Fine-Grained and Lightweight OSA Detection: A CRNN-Based Model for Precise Temporal Localization of Respiratory Events in Sleep Audio

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. The authors wrote, "Obstructive sleep apnea (OSA) is a chronic disorder characterized by recurrent upper airway collapse during sleep, manifesting as snoring, intermittent apnea, and sleep fragmentation. These abnormalities are strongly associated with an increased risk of hypertension, stroke, coronary artery disease, and other severe comorbidities [1]. Currently, the clinical gold standard for diagnosing OSA is polysomnography (PSG). However, PSG involves overnight monitoring of physiological signals in specialized laboratories and requires time-consuming manual scoring by trained technicians." This is the crucial point for the Introduction of the OSA manuscript. However, the authors attributed OSA to a simple sleep disorder with a purely biomechanical basis, which is inaccurate and cited only 1 reference published more than 10 years ago to justify that issue. Please do not cite articles older than 10 years bcause mostly they are outdated. According to the latest findings, OSA has a strong genetic contribution and is associated with neurotransmission: doi: 10.1038/s41598-020-64615-y, doi: 10.1080/07853890.2025.2548386, doi: 10.1038/s41598-023-38842-y. Please justify this sentence "Obstructