REVIEW 3 major objections 5 minor 15 references
Enhancing Lung Disease Diagnosis via Semi-Supervised Machine Learning
T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Semi-supervised learning modules added to an MFCC+CNN model raise six-class lung-sound classification accuracy on an open respiratory sound dataset from 89.1% to 92.9%.
desk verdict The evaluation is circular: the text says the final model is trained on the validation set, so the headline 92.9% is not an honest estimate of test accuracy. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The pipeline consists of an MFCC feature front end, a four-layer 2x2-filter CNN classifier with six output nodes, and three semi-supervised modules applied each epoch. Mix-Match blends labeled and unlabeled samples through Mix-Up augmentation and sharpened pseudo-labels; Co-Refinement trains the model on its own predictions for unlabeled data merged with labeled data; Co-Refurbishing mixes unlabeled predictions with a subset of labeled data to update model parameters. The open-access respiratory sound dataset used in the study supplies 920 recordings from 126 participants across six respiratory conditions, and the semi-supervised modules are what carry the claimed gain by turning unlabeled recordings into training signal.
What would settle it
Retrain the identical MFCC+CNN and semi-supervised pipeline with a strict split where the test recordings are withheld until after training, then compare the held-out accuracy to the reported 92.9% and 89.1% baselines.
Extended reading notes
Core claim
The central claim is that the combination of Mix-Match, Co-Refinement, and Co-Refurbishing with an MFCC+CNN backbone improves six-class respiratory sound classification from 89.1% to 92.9% accuracy. The paper reports that pneumonia precision and recall improve from 0.36 and 0.71 to 0.78 and 1.00, and that COPD, the most frequent class, reaches precision 0.97 and recall 1.00. Removing Co-Refinement drops accuracy to 89.7%, and removing Co-Refurbishing drops it to 90.7%, so the authors argue both modules contribute to the final gain. Classes with very few samples, bronchiectasis and bronchiolitis, improve less, which the paper attributes to the small number of training examples available for them.
Load-bearing premise
The load-bearing premise is that the 92.9% accuracy is measured on recordings the model never trained on, since the paper says the model is trained again on the full labeled and validation sets before evaluation.
Editorial extensions
If this is right
- If the gain is real, semi-supervised training can lift lung-sound classification by roughly four accuracy points without new labeled recordings.
- The per-class results suggest that classes with enough training samples, such as COPD, pneumonia, and healthy, benefit most, while very small classes remain hard to improve.
- The reported sensitivity to removing either Co-Refinement or Co-Refurbishing implies that the two modules are complementary rather than redundant.
- A working semi-supervised pipeline would lower the annotation burden for building lung-disease screening tools from stethoscope recordings.
Reading between the lines
- My inference: the 3.8-point gain should be treated with caution until the model is evaluated on a test split that never participates in training, because the paper's own wording suggests the validation set was used during training.
- My inference: the same module stack could be tested on other respiratory sound benchmarks or on spectrogram features besides MFCC to see whether the gain transfers.
- My inference: the poor results for bronchiectasis and bronchiolitis suggest that class-balanced sampling or synthetic augmentation, not just semi-supervision, would be needed to improve the smallest classes.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes combining an MFCC+CNN baseline with three existing semi-supervised learning modules—MixMatch, Co-Refinement, and Co-Refurbishing—for six-class lung sound classification on the ICHBI dataset. The abstract and conclusion claim that this combination raises accuracy from 89.1% to 92.9% (a 3.8 percentage-point gain), with per-class improvements for COPD, pneumonia, and healthy classes. Section 4 describes the experimental setup and Section 5 reports the evaluation and ablations.
Significance. If the reported accuracy were a valid held-out estimate, the result would provide a useful empirical demonstration that standard SSL modules can benefit lung-sound classification with limited labels. The paper does not propose a new method, but a careful evaluation of three existing modules on a medical audio task could be of interest to the community. The paper also has some strengths: it uses a public dataset (ICHBI) and reports confusion matrices and classification reports. However, the evaluation protocol as described invalidates the headline result, so the current manuscript does not establish its central claim.
major comments (3)
- [Section 4, paragraph 1] The evaluation is circular. The text states that after the semi-supervised loop, "the model is trained again, this time using the full labeled dataset and the validation dataset," and Section 5 states "we trained and validated on the labeled dataset." No held-out test set is defined or used. Consequently, the reported 92.9% accuracy and the +3.8% improvement over baseline are measurements on training data, not estimates of generalization. This is a load-bearing flaw because the central claim of the paper depends entirely on this number.
