Deep Learning-Based Classification of Common Lung Sounds via Auto-Detected Respiratory Cycles

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for submitting this interesting manuscript on deep learning-based classification of lung sounds. The work is well organized and addresses a clinically relevant problem.

Major Comments

1. Novelty and Contribution Need Clearer Articulation

The manuscript does not clearly state how it advances beyond existing high-accuracy models in the literature. Please specify the unique contribution (automatic cycle segmentation, dataset balancing, gammatonegram + DenseNet201 combination, etc.) and explicitly compare novelty with recent studies.

2. Insufficient Dataset Description

More information is needed on patient demographics, recording environment, comorbidities, and clinical context. The manuscript should clarify inclusion/exclusion criteria and provide a demographic summary table.

3. Automatic Respiratory Cycle Detection Requires More Technical Detail

The segmentation algorithm is a core component, yet insufficiently described. Please elaborate on thresholds, boundary detection criteria, error correction, and whether segmentation outputs were manually validated.

4. Risk of Overfitting Not Adequately Addressed

With only 400 cycles, a 97% accuracy suggests possible overfitting. Provide training/validation curves, explanation of regularization strategies, and variance across multiple training runs.

5. Lack of Statistical Comparison Between Models

The analysis lists accuracies, but statistical significance testing (e.g., McNemar test, CI for accuracy) is missing. Adding confidence intervals or comparative statistics would improve scientific rigor.

6. Too Small Test Set

Each class has only 10 test samples. It’s very small. This severely limits generalization. Please justify this choice and consider evaluation with cross-validation or larger test splits.

 

⸻

 

Minor Comments

1. Figures (especially Figures 1–4) need higher resolution.

2. Some terms are used inconsistently (“rhonchi/ronchi,” “coarse crackle/coarse wheeze”).

3. Table 5 formatting is difficult to follow; please reorganize or split.

4. Several line-break and spacing errors appear throughout the text—please edit for readability.

5. Expand the Discussion with more clinical relevance and potential real-world implementation pathways.

Author Response

Major Comments

Comments 1. [Novelty and Contribution Need Clearer Articulation]. The manuscript does not clearly state how it advances beyond existing high-accuracy models in the literature. Please specify the unique contribution (automatic cycle segmentation, dataset balancing, gammatonegram + DenseNet201 combination, etc.) and explicitly compare novelty with recent studies.

Response 1: Thank you for your valuable comment. We agree that the original aspects of the study need to be expressed more clearly. In the revised article, we have explicitly stated the innovations of our study and the progress it has made compared to recent high-accuracy studies in the literature at the end of the Introduction.

Comments 2.  [Insufficient Dataset Description] More information is needed on patient demographics, recording environment, comorbidities, and clinical context. The manuscript should clarify inclusion/exclusion criteria and provide a demographic summary table.

Response 2: Thank you for your valuable comment. We have added the text below to Table 1 in the revised article.

Comments 3. [Automatic Respiratory Cycle Detection Requires More Technical Detail] The segmentation algorithm is a core component, yet insufficiently described. Please elaborate on thresholds, boundary detection criteria, error correction, and whether segmentation outputs were manually validated.

Response 3: Thank you for highlighting the importance of the segmentation algorithm. In the revised manuscript, we have included detailed explanations of the basic steps of automated respiratory cycle segmentation in section 2.1. Data acquisition.

Comments 4. [Risk of Overfitting Not Adequately Addressed] With only 400 cycles, a 97% accuracy suggests possible overfitting. Provide training/validation curves, explanation of regularization strategies, and variance across multiple training runs.

Response 4: We thank the reviewer for raising this important concern regarding potential overfitting. In response, we have revised the manuscript to explicitly address all three points. First, training and validation learning curves for both accuracy and loss have been added as Figure 8. These curves demonstrate stable convergence with closely aligned training and validation performance, indicating the absence of severe overfitting. The epoch corresponding to the minimum validation loss, selected via early stopping, is explicitly highlighted. Second, the regularization strategies employed in the proposed framework are now clearly described in the Methods section. These include early stopping with a patience of 15 epochs, dropout with a rate of 0.5, L2 regularization in fully connected layers, and partial fine-tuning of ImageNet-pretrained CNN backbones, where only the upper layers are trainable. Third, to quantify performance variability, all classification experiments were repeated ten times using different random data partitions. The results are reported as mean accuracy and standard deviation. In addition, the robustness of the most successful gammatonegram–DenseNet201 model under different data splitting ratios is analyzed in Table 6. The observed increase in variance under more stringent splits further reflects expected statistical behavior rather than memorization. Taken together, the learning curves, explicit regularization strategies, and low variance across repeated runs support that the reported high accuracy reflects stable generalization rather than overfitting.

