Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses

Seong Eun Kim; Seung Gyu Jeong

REVIEW 4 major objections 5 minor 1 cited by

Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Two patient-identity losses lift lung sound classification on ICBHI.

desk verdict A modest, clearly explained metric-learning tweak for lung sound classification whose reported gain is credible only if the missing validation split existed. read the letter →

arxiv 2505.23834 v1 pith:GJ5NTDSC submitted 2025-05-28 cs.SD cs.LGeess.AS

Seung Gyu Jeong , Seong Eun Kim This is my paper

classification cs.SDcs.LGeess.AS

keywords lungsoundclassificationpatientvariabilityfeaturealignmentcohesion-separationlossglobalBEATsICBHI2017biomedicalaudio

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

Lung sound classifiers are usually trained to tell disease classes apart, and the authors argue this ignores a second source of structure: each patient's acoustic identity. The paper proposes training with two extra losses that make the learned features respect patient identity, one pulling the same patient's recordings together and pushing different patients apart, and the other anchoring every patient cluster near a common center. On the ICBHI respiratory sound benchmark, adding these losses to a pretrained BEATs audio transformer raises the four-class score from 63.49 to 64.84 and the two-class score from 70.76 to 72.08 compared with cross-entropy alone. The claim is that preserving patient-specific structure, rather than letting it collapse into large class clusters, improves accuracy especially for underrepresented patient profiles. If true, it means patient identity is useful training signal, not nuisance variation.

What carries the argument

The engine of the method is a training objective that adds two patient-level losses to cross-entropy: $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{CE}} + \lambda_{\text{pcsl}}\mathcal{L}_{\text{PCSL}} + \lambda_{\text{gpal}}\mathcal{L}_{\text{GPAL}}$. PCSL is the ratio of within-patient scatter to between-patient scatter, $\mathcal{L}_{\text{PCSL}} = S_W/(S_B+\epsilon)$, a patient-level analogue of linear discriminant analysis, and it clusters embeddings of the same patient while separating different patients. GPAL is the mean squared distance from each patient centroid to the global centroid of all patients in the batch, $\mathcal{L}_{\text{GPAL}} = \frac{1}{|P|}\sum_{p\in P}\|\mu_p - \mu_G\|_2^2$, which stops the patient clusters from drifting apart and fragmenting the class structure. A projection head computes these losses during training and is discarded at inference, so the deployment model is unchanged in size.

What would settle it

Retrain on the official ICBHI split while holding out a subset of the test patients solely for choosing $\lambda_{\text{pcsl}}$ and $\lambda_{\text{gpal}}$, then score only the untouched test patients; if the score gain over cross-entropy does not reappear, the reported 1.3-point improvement was at least partly a fitted value from test-set tuning.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that adding patient-identity constraints to a class-labeled lung sound classifier improves classification without any extra information at inference. With the BEATs backbone, the combined PAFA objective reaches 64.84% on the four-class ICBHI score and 72.08% on the two-class score, against 63.49% and 70.76% for the same backbone trained with cross-entropy only, averaged over five seeds. The gain is consistent across three backbones, and ablations show that removing either patient-aware loss drops the score to roughly 63.9%, so the two losses are complementary. Visualizations and per-patient accuracy comparisons support the authors' explanation that PAFA works by keeping the features of less-represented patients from being absorbed into dominant class clusters.

Load-bearing premise

The load-bearing premise is that the penalty weights $\lambda_{\text{pcsl}}=50$ and $\lambda_{\text{gpal}}=0.0005$, found by grid search, were tuned on a validation split that does not include test patients; the paper never reports such a split.

Editorial extensions

If this is right

On the official ICBHI split, PAFA raises the four-class score from 63.49 to 64.84 and the two-class score from 70.76 to 72.08 over the same BEATs baseline.
The improvement is not backbone-specific: PAFA adds 0.41, 1.26, and 1.35 score points on CNN6, AST, and BEATs respectively.
Ablations show the two losses are complementary: dropping either PCSL or GPAL leaves the score near 63.9, confirming both are needed for the full gain.
At inference the projection head is removed, so the patient-aware training cost does not add parameters or compute at deployment.
The t-SNE and per-patient analyses indicate the gain concentrates in patients whose features would otherwise merge into dominant class clusters.