- [Sections 4-5] The paper never specifies the labeled/unlabeled split, the number of labeled samples per class, or how the unlabeled pool was selected. Without this information, the reader cannot verify that the semi-supervised setting is meaningful or that the claimed per-class gains (e.g., pneumonia recall from 0.71 to 1.0) are not artifacts of an unbalanced or inadvertently informative split.
- [Section 5, ablation paragraph] The ablation results (89.7% without Co-Refinement, 90.7% without Co-Refurbishing) are reported as point estimates with no measure of variance, no repeated runs, and no significance test. Given the very small sample sizes in some classes (13-16 total recordings), these differences are within plausible random variation, so the claim that the combination "leads to an improvement" is not supported.
minor comments (5)
- [Abstract] The phrase "feature insufficient labeled data" appears to be a typo; it should likely read "feature complexity and insufficient labeled data."
- [Abstract and Conclusion] The claimed "increase of 3.8%" should be stated as "3.8 percentage points" to avoid ambiguity.
- [Section 3.1] The CNN training hyperparameters (learning rate, batch size, number of epochs, optimizer) are not reported, nor is the MFCC configuration (e.g., number of filters, window length) beyond the coefficient count; this hinders reproducibility.
- [Section 5] The baseline accuracy (89.1%) appears only in the Conclusion; reporting it in Section 5 alongside the baseline classification report would improve clarity.
- [Figures 4-7] The figures are referenced without explanation of what is plotted (e.g., normalized or raw values); the captions are also minimal, making them difficult to interpret independently.
Circularity Check
Central accuracy claim is in-sample: the validation set is used in final training, so the 92.9% and +3.8% improvement are not held-out predictions.
-
fitted input called prediction
[Section 4 (Experiments) and Section 5 (Evaluation and Analysis); conclusion restates the numbers]
"Once the loop is completed, the model is trained again, this time using the full labeled dataset and the validation dataset. ... During the evaluation of the complete model, we trained and validated on the labeled dataset ... However, the final overall model achieves an accuracy of 92.9%."
The reported 92.9% accuracy is measured on data that the final training pass explicitly includes: the validation dataset is folded into training, and no held-out test split is described anywhere in the paper. The accuracy therefore measures the model's fit to the evaluation data itself, not its generalization to unseen recordings. The claimed +3.8% improvement over the 89.1% baseline is the difference between two in-sample fits, so the 'prediction' of improved accuracy reduces to memorization or transductive fitting of the very data used for evaluation. This is circular by construction: the evaluation target is an input to the final training run.
full rationale
The manuscript has no self-citation chain and no imported uniqueness theorem; the semi-supervised modules are standard external methods (MixMatch, Co-Refinement, Co-Refurbishing). The only load-bearing circularity is in the evaluation protocol. Section 4 says the model is 'trained again, this time using the full labeled dataset and the validation dataset,' and Section 5 says the authors 'trained and validated on the labeled dataset.' With no test set or split sizes given, the headline 92.9% and the +3.8% gain are not established as held-out generalization: they are accuracy on training/evaluation data. Co-Refinement's use of the model's own predictions as pseudo-labels is a standard self-training mechanism and is not, by itself, a circular derivation of the accuracy claim; the fatal circular step is the in-sample evaluation. If the authors add a genuine held-out test split, the same experiment could be valid, but as written the central result is not an independent prediction.
Assumptions & free parameters
free parameters (6)
- labeled/unlabeled data split =
not reported
- MFCC coefficient count =
40
- CNN architecture hyperparameters =
4 Conv2D layers, 2x2 filters, dropout 0.2, 2x2 stride-2 pooling
- labeled/unlabeled loss weight =
not reported
- training hyperparameters =
not reported
- MixMatch hyperparameters =
not reported
assumptions (5)
- domain assumption ICHBI labels are correct and the six classes are mutually exclusive
- domain assumption Unlabeled recordings come from the same distribution and class set as labeled recordings
- domain assumption Model-generated pseudo-labels on unlabeled data are accurate enough to improve training
- domain assumption The validation set is not used to train the final model
- ad hoc to paper The sequential order of MixMatch, Co-Refinement, then Co-Refurbishing within each epoch is beneficial
Cite this review
Pith. "Pith review of Enhancing Lung Disease Diagnosis via Semi-Supervised Machine Learning." pith.science (2026). https://pith.science/paper/I5DUI7EL
@misc{pith2026250716845,
author = {Pith},
title = {Pith review of: Enhancing Lung Disease Diagnosis via Semi-Supervised Machine Learning},
year = {2026},
howpublished = {\url{https://pith.science/paper/I5DUI7EL}},
note = {Machine review of arXiv:2507.16845}
}
read the original abstract
Lung diseases, including lung cancer and COPD, are significant health concerns globally. Traditional diagnostic methods can be costly, time-consuming, and invasive. This study investigates the use of semi supervised learning methods for lung sound signal detection using a model combination of MFCC+CNN. By introducing semi supervised learning modules such as Mix Match, Co-Refinement, and Co Refurbishing, we aim to enhance the detection performance while reducing dependence on manual annotations. With the add-on semi-supervised modules, the accuracy rate of the MFCC+CNN model is 92.9%, an increase of 3.8% to the baseline model. The research contributes to the field of lung disease sound detection by addressing challenges such as individual differences, feature insufficient labeled data.