Comments 5. [Lack of Statistical Comparison Between Models] The analysis lists accuracies, but statistical significance testing (e.g., McNemar test, CI for accuracy) is missing. Adding confidence intervals or comparative statistics would improve scientific rigor.

Response 5: We thank the reviewer for this suggestion. To address uncertainty in the reported accuracies, all experiments were repeated ten times, and results are reported as mean accuracy and standard deviation, providing an empirical measure of performance variability.

Comments 6. [Too Small Test Set] Each class has only 10 test samples. It’s very small. This severely limits generalization. Please justify this choice and consider evaluation with cross-validation or larger test splits.

Response 6: We thank the reviewer for this important comment. The 80/10/10 data split was selected to maintain a balance between training data sufficiency and independent testing, given the limited size of the dataset. To mitigate the effect of the small test set, all experiments were repeated ten times, and results are reported as mean accuracy and standard deviation. In addition, the robustness of the most successful model was further evaluated under larger test splits (up to 40%), as reported in Table 6. These analyses collectively support the generalization capability of the proposed approach despite the limited number of samples per class.

Minor Comments

Comments 1. Figures (especially Figures 1–4) need higher resolution.

Response 1: The figures mentioned have been updated.

Comments 2. Some terms are used inconsistently (“rhonchi/ronchi,” “coarse crackle/coarse wheeze”).

Response 2: The entire text has been reviewed and the term ‘rhonchi’ has been used consistently, with the use of ‘ronchi’ removed. “fine/coarse crackles”. Similarly, the incorrect usage in the text has been corrected to the correct form, ‘fine/coarse crackles’.

Comments 3. Table 5 formatting is difficult to follow; please reorganize or split.

Response 3: Table 5 has been made more followable by changing the border style.

Comments 4. Several line-break and spacing errors appear throughout the text—please edit for readability.

Response 4: The entire text has been reviewed.

Comments 5. Expand the Discussion with more clinical relevance and potential real-world implementation pathways.

Response 5: The discussion section has been expanded.

 

 

Reviewer 2 Report

Comments and Suggestions for Authors

The article presents the classification of lung sounds (LFS) using CNNs after converting them into various time-frequency representations such as spectrograms, scalograms, Mel-spectrograms, and gammatonegrams. Furthermore, hybrid architectures of Support Vector Machines (SVM) and Long Short-Term Memory (LSTM) were used in the classification process. There are a total of four classes, each with 100 data points. In addition to observing changes in the time domain, it is valuable to examine the changes in the frequency domain of the signal. I have a few suggestions:

1. From the confusion matrix, it appears that 10% of the data was used for testing. The test data should be 20% or 30%. Using 10% of the test data results in a 2.5% change in accuracy. This rate is very high. Using 40 data points for testing hinders the generalization of the results.

2. The loss function change for the model that yielded the best results should be included in the article. 

3. Figure 4 shows that not all signals are the same length. Does this affect the results? Although the signals are of different lengths, the same size image (feature) is obtained for each signal.

 4. Why wasn't data augmentation used? More labeled data from a single patient could have been obtained by choosing a fixed window length.

Author Response

Comments 1. From the confusion matrix, it appears that 10% of the data was used for testing. The test data should be 20% or 30%. Using 10% of the test data results in a 2.5% change in accuracy. This rate is very high. Using 40 data points for testing hinders the generalization of the results.

Response 1: We thank the reviewer for this observation. While a 10% test split results in discrete accuracy steps due to the limited number of test samples, this configuration was initially adopted to preserve sufficient training data given the small dataset size. To address concerns regarding generalization, we have extended the evaluation to include larger test splits (20%, 30%, and 40%), and the corresponding results are reported in Table 6. The observed performance trends remain consistent across these splits, indicating that the reported results are not dependent on a single test partition and supporting the robustness of the proposed method.