Reading between the lines

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

Because the method only needs patient labels during training, a direct test of the same idea is to run PAFA on other patient-centric biomedical audio benchmarks, such as heart sound or pathological speech classification, where feature collapse into majority classes is also a known failure mode.
The paper does not report a patient-disjoint validation split for the hyperparameter grid search; a stricter test would tune the two weights on a held-out patient subset and report the untouched test score, which could be lower than 64.84.
The patient centroids PAFA learns suggest an inexpensive personalization scheme: after deployment, a new patient's few recordings could be matched to the nearest learned patient cluster and then refined with only a handful of labeled examples.
Since PCSL is a ratio of scatter terms, its signal depends on which patients appear in a batch; varying batch composition or using per-patient sampling may change the gain more than it would for cross-entropy alone.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: the contribution is a patient-level Fisher-LDA-style loss (PCSL) paired with a global centroid penalty (GPAL), added to a BEATs transformer. The combination is new in this domain, and the authors are upfront that PCSL is Fisher-inspired. What the paper does well: it gives clean equations, tests three backbones, includes an ablation, reports means and stds over five seeds, and links code. The ablation shows both losses matter relative to leaving one out, and the gain direction is consistent across CNN6, AST, and BEATs, which is more than many papers at this level do.

The soft spots are real. The lone sentence on hyperparameters says a grid search set lambda_pcsl=50 and lambda_gpal=0.0005. There is no validation split described anywhere; Section 3.1 only reports the official 60/40 train/test split. If those lambdas were chosen by evaluating on the test set, the 1.35-point gain is a fitted maximum over whatever pairs were tried, not a predictive result. With five-seed stds around 0.6–1.1 points, that is a load-bearing problem, not a cosmetic one. The paper also skips significance testing entirely, and the t-SNE/patient bar charts in Figures 2–5 are qualitative; Figure 5 explicitly selects six test patients nearest to sub-cluster centroids, so it cannot support a general claim about test-phase gains. None of this refutes the method; it means the central empirical claim is not established as written.

I do not see a theoretical or citation problem worth worrying about. PCSL is a direct analogue of Fisher's ratio with ‘class’ replaced by ‘patient,’ and GPAL is a simple variance penalty; the authors cite Fisher and neural collapse appropriately. The code link is a plus, assuming it actually runs.

Who is this for? Someone working on respiratory sound classification or patient-conditional metric learning who wants a simple baseline to compare against. It is not a field-changer. A serious referee could fix the validation issue with one paragraph if the authors actually used a patient-disjoint validation split. In its current form I would not trust the headline number, but I would send it to review rather than desk reject; the method is coherent and the experiments are reproducible in principle.

Referee Report

4 major / 5 minor

Summary. The paper proposes PAFA, a patient-aware feature alignment method for lung sound classification, which augments a BEATs backbone with two auxiliary losses: Patient Cohesion-Separation Loss (PCSL), which clusters embeddings from the same patient and pushes different patient centroids apart, and Global Patient Alignment Loss (GPAL), which draws patient centroids toward a common global centroid. On the ICBHI 2017 dataset, the authors report that PAFA improves the four-class ICBHI score from 63.49% to 64.84% and the two-class score from 70.76% to 72.08% relative to the same BEATs backbone trained with cross-entropy alone. They also report backbone comparisons on CNN6 and AST, an ablation of PCSL and GPAL, and t-SNE visualizations intended to show that PAFA preserves patient-specific sub-clusters.

Significance. The core idea is simple, plausible, and potentially reusable for other patient-centric biomedical audio tasks. The paper provides code, reports five-seed means with standard deviations, and includes ablations of the two loss components as well as experiments across three backbones. However, the reported gains are small relative to run-to-run variation, no statistical significance tests are provided, and the hyperparameter selection protocol is not documented. The central quantitative claim therefore needs additional verification before the result can be considered established.

major comments (4)

