous non-contact vital sign monitoring,” in 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017). IEEE, 2017, pp. 266–272

  6. [6]

    A novel multi-task deep learning model for skin lesion segmentation and classification,

    X. Yang, Z. Zeng, S. Y . Yeo, C. Tan, H. L. Tey, and Y . Su, “A novel multi-task deep learning model for skin lesion segmentation and classification,” arXiv preprint arXiv:1703.01025 , 2017

  7. [7]

    Canet: cross- disease attention network for joint diabetic retinopathy and diabetic macular edema grading,

    X. Li, X. Hu, L. Yu, L. Zhu, C.-W. Fu, and P.-A. Heng, “Canet: cross- disease attention network for joint diabetic retinopathy and diabetic macular edema grading,” IEEE transactions on medical imaging , vol. 39, no. 5, pp. 1483–1493, 2019

  8. [8]

    Multi-task deep convolutional neural network for cancer diagnosis,

    Q. Liao, Y . Ding, Z. L. Jiang, X. Wang, C. Zhang, and Q. Zhang, “Multi-task deep convolutional neural network for cancer diagnosis,” Neurocomputing, vol. 348, pp. 66–73, 2019

Show all 33 references
  1. [9]

    Multi-task deep learning based ct imaging analysis for covid-19 pneumonia: Classi- fication and segmentation,

    A. Amyar, R. Modzelewski, H. Li, and S. Ruan, “Multi-task deep learning based ct imaging analysis for covid-19 pneumonia: Classi- fication and segmentation,” Computers in biology and medicine , vol. 126, p. 104037, 2020

  2. [10]

    A feature transfer enabled multi-task deep learning model on medical imaging,

    F. Gao, H. Yoon, T. Wu, and X. Chu, “A feature transfer enabled multi-task deep learning model on medical imaging,” Expert Systems with Applications, vol. 143, p. 112957, 2020

  3. [11]

    An end-to-end multi-task deep learning framework for skin lesion analysis,

    L. Song, J. Lin, Z. J. Wang, and H. Wang, “An end-to-end multi-task deep learning framework for skin lesion analysis,” IEEE journal of biomedical and health informatics , vol. 24, no. 10, pp. 2912–2921, 2020

  4. [12]

    Multi-task learning for lung sound and lung disease classification,

    K. Suma, D. Koppad, P. Kumar, N. A. Kantikar, and S. Ramesh, “Multi-task learning for lung sound and lung disease classification,” SN Computer Science , vol. 6, no. 1, p. 51, 2024

  5. [13]

    A multi-task learning based applicable ai model simultaneously predicts stage, histology, grade and lnm for cervical cancer before surgery,

    Z. Wang, H. Gao, X. Wang, M. Grzegorzek, J. Li, H. Sun, Y . Ma, X. Zhang, Z. Zhang, A. Dekker et al. , “A multi-task learning based applicable ai model simultaneously predicts stage, histology, grade and lnm for cervical cancer before surgery,” BMC women’s health, vol. 24, no....

  6. [14]

    Cross-task attention network: Improving multi-task learning for medical imaging applications,

    S. Kim, T. G. Purdie, and C. McIntosh, “Cross-task attention network: Improving multi-task learning for medical imaging applications,” inIn- ternational Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2023, pp. 119–128

  7. [15]

    Intra-person multi-task learning method for chronic-disease prediction,

    G. Kim, H. Lim, Y . Kim, O. Kwon, and J.-H. Choi, “Intra-person multi-task learning method for chronic-disease prediction,” Scientific Reports, vol. 13, no. 1, p. 1069, 2023

  8. [16]

    Revisiting multi-task learning in the deep learning era,

    S. Vandenhende, S. Georgoulis, M. Proesmans, D. Dai, and L. Van Gool, “Revisiting multi-task learning in the deep learning era,” arXiv preprint arXiv:2004.13379, vol. 2, no. 3, pp. 5491–5500, 2020

  9. [17]

    Multi- task learning for lung sound & lung disease classification,

    S. KV , D. Koppad, P. Kumar, N. A. Kantikar, and S. Ramesh, “Multi- task learning for lung sound & lung disease classification,” arXiv preprint arXiv:2404.03908, 2024

  10. [18]

    Automated multi-task learning for joint disease prediction on electronic health records,

    S. Cui and P. Mitra, “Automated multi-task learning for joint disease prediction on electronic health records,” arXiv preprint arXiv:2403.04086, 2024

  11. [19]

    Pretraining respiratory sound repre- sentations using metadata and contrastive learning,

    I. Moummad and N. Farrugia, “Pretraining respiratory sound repre- sentations using metadata and contrastive learning,” in 2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2023, pp. 1–5

  12. [20]

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

    J.-W. Kim, S. Bae, W.-Y . Cho, B. Lee, and H.-Y . Jung, “Stethoscope- guided supervised contrastive learning for cross-domain adaptation on respiratory sound classification,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)....