Comments 2. The loss function change for the model that yielded the best results should be included in the article. 

Response 2: We thank the reviewer for this comment. The loss function used for the best-performing model has now been explicitly stated in the revised manuscript in Figure 8.

Comments 3. Figure 4 shows that not all signals are the same length. Does this affect the results? Although the signals are of different lengths, the same size image (feature) is obtained for each signal.

Response 3: We thank the reviewer for this observation. Although the respiratory signals have different temporal lengths, all signals are transformed into fixed-size feature images during the preprocessing stage. Therefore, the variability in signal length does not affect the model input or the classification results, as the network always receives features with identical dimensions. The text has been added under Figure 4 in the revised article.

Comments 4. Why wasn't data augmentation used? More labeled data from a single patient could have been obtained by choosing a fixed window length.

Response 4: We thank the reviewer for this valuable suggestion. While fixed-length windowing could increase the number of labeled samples from a single subject, we deliberately avoided this form of data augmentation to prevent generating highly correlated samples that may inflate performance. Instead, we focused on physiologically meaningful respiratory cycle segmentation and enforced strict subject-level separation between training and test sets.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript titled “Deep Learning-Based Classification of Common Lung Sounds via Auto-Detected Respiratory Cycles” presents a deep learning pipeline for classification of single-channel lung sounds into four classes: normal, rhonchi, fine crackles, and coarse crackles. Recordings from 94 subjects are segmented into respiratory cycles by an automatic method, then converted into time-frequency images (spectrogram, scalogram, Mel-spectrogram, gammatonegram). Pre-trained CNN backbones with transfer learning extract features, and three classifier heads (CNN, CNN-LSTM, CNN-SVM) are compared. The best result uses gammatonegrams with DenseNet201 and a CNN classifier, with 97% accuracy on the test set.

The topic fits the scope of the journal and the manuscript shows solid engineering work. The main concerns relate to dataset size, evaluation design, risk of overfitting, and a somewhat incremental level of novelty.

1. Novelty and contribution

Automatic cycle detection, balanced class counts, and comparison of several time-frequency transforms form a useful contribution. The use of gammatonegrams with DenseNet201 for four-way classification is interesting, and Table 6 gives a helpful comparison with several recent studies.

Many core elements already appear in prior work. Your previous paper [19] already combines automatic cycle detection with classical features and classifiers. Gupta et al. [15] already use gammatonegrams with deep CNNs and transfer learning. The present study seems to move from hand-crafted features to deep features, add several time-frequency variants, and apply this to a new local dataset.

I suggest you sharpen the contribution statement. Spell out, in the Introduction and again in the Conclusion, exactly what is new compared with [15], [19], [21], [22], and [23]. For example, is the key advance the combo of auto-detected cycles, gammatonegrams, and DenseNet201 on four common lung sound types, or a systematic benchmark of several image transforms on a balanced dataset? Right now that point feels vague.

2. Dataset description and potential bias

Table 1 describes a dataset of 400 respiratory cycles from 94 subjects, with 100 cycles per class. Thirty subjects contribute normal sounds, and 23, 20, and 21 subjects contribute rhonchi, fine crackles, and coarse crackles, respectively. This balances cycle counts but not subject counts.

This raises questions that affect generalisation:

How many cycles does each subject contribute on average, and what is the range?

Do some subjects contribute to more than one class (mixed pathology)?

Are the 100 cycles per class selected in a way that avoids heavy dominance by a few subjects?

You briefly note that wheeze sounds are excluded. They can occur in healthy subjects during forced expiration and can be audible without a stethoscope. It would help to expand the clinical reasoning here. List typical diagnoses for each class in your dataset and explain how a four-class model without wheezes would support real clinical decisions.

The recording protocol also needs more detail. Please specify:

subject demographics (age range, sex distribution, any key comorbidities)

clinical diagnoses linked to each sound class

body position during recording and auscultation locations (number and placement of sites)

sampling rate, bit depth, and approximate recording length per subject

setting (quiet room, bedside, emergency department, etc.)