[2.2, Eq. (1), Table 1] The central quantitative claim rests on the two loss weights λ_pcsl and λ_gpal, which Section 2.2 says were set by grid search to 50 and 0.0005. The paper does not report the search range, the selection criterion, or, most importantly, whether the grid search was performed on a validation split disjoint from the official ICBHI test set. Sections 3.1 and 4.1 name only the official 60/40 train/test split, so the only held-out labels available for selection appear to be the test labels. If that was the case, the reported PAFA gains of 1.35 and 1.32 points are fitted values rather than predictive improvements. Please specify the validation protocol, and if none was used, re-run the selection on a patient-disjoint validation set and report the test score for the chosen weights.
[Table 1, Section 4.1] The reported differences between BEATs+CE and BEATs+PAFA are small relative to the run-to-run variation: four-class 64.84±0.60 vs 63.49±1.08 and two-class 72.08±0.55 vs 70.76±0.42. No significance test or per-seed results are provided, so it is not possible to tell whether the improvement is beyond noise. Please add paired significance tests (e.g., a paired t-test or Wilcoxon signed-rank over the five seeds, or a patient-level bootstrap) and report per-seed scores.
[4.4, Figure 5] The claim that PAFA 'significantly improves' test performance for the patients most affected by feature collapse is based on a post-hoc selection of six test patients nearest to centroids of training sub-clusters. This selection can inflate the apparent benefit, and the figure shows no error bars or aggregate statistics. Please report the distribution of per-patient accuracy changes across all test patients, with a paired test over patients, and clearly label Figure 5 as an illustrative case study rather than evidence of overall improvement.
[3.1, 4.1, Section 5] The paper repeatedly refers to 'novel test patients' (e.g., Section 5), but it does not state whether the official 60/40 split is patient-disjoint. If the same patients appear in both train and test, the patient-aware losses could exploit patient identity rather than learning generalizable representations. Please state explicitly whether the split is patient-disjoint, and if not, evaluate under a patient-disjoint split.

minor comments (5)

[2.2.1, Eq. (5)] The constant ϵ in Eq. (5) is not given a value; specify it for reproducibility.
[3.1] The train/test class counts are useful; please also report the number of patients in each split, since this is relevant to the patient-disjointness concern.
[4.4, Figures 2–4] The t-SNE visualizations lack axis labels, legends, and an explanation of the color code; clarify what the colors represent and note that they are computed on training data.
[Table 1 caption] The symbols 'X' and 'O' in the Metadata column are not defined in the caption; define them explicitly.
[Abstract] The phrase 'outstanding results' is subjective; replace it with the actual numerical comparison or a more neutral characterization.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PAFA's reported gains are empirical measurements on the official ICBHI test split, and no load-bearing step reduces to its own inputs.

full rationale

The paper's central claims are empirical: Table 1 reports that BEATs+PAFA scores 64.84% (four-class) and 72.08% (two-class) on the ICBHI dataset, compared with 63.49% and 70.76% for the CE-only baseline. These are measured results after training with the combined loss in Eq. (1), not derived predictions. PCSL and GPAL (Eqs. 5 and 7) are defined directly from training-batch patient centroids, and the projection head is removed at inference (Section 2.3), so the reported classification scores are not constructed to equal the losses. The only potentially circular element is the hyperparameter selection: Section 2.2 states that 'a grid search determined that λpcsl = 50 and λgpal = 0.0005 yield the best performance,' while Section 3.1 and Section 4.1 describe only the official 60/40 train/test split. If those weights had been chosen on the test set, the 1.35-point and 1.32-point gains would be fitted values rather than predictions. However, the paper does not state that the grid search used the test set, so this is an unproven protocol ambiguity, not a demonstrated circular step. There is no load-bearing self-citation, no imported uniqueness theorem, no ansatz smuggled in by citation, and no known result merely renamed: PCSL is explicitly identified as Fisher LDA with 'class' replaced by 'patient,' and GPAL is a centroid-variance regularizer. The ablation in Table 3 is an empirical sensitivity check, not a circular prediction. Accordingly, the derivation chain is self-contained and the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The central claim depends on the two tuned loss weights and on the ICBHI split being patient-disjoint. No new entities are introduced.

free parameters (2)

lambda_pcsl = 50
Selected by grid search to maximize ICBHI score; no validation split reported.
lambda_gpal = 0.0005
Selected by grid search; same issue.