  13. [21]

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

    J.-W. Kim, M. Toikkanen, Y . Choi, S.-E. Moon, and H.-Y . Jung, “Bts: Bridging text and sound modalities for metadata-aided respiratory sound classification,” in Interspeech 2024, 2024, pp. 1690–1694

  14. [22]

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

    J.-W. Kim, M. Toikkanen, A. Jalali, M. Kim, H.-J. Han, H. Kim, W. Shin, H.-Y . Jung, and K. Kim, “Adaptive metadata-guided super- vised contrastive learning for domain adaptation on respiratory sound classification,” IEEE Journal of Biomedical and Health Informatics , 2025

  15. [23]

    AST: Audio Spectrogram Transformer,

    Y . Gong, Y .-A. Chung, and J. Glass, “AST: Audio Spectrogram Transformer,” in Proc. Interspeech 2021 , 2021, pp. 571–575

  16. [24]

    Patch-Mix Contrastive Learning with Audio Spectrogram Transformer on Respiratory Sound Classification,

    S. Bae, J.-W. Kim, W.-Y . Cho, H. Baek, S. Son, B. Lee, C. Ha, K. Tae, S. Kim, and S.-Y . Yun, “Patch-Mix Contrastive Learning with Audio Spectrogram Transformer on Respiratory Sound Classification,” in Proc. INTERSPEECH 2023 , 2023, pp. 5436–5440

  17. [25]

    A respiratory sound database for the development of automated classification,

    B. Rocha, D. Filos, L. Mendes, I. V ogiatzis, E. Perantoni, E. Kaimakamis, P. Natsiavas, A. Oliveira, C. Jácome, A. Marques et al., “A respiratory sound database for the development of automated classification,” in Precision Medicine Powered by pHealth and Con- nected Health: ...

  18. [26]

    An overview of multi-task learning in deep neural net- works,

    S. Ruder, “An overview of multi-task learning in deep neural net- works,” arXiv preprint arXiv:1706.05098 , 2017

  19. [27]

    A bayesian/information theoretic model of learning to learn via multiple task sampling,

    J. Baxter, “A bayesian/information theoretic model of learning to learn via multiple task sampling,”Machine learning, vol. 28, pp. 7–39, 1997

  20. [28]

    Multitask learning: A knowledge-based source of induc- tive bias1,

    R. Caruana, “Multitask learning: A knowledge-based source of induc- tive bias1,” in Proceedings of the Tenth International Conference on Machine Learning. Citeseer, 1993, pp. 41–48

  21. [29]

    Low resource dependency parsing: Cross-lingual parameter sharing in a neural network parser,

    L. Duong, T. Cohn, S. Bird, and P. Cook, “Low resource dependency parsing: Cross-lingual parameter sharing in a neural network parser,” in Proceedings of the 53rd annual meeting of the Association for Computational Linguistics and the 7th international joint conference on natu...

  22. [30]

    Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,

    S. Gairola, F. Tom, N. Kwatra, and M. Jain, “Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,” in 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) . IEEE, 2021, pp. 527–530

  23. [31]

    Adversarial fine-tuning using generated respiratory sound to address class imbalance,

    J.-W. Kim, C. Yoon, M. Toikkanen, S. Bae, and H.-Y . Jung, “Adversarial fine-tuning using generated respiratory sound to address class imbalance,” in Deep Generative Models for Health Workshop NeurIPS 2023 , 2023. [Online]. Available: https://openreview.net/ forum?id=z1A VG5LDQ7

  24. [32]

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

    J.-W. Kim, M. Toikkanen, S. Bae, M. Kim, and H.-Y . Jung, “Repaug- ment: Input-agnostic representation-level augmentation for respiratory sound classification,” in 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) , 2024, pp. 1–6

  25. [2017]

    Springer, 2018, pp. 33–37

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.