These details are important for judging external validity and for readers who want to replicate or extend your work.

3. Segmentation and pre-processing

Figure 2 on page 6 shows the automatic respiratory cycle detection pipeline, with overlapping frames, Hamming window, NFFT, a band-pass filter (80–1000 Hz), spectral energy density, smoothing, and dynamic time warping. The text points to [27] for more detail. That helps, but reproducibility of the present study still needs explicit parameters.

Please add in the Methods:

sampling rate of raw recordings

frame length and overlap used in segmentation

design of the 80–1000 Hz band-pass filter (filter type and order)

wavelet family and scale range used in the discrete wavelet transform

dynamic time warping settings (distance measure, constraints)

minimum and maximum allowed cycle length and handling of outliers

A short justification for the 80 Hz lower cutoff would be useful, since some work keeps more low-frequency content for crackles and rhonchi. A sentence that links your choice to prior studies would be enough.

4. Time-frequency transforms

Figure 4 on page 7 compares amplitude plots, spectrograms, scalograms, Mel-spectrograms, and gammatonegrams for all four classes. The visual differences are clear and support the image-based strategy. Section 2.2 explains each transform at a conceptual level, but practical details are missing.

For each transform, please specify:

image size passed to the CNNs (for example 224 × 224 pixels)

window length, overlap, and FFT size for spectrograms and Mel-spectrograms

number of Mel filters and frequency range

mother wavelet and scale range for scalograms

number of gammatone filters, center frequencies, and spacing

use of log-magnitude or dB scaling

any per-image normalization applied before feeding the CNNs

The caption of Figure 4 could mention the key parameter values or refer clearly to a subsection with these details. A brief note on computational cost of each transform would help readers who plan real-time or embedded systems.

5. Experimental design and data splitting

Section 2.4 states that the dataset is split into 80% training, 10% validation, and 10% testing, with 100 epochs, early stopping with patience 15, a batch size of 16, and categorical cross-entropy with Adam. The text claims that this avoids overlap between training and testing, but the unit of splitting is not stated.

From Table 4, the confusion matrix of the best method, each row sums to 10. That implies only 10 test cycles per class, so 40 test samples in total. If cycles from the same subject appear in both the training and test sets, subject-specific patterns can inflate accuracy.

I strongly recommend that you:

State clearly whether splitting is done at the subject level or at the cycle level.

If the split is at the cycle level, repeat experiments with subject-wise cross-validation (for example leave-one-subject-out or k-fold by subject) and report mean ± standard deviation.

For each configuration in Table 3, run several trials with different random seeds and report mean ± standard deviation, not only a single number.

The small test size also means that one misclassified cycle changes accuracy by 2.5 percentage points. A subject-wise cross-validation scheme will give more stable and credible estimates.

6. Model complexity and risk of overfitting

DenseNet201 is a large CNN with many parameters, built for ImageNet-scale data. Your training set contains only 320 cycles. Even with transfer learning and early stopping, the 97% test accuracy and near-perfect confusion matrix raise concern about overfitting.

To address this, please:

report learning curves for training and validation accuracy or loss

state exactly which layers are frozen for each backbone and classifier variant

give the number of trainable parameters for each model (CNN, CNN-LSTM, CNN-SVM)

describe any regularisation (L2 penalties, dropout rates) used in the CNN-only case

You write that only the last 100 layers are trainable in the CNN-LSTM variant. A short rationale for this choice and, if possible, a small ablation (for example last 50 vs last 100 vs last 150 layers) would strengthen the method section.

Given the dataset size, a comparison of performance versus parameter count would be very informative. It might turn out that a smaller backbone, such as MobileNetV2, offers similar accuracy with a better trade-off.

7. Evaluation metrics and clinical interpretation

The paper reports accuracy, precision, recall, and F1 scores per class. This is useful for method comparison. For clinical readers, a stronger link to sensitivity and specificity for key clinical questions would help.

You could define and evaluate binary tasks such as:

normal vs any abnormal sound

crackles (fine + coarse) vs non-crackle

rhonchi vs non-rhonchi

Reporting ROC curves and AUC values for these tasks would give extra insight. A table may suffice; full plots are optional.