assumptions (2)

domain assumption The official ICBHI 60/40 split is patient-disjoint.
The paper claims cross-subject generalization; if patients appear in both train and test, PCSL's patient clustering would leak identity into evaluation.
ad hoc to paper The two loss weights selected by grid search generalize to the test distribution.
The weights are tuned and the paper does not document a validation split.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses." pith.science (2026). https://pith.science/paper/GJ5NTDSC

@misc{pith2026250523834,
  author       = {Pith},
  title        = {Pith review of: Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GJ5NTDSC}},
  note         = {Machine review of arXiv:2505.23834}
}

read the original abstract

Lung sound classification is vital for early diagnosis of respiratory diseases. However, biomedical signals often exhibit inter-patient variability even among patients with the same symptoms, requiring a learning approach that considers individual differences. We propose a Patient-Aware Feature Alignment (PAFA) framework with two novel losses, Patient Cohesion-Separation Loss (PCSL) and Global Patient Alignment Loss (GPAL). PCSL clusters features of the same patient while separating those from other patients to capture patient variability, whereas GPAL draws each patient's centroid toward a global center, preventing feature space fragmentation. Our method achieves outstanding results on the ICBHI dataset with a score of 64.84\% for four-class and 72.08\% for two-class classification. These findings highlight PAFA's ability to capture individualized patterns and demonstrate performance gains in distinct patient clusters, offering broader applications for patient-centered healthcare.

Figures

Figures reproduced from arXiv: 2505.23834 by the authors.

**Figure 1.** Overview of the PAFA framework. The input audio is transformed into a Mel Spectrogram. The BEATs backbone encodes the spectrogram into feature embeddings. The projection head computes patient-aware losses (PCSL and GPAL). 2.2. Loss Functions The overall training objective comprises the standard crossentropy loss and two novel patient-aware losses designed to maintain patient-specific consistency while maintaining a… view at source ↗

**Figure 4.** presents another perspective by examining patients with the largest number of samples. In both CE-only and PAFA, these highly sampled patients still tend to reside within the main class-centric clusters, presumably because their abundant data ensures robust representation in either method [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 3.** t-SNE visualization focusing on training samples from a subset of patients who form new sub-clusters under PAFA. (Left) CE-only. (Right) PAFA. 4.3. Ablation Study on Patient-Aware Loss Components We conducted an ablation study to evaluate the separate contributions of PCSL and GPAL [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Adaptive Knowledge Distillation using a Device-Aware Teacher for Low-Complexity Acoustic Scene Classification
cs.SD 2025-09 conditional novelty 4.0 of 10

A low-complexity scene classifier trained by two-teacher knowledge distillation and device-specific fine-tuning reaches 57.93% accuracy on the DCASE 2025 development set.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · cited by 1 Pith paper

[1]

Feature Collapse

Introduction Lung sound analysis is essential in early respiratory disease de- tection, offering a non-invasive means to identify critical con- ditions such as chronic obstructive pulmonary disease (COPD), asthma, and pneumonia [1,2]. However, analyzing these sounds remains challenging due to the diverse patient-specific charac- teristics that can influen...

work page 2017
[2]

Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses

Method 2.1. PAFA Framework Our proposed PAFA framework builds upon the BEATs model as its backbone. A projection head is introduced to impose patient-centric constraints during training through two special- ized losses parallel to the primary classification head. Fig- ure 1 illustrates the complete architecture. By balancing the tra- ditional cross-entrop...

work page Pith review arXiv 2025
[3]

Dataset This study uses the ICBHI 2017 dataset [16], comprising 6,898 recordings from 126 subjects

Experimental Setup 3.1. Dataset This study uses the ICBHI 2017 dataset [16], comprising 6,898 recordings from 126 subjects. The data is categorized into four classes—Normal, Crackle, Wheeze, and Both—with sampling rates between 4 kHz and 44.1 kHz. The training set (4,142 samples) comprises 2,063 Normal, 1,215 Crackle, 501 Wheeze, and 363 Both recordings, ...

work page 2017
[4]

Performance on the ICBHI Dataset We evaluated our method on the ICBHI dataset using the of- ficial 60/40 (train/test) split

Result 4.1. Performance on the ICBHI Dataset We evaluated our method on the ICBHI dataset using the of- ficial 60/40 (train/test) split. Table 1 shows Sp, Se, and the overall Score for both four- and two-class tasks. Compared with recent approaches, our BEATs-based models achieve state- of-the-art Scores in both tasks. Specifically, for the four-class tas...