Figures
Reference graph
Works this paper leans on
-
[1]
Ardakani, Ali Abb asian, et al. "Application of deep learning technique to manage COVID -19 in routine clinical practice using CT images: Results of 10 convolutional neural networks." Computers in Biology and Medicine (2020): 103795
work page 2020
-
[2]
Mixmatch: A ho listic approach to semi-supervised learning
Berthelot, David, et al. "Mixmatch: A ho listic approach to semi-supervised learning." Advances in neural information processing systems 32 (2019)
work page 2019
-
[3]
Dunne, Rob, Tim Morris, and Simon Harper. "High accuracy classification of COVID -19 coughs using Mel-frequency cepstral coefficients and a convoluti onal neural network with a use case for smart home devices." (2020). 175
work page 2020
-
[4]
Unsupervised clustering using pseudo -semi-supervised learning
Gupta, Divam, et al. "Unsupervised clustering using pseudo -semi-supervised learning." International Conference on Learning Representations. 2020
work page 2020
-
[5]
CNN architectur es for large-scale audio classification
Hershey, Shawn, et al. "CNN architectur es for large-scale audio classification." 2017 ieee international conference on acoustics, speech and signal processing (icassp). IEEE, 2017
work page 2017
-
[6]
An automatic approach based on CNN architecture to detect Covid -19 disease from chest X -ray images
Hira, Swati, Anita Bai, and Sanchit Hira. "An automatic approach based on CNN architecture to detect Covid -19 disease from chest X -ray images." Applied Intelligence 51 (2021): 2864 - 2889
work page 2021
-
[7]
Graph -based semi - supervised one class support vector machine for detecting abnormal lung sounds
Lang, Rongling, et al. "Graph -based semi - supervised one class support vector machine for detecting abnormal lung sounds." Applied Mathematics and Computation 364 (2020): 124487
work page 2020
-
[8]
Analysis of unlabeled lung sound samples using semi -supervised convolutional neural networks
Lang, Rongling, et al. "Analysis of unlabeled lung sound samples using semi -supervised convolutional neural networks." Applied Mathematics and Computation 411 (2021): 126511
work page 2021
Show all 15 references
-
[9]
Dividemix: Learning with noisy labels as semi - supervised learning
Li, Junnan, Richard Socher, and Steven CH Hoi. "Dividemix: Learning with noisy labels as semi - supervised learning." arXiv preprint arXiv:2002.07394 (2020)
2020 arXiv
-
[10]
Improving unsupervised image clustering with robust learning
Park, Sungwon, et al. "Improving unsupervised image clustering with robust learning." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021
2021
-
[11]
An open access database for the evaluation of respiratory sound classification algorithms
Rocha, Bruno M., et al. "An open access database for the evaluation of respiratory sound classification algorithms." Physiological measurement 40.3 (2019): 035001
2019
-
[12]
Selfie: Refurbishing unclean samples for robu st deep learning
Song, Hwanjun, Minseok Kim, and Jae -Gil Lee. "Selfie: Refurbishing unclean samples for robu st deep learning." International Conference on Machine Learning. PMLR, 2019
2019
-
[13]
A survey on semi -supervised learning
Van Engelen, Jesper E., and Holger H. Hoos. "A survey on semi -supervised learning." Machine learning 109.2 (2020): 373-440
2020
-
[14]
Scan: Learning to classify images without labels
Van Gansbeke, Wouter, et al. "Scan: Learning to classify images without labels." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part X. Cham: Springer International Publishing, 2020
2020
-
[15]
World health statistics 2020: monitoring health for the SDGs sustainable development goals
World Health Organization. World health statistics 2020: monitoring health for the SDGs sustainable development goals. World Health Organization, 2020. 176
2020
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.