The test set is small, so confidence intervals for accuracy and for per-class recall would convey uncertainty more clearly. Simple binomial confidence intervals already help.

A subject-level evaluation would be valuable. For example, apply majority voting across cycles for each subject and report subject-level accuracy. That matches real use, where clinicians listen to several cycles per subject.

8. Details of CNN-LSTM and CNN-SVM pipelines

The description of the CNN-LSTM and CNN-SVM models in Section 2.4 reads quite densely. A clearer explanation and a schematic would help readers.

For CNN-LSTM you state that:

a 7 × 7 × 1920 feature map is extracted

this map is reshaped into 9 time steps of 64-dimensional vectors

a bidirectional LSTM with 256 neurons processes the sequence

a dense layer with 512 neurons, L2 regularisation, and 50% dropout follows

a softmax layer outputs four class probabilities

Please confirm and clarify how 7 × 7 × 1920 turns into 9 × 64. Some pooling or grouping is happening, but the text does not spell it out. A simple block diagram, similar to Figure 5 on page 8, would be very helpful.

For CNN-SVM, state the SVM kernel type (linear, RBF, etc.), the values or search ranges for key hyperparameters, and how you tune them. Right now the description of SVM use feels very brief.

9. Comparison with related work

Table 6 provides a useful summary of deep learning studies on lung sound classification, including sample sizes, channel types, segmentation strategies, features, and performance. This is a strong point of the paper.

The metrics across studies are not uniform. Some works on the ICBHI dataset report the challenge score, which mixes sensitivity and specificity and differs from plain accuracy. Others report AUC. Your own dataset uses a different class set and excludes wheezes.

Please add a short explanation of the ICBHI score and state explicitly that it is not directly comparable to your accuracy numbers. Then frame your claim more carefully, for example as “high accuracy on a balanced four-class single-channel dataset with auto-detected cycles”, rather than direct outperformance of multi-channel or larger datasets.

10. Data availability

You state that the data are not openly available for reasons of sensitivity, and can be obtained from the corresponding author on reasonable request. Open data are valuable in this field, since public lung sound datasets remain limited.

If full release of audio is not possible, consider:

sharing time-frequency images and labels in anonymised form

releasing a smaller anonymised subset of audio

describing the request process in more detail (required approvals, typical response time, data format)

This would make the work more useful for the community and align well with the journal’s emphasis on reproducibility.

Author Response

Comments 1. [Novelty and contribution] The present study seems to move from hand-crafted features to deep features, add several time-frequency variants, and apply this to a new local dataset. I suggest you sharpen the contribution statement. Spell out, in the Introduction and again in the Conclusion, exactly what is new compared with [15], [19], [21], [22], and [23]. For example, is the key advance the combo of auto-detected cycles, gammatonegrams, and DenseNet201 on four common lung sound types, or a systematic benchmark of several image transforms on a balanced dataset? Right now that point feels vague.

Response 1: Thank you for your valuable comment. We agree that the original aspects of the study need to be expressed more clearly. In the revised article, we have explicitly stated the innovations of our study and the progress it has made compared to recent high-accuracy studies in the literature at the end of the Introduction.

Comments 2. [Dataset description and potential bias] Table 1 describes a dataset of 400 respiratory cycles from 94 subjects, with 100 cycles per class. Thirty subjects contribute normal sounds, and 23, 20, and 21 subjects contribute rhonchi, fine crackles, and coarse crackles, respectively. This balances cycle counts but not subject counts. This raises questions that affect generalisation: How many cycles does each subject contribute on average, and what is the range? Do some subjects contribute to more than one class (mixed pathology)? Are the 100 cycles per class selected in a way that avoids heavy dominance by a few subjects? You briefly note that wheeze sounds are excluded. They can occur in healthy subjects during forced expiration and can be audible without a stethoscope. It would help to expand the clinical reasoning here. List typical diagnoses for each class in your dataset and explain how a four-class model without wheezes would support real clinical decisions. The recording protocol also needs more detail. Please specify: subject demographics (age range, sex distribution, any key comorbidities) clinical diagnoses linked to each sound class body position during recording and auscultation locations (number and placement of sites) sampling rate, bit depth, and approximate recording length per subject setting (quiet room, bedside, emergency department, etc.) These details are important for judging external validity and for readers who want to replicate or extend your work.