work page
[5]

Conclusion We introduced Patient-Aware Feature Alignment (PAFA), a framework that explicitly incorporates patient-centric con- straints into lung sound classification. By simultaneously ap- plying Patient Cohesion-Separation Loss (PCSL) and Global Patient Alignment Loss (GPAL), PAFA enforces both intra- patient cohesion and an overarching global structure...

work page
[6]

Acknowledgement This work was supported in part by the National Research Foun- dation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) of the Korean government under Grant RS-2023- 00221365 and in part by Seoul National University of Science and Technology

work page 2023
[7]

Computerized respiratory sounds in patients with copd: a systematic review,

C. Jacome and A. Marques, “Computerized respiratory sounds in patients with copd: a systematic review,” COPD: Journal of Chronic Obstructive Pulmonary Disease, vol. 12, no. 1, pp. 104– 112, 2015

work page 2015
[8]

Automatic detection of patient with respiratory diseases using lung sound analysis,

G. Chambres, P. Hanna, and M. Desainte-Catherine, “Automatic detection of patient with respiratory diseases using lung sound analysis,” in 2018 International Conference on Content-Based Multimedia Indexing (CBMI). IEEE, 2018, pp. 1–6

work page 2018

Show all 29 references

[9]

Intra-patient vari- ability in tacrolimus exposure: causes, consequences for clinical management,

N. Shuker, T. van Gelder, and D. A. Hesselink, “Intra-patient vari- ability in tacrolimus exposure: causes, consequences for clinical management,” Transplantation Reviews, vol. 29, no. 2, pp. 78–84, 2015

2015
[10]

Inter-relationships between inflammatory markers in patients with stable copd with bronchitis: intra-patient and inter- patient variability,

E. Sapey, D. Bayley, A. Ahmad, P. Newbold, N. Snell, and R. A. Stockley, “Inter-relationships between inflammatory markers in patients with stable copd with bronchitis: intra-patient and inter- patient variability,”Thorax, vol. 63, no. 6, pp. 493–499, 2008

2008
[11]

Lung sound classifi- cation using cepstral-based statistical features,

N. Sengupta, M. Sahidullah, and G. Saha, “Lung sound classifi- cation using cepstral-based statistical features,” Computers in Bi- ology and Medicine, vol. 75, pp. 118–129, 2016

2016
[12]

Automatic classification of lung sounds using machine learning algorithms,

A. Ullah, M. S. Khan, M. U. Khan, and F. Mujahid, “Automatic classification of lung sounds using machine learning algorithms,” in 2021 International Conference on Frontiers of Information Technology (FIT). IEEE, 2021, pp. 131–136

2021
[13]

Lung sounds classifica- tion using convolutional neural networks,

D. Bardou, K. Zhang, and S. M. Ahmad, “Lung sounds classifica- tion using convolutional neural networks,” Artificial Intelligence in Medicine, vol. 88, pp. 58–69, 2018

2018
[14]

Automated lung sound classification using a hybrid cnn-lstm network and fo- cal loss function,

G. Petmezas, G.-A. Cheimariotis, L. Stefanopoulos, B. Rocha, R. P. Paiva, A. K. Katsaggelos, and N. Maglaveras, “Automated lung sound classification using a hybrid cnn-lstm network and fo- cal loss function,” Sensors, vol. 22, no. 3, p. 1232, 2022

2022
[15]

Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classifi- cation,

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 classifi- cation,” in Interspeech 2023, 2023, pp. 5436–5440

2023
[16]

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 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024,...

2024
[17]

Pretraining respiratory sound rep- resentations using metadata and contrastive learning,

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

2023
[18]

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

2024
[19]

Prevalence of neural col- lapse during the terminal phase of deep learning training,

V . Papyan, X. Han, and D. L. Donoho, “Prevalence of neural col- lapse during the terminal phase of deep learning training,” Pro- ceedings of the National Academy of Sciences , vol. 117, no. 40, pp. 24 652–24 663, 2020

2020
[20]

BEATs: Audio pre-training with acoustic to- kenizers,

S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “BEATs: Audio pre-training with acoustic to- kenizers,” in Proceedings of the 40th International Conference on Machine Learning (ICML) , vol. 202. PMLR, 23–29 Jul 2023, pp. 5178–5193

2023
[21]

Audio set: An ontology and human-labeled dataset for audio events,

J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in 2017 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017,...