Response 2: Thank you for your valuable comment. Although the number of respiratory cycles is balanced across classes in Table 1, we acknowledge that the number of patients per class may vary depending on the distribution of clinical visits. In the revised manuscript, we specified the respiratory cycle statistics per subject (3–7 cycles per recording) and clarified that a subject-level contribution limit was applied to prevent a small number of individuals from dominating the balanced selection of 100 cycles per class. We also clarified that each individual was assigned to a single target class based on physician labeling according to the predominant sound type, and that mixed or ambiguous cases were not included in the balanced subset. The wheeze class was excluded from the scope of this study because wheeze sounds can also be observed during forced expiration in otherwise healthy individuals and can be sufficiently distinct to be audible without a stethoscope. In preliminary experiments, inclusion of this outlier sound group was observed to inflate overall performance metrics, potentially confounding clinically meaningful comparisons. To strengthen the description of the recording protocol, we provided detailed information on subject demographics, recording environment, and auscultation position. Section 2.1 (Data Acquisition) has been revised.

Comments 3. [Segmentation and pre-processing] Figure 2 on page 6 shows the automatic respiratory cycle detection pipeline, with overlapping frames, Hamming window, NFFT, a band-pass filter (80–1000 Hz), spectral energy density, smoothing, and dynamic time warping. The text points to [27] for more detail. That helps, but reproducibility of the present study still needs explicit parameters. Please add in the Methods: sampling rate of raw recordings frame length and overlap used in segmentation design of the 80–1000 Hz band-pass filter (filter type and order) wavelet family and scale range used in the discrete wavelet transform dynamic time warping settings (distance measure, constraints) minimum and maximum allowed cycle length and handling of outliers A short justification for the 80 Hz lower cutoff would be useful, since some work keeps more low-frequency content for crackles and rhonchi. A sentence that links your choice to prior studies would be enough.

Response 3: We thank the reviewer for emphasizing reproducibility. In response, we have added a concise summary of all key parameters of the automatic respiratory cycle detection pipeline to the Methods section. These parameters are identical to those defined in our previous study [27], on which the current work directly builds. Specifically, the sampling rate, frame length and overlap, band-pass filter design, wavelet configuration, dynamic time warping settings, respiratory cycle duration constraints, and outlier handling strategy are now explicitly stated. In addition, a brief justification for the 80 Hz lower cutoff frequency has been included, consistent with prior lung sound analysis studies that suppress low-frequency motion artifacts and heart sounds while preserving the dominant spectral content of crackles and rhonchi. Section 2.1 (Data Acquisition) has been revised.

Comments 4. [Time-frequency transforms] Figure 4 on page 7 compares amplitude plots, spectrograms, scalograms, Mel-spectrograms, and gammatonegrams for all four classes. The visual differences are clear and support the image-based strategy. Section 2.2 explains each transform at a conceptual level, but practical details are missing. For each transform, please specify: image size passed to the CNNs (for example 224 × 224 pixels) window length, overlap, and FFT size for spectrograms and Mel-spectrograms number of Mel filters and frequency range mother wavelet and scale range for scalograms number of gammatone filters, center frequencies, and spacing use of log-magnitude or dB scaling any per-image normalization applied before feeding the CNNs The caption of Figure 4 could mention the key parameter values or refer clearly to a subsection with these details. A brief note on computational cost of each transform would help readers who plan real-time or embedded systems.

Response 4: We thank the reviewer for this constructive comment. In the revised manuscript, we explicitly added the practical parameter settings for each time–frequency representation directly within the corresponding subsections of Section 2.2. This includes the fixed image size used for CNN input (256 × 256 × 3), as well as window length, overlap, and FFT size for spectrograms and Mel-spectrograms; the number of Mel filters and frequency range; the wavelet configuration for scalograms; and the gammatone filterbank parameters. A brief theoretical comparison of the relative computational complexity of the feature representations has been added to the manuscript.

Comments 5. [Experimental design and data splitting] Section 2.4 states that the dataset is split into 80% training, 10% validation, and 10% testing, with 100 epochs, early stopping with patience 15, a batch size of 16, and categorical cross-entropy with Adam. The text claims that this avoids overlap between training and testing, but the unit of splitting is not stated. From Table 4, the confusion matrix of the best method, each row sums to 10. That implies only 10 test cycles per class, so 40 test samples in total. If cycles from the same subject appear in both the training and test sets, subject-specific patterns can inflate accuracy. I strongly recommend that you: State clearly whether splitting is done at the subject level or at the cycle level. If the split is at the cycle level, repeat experiments with subject-wise cross-validation (for example leave-one-subject-out or k-fold by subject) and report mean ± standard deviation. For each configuration in Table 3, run several trials with different random seeds and report mean ± standard deviation, not only a single number. The small test size also means that one misclassified cycle changes accuracy by 2.5 percentage points. A subject-wise cross-validation scheme will give more stable and credible estimates.

Response 5: We thank the reviewer for this important clarification. In the present study, data splitting was performed at the respiratory cycle level rather than at the subject level. This choice was made to ensure sufficient training data for deep learning models given the limited dataset size. We agree that cycle-level splitting may allow subject-specific characteristics to appear in different subsets and potentially influence classification performance. However, respiratory cycles recorded from the same individual may still exhibit substantial intra-subject variability. Nevertheless, this does not fully eliminate the possibility of subject-specific bias, which is acknowledged as a limitation of the present study. This has now been explicitly stated in the revised manuscript. To mitigate this effect, all experiments were repeated ten times using different random partitions, and results are reported as mean accuracy and standard deviation. In addition, the robustness of the best-performing model was further evaluated under larger test splits (up to 40%), as reported in Table 6, showing consistent performance trends. The limitation of not employing subject-wise cross-validation has now been explicitly acknowledged in the manuscript, and subject-wise evaluation schemes such as leave-one-subject-out cross-validation will be considered in future work to further strengthen generalization assessment.

Comments 6. [Model complexity and risk of overfitting] DenseNet201 is a large CNN with many parameters, built for ImageNet-scale data. Your training set contains only 320 cycles. Even with transfer learning and early stopping, the 97% test accuracy and near-perfect confusion matrix raise concern about overfitting. To address this, please: report learning curves for training and validation accuracy or loss state exactly which layers are frozen for each backbone and classifier variant give the number of trainable parameters for each model (CNN, CNN-LSTM, CNN-SVM) describe any regularisation (L2 penalties, dropout rates) used in the CNN-only case You write that only the last 100 layers are trainable in the CNN-LSTM variant. A short rationale for this choice and, if possible, a small ablation (for example last 50 vs last 100 vs last 150 layers) would strengthen the method section. Given the dataset size, a comparison of performance versus parameter count would be very informative. It might turn out that a smaller backbone, such as MobileNetV2, offers similar accuracy with a better trade-off.

Response 6: We thank the reviewer for the valuable comments regarding model complexity and potential overfitting. In the revised manuscript, training and validation learning curves have been added to demonstrate stable convergence and closely aligned performance. The fine-tuning strategy is now explicitly stated: in the CNN-only and CNN–SVM configurations, the pretrained CNN backbone is fully frozen and used strictly as a fixed feature extractor, whereas in the CNN–LSTM configuration only the last 100 layers of DenseNet201 are trainable. Regularization strategies are also clarified; dropout (0.5) and L2 regularization are applied in the CNN–LSTM model, while overfitting control in the CNN-only case relies on transfer learning and early stopping. Although an explicit ablation on fine-tuning depth was not performed, a comparative evaluation across multiple backbones, including smaller models such as MobileNetV2, is already reported, providing insight into performance versus model capacity under limited data conditions.

Comments 7. [Evaluation metrics and clinical interpretation] The paper reports accuracy, precision, recall, and F1 scores per class. This is useful for method comparison. For clinical readers, a stronger link to sensitivity and specificity for key clinical questions would help. You could define and evaluate binary tasks such as: normal vs any abnormal sound crackles (fine + coarse) vs non-crackle rhonchi vs non-rhonchi Reporting ROC curves and AUC values for these tasks would give extra insight. A table may suffice; full plots are optional. The test set is small, so confidence intervals for accuracy and for per-class recall would convey uncertainty more clearly. Simple binomial confidence intervals already help. A subject-level evaluation would be valuable. For example, apply majority voting across cycles for each subject and report subject-level accuracy. That matches real use, where clinicians listen to several cycles per subject.