2017
[22]

A respiratory sound database for the development of auto- mated classification,

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

2017
[23]

Fisher linear discriminant analysis,

C. Li and B. Wang, “Fisher linear discriminant analysis,” CCIS Northeastern University, vol. 6, 2014

2014
[24]

Im- agenet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Im- agenet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2009, pp. 248–255

2009
[25]

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

Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” in 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE...

2023
[26]

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 Confer- ence of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 2021, pp. 527–530

2021
[27]

Lung sound classification using co-tuning and stochastic normalization,

T. Nguyen and F. Pernkopf, “Lung sound classification using co-tuning and stochastic normalization,” IEEE Transactions on Biomedical Engineering, vol. 69, no. 9, pp. 2872–2882, 2022

2022
[28]

PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,

Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumb- ley, “PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

2020
[29]

AST: Audio spectrogram transformer,

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

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Feature Collapse

Introduction Lung sound analysis is essential in early respiratory disease de- tection, offering a non-invasive means to identify critical con- ditions such as chronic obstructive pulmonary disease (COPD), asthma, and pneumonia [1,2]. However, analyzing these sounds remains challenging due to the diverse patient-specific charac- teristics that can influen...

work page 2017

[2] [2]

Patient-Aware Feature Alignment for Robust Lung Sound Classification:Cohesion-Separation and Global Alignment Losses

Method 2.1. PAFA Framework Our proposed PAFA framework builds upon the BEATs model as its backbone. A projection head is introduced to impose patient-centric constraints during training through two special- ized losses parallel to the primary classification head. Fig- ure 1 illustrates the complete architecture. By balancing the tra- ditional cross-entrop...

work page Pith review arXiv 2025

[3] [3]

Dataset This study uses the ICBHI 2017 dataset [16], comprising 6,898 recordings from 126 subjects

Experimental Setup 3.1. Dataset This study uses the ICBHI 2017 dataset [16], comprising 6,898 recordings from 126 subjects. The data is categorized into four classes—Normal, Crackle, Wheeze, and Both—with sampling rates between 4 kHz and 44.1 kHz. The training set (4,142 samples) comprises 2,063 Normal, 1,215 Crackle, 501 Wheeze, and 363 Both recordings, ...

work page 2017

[4] [4]

Performance on the ICBHI Dataset We evaluated our method on the ICBHI dataset using the of- ficial 60/40 (train/test) split

Result 4.1. Performance on the ICBHI Dataset We evaluated our method on the ICBHI dataset using the of- ficial 60/40 (train/test) split. Table 1 shows Sp, Se, and the overall Score for both four- and two-class tasks. Compared with recent approaches, our BEATs-based models achieve state- of-the-art Scores in both tasks. Specifically, for the four-class tas...

work page

[5] [5]

Conclusion We introduced Patient-Aware Feature Alignment (PAFA), a framework that explicitly incorporates patient-centric con- straints into lung sound classification. By simultaneously ap- plying Patient Cohesion-Separation Loss (PCSL) and Global Patient Alignment Loss (GPAL), PAFA enforces both intra- patient cohesion and an overarching global structure...

work page

[6] [6]

Acknowledgement This work was supported in part by the National Research Foun- dation of Korea (NRF) funded by the Ministry of Science and ICT (MSIT) of the Korean government under Grant RS-2023- 00221365 and in part by Seoul National University of Science and Technology

work page 2023

[7] [7]

Computerized respiratory sounds in patients with copd: a systematic review,

C. Jacome and A. Marques, “Computerized respiratory sounds in patients with copd: a systematic review,” COPD: Journal of Chronic Obstructive Pulmonary Disease, vol. 12, no. 1, pp. 104– 112, 2015

work page 2015

[8] [8]

Automatic detection of patient with respiratory diseases using lung sound analysis,

G. Chambres, P. Hanna, and M. Desainte-Catherine, “Automatic detection of patient with respiratory diseases using lung sound analysis,” in 2018 International Conference on Content-Based Multimedia Indexing (CBMI). IEEE, 2018, pp. 1–6

work page 2018