Response 7: We thank the reviewer for emphasizing the importance of clinical interpretability. In response, we have expanded the Discussion section by explicitly interpreting the confusion matrix of the best-performing model from a clinical perspective. The added text links class-wise recall and precision to clinical sensitivity and specificity, and highlights the model’s ability to reliably distinguish normal from abnormal respiratory cycles. In addition, misclassifications between acoustically similar classes (e.g., fine and coarse crackles) are discussed in the context of clinical ambiguity. We believe that this confusion-matrix-based interpretation strengthens the clinical relevance of the proposed method without introducing unreliable statistics given the limited test set size.

Comments 8. [Details of CNN-LSTM and CNN-SVM pipelines] The description of the CNN-LSTM and CNN-SVM models in Section 2.4 reads quite densely. A clearer explanation and a schematic would help readers. For CNN-LSTM you state that: a 7 × 7 × 1920 feature map is extracted this map is reshaped into 9 time steps of 64-dimensional vectors a bidirectional LSTM with 256 neurons processes the sequence a dense layer with 512 neurons, L2 regularisation, and 50% dropout follows a softmax layer outputs four class probabilities Please confirm and clarify how 7 × 7 × 1920 turns into 9 × 64. Some pooling or grouping is happening, but the text does not spell it out. A simple block diagram, similar to Figure 5 on page 8, would be very helpful. For CNN-SVM, state the SVM kernel type (linear, RBF, etc.), the values or search ranges for key hyperparameters, and how you tune them. Right now the description of SVM use feels very brief.

Response 8: Thank you for the constructive comment. Section 2.4 has been revised to improve clarity and readability of both the CNN–LSTM and CNN–SVM models.

Comments 9. [Comparison with related work] Table 6 provides a useful summary of deep learning studies on lung sound classification, including sample sizes, channel types, segmentation strategies, features, and performance. This is a strong point of the paper. The metrics across studies are not uniform. Some works on the ICBHI dataset report the challenge score, which mixes sensitivity and specificity and differs from plain accuracy. Others report AUC. Your own dataset uses a different class set and excludes wheezes. Please add a short explanation of the ICBHI score and state explicitly that it is not directly comparable to your accuracy numbers. Then frame your claim more carefully, for example as “high accuracy on a balanced four-class single-channel dataset with auto-detected cycles”, rather than direct outperformance of multi-channel or larger datasets.

Response 9: We appreciate this valuable comment. In the revised version, we have added a brief explanation in Table 6, specifically regarding the structure of the ICBHI challenge score and why it cannot be directly compared with accuracy.

Comments 10. [Data availability] You state that the data are not openly available for reasons of sensitivity, and can be obtained from the corresponding author on reasonable request. Open data are valuable in this field, since public lung sound datasets remain limited. If full release of audio is not possible, consider: sharing time-frequency images and labels in anonymised form releasing a smaller anonymised subset of audio describing the request process in more detail (required approvals, typical response time, data format) This would make the work more useful for the community and align well with the journal’s emphasis on reproducibility.

Response 10: We thank the reviewer for highlighting the importance of open data and reproducibility in lung sound research. We fully agree that publicly available datasets are highly valuable for the community. However, due to legal and local ethical restrictions associated with patient privacy and institutional regulations, the raw lung sound recordings cannot be openly shared at this stage. The recordings were collected in a clinical setting under ethics approval that does not permit public release of audio data, even in anonymized form, without obtaining additional ethical clearance. We have therefore indicated that the data can be made available from the corresponding author upon reasonable request, subject to institutional approval. We would like to note that an application for an updated ethics approval is currently in progress to enable broader data sharing in the future. Upon completion of this process, we plan to explore the release of derived representations or a suitably anonymized subset of the dataset, in line with ethical and legal requirements.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Since the images are the same size, the extracted features are also the same. We agree on that. However, your signals are of different lengths. Therefore, you are not evaluating each signal under the same conditions.

I have no further comments.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have adequately addressed the comments raised by the reviewers in the previous round. I recommend this manuscript for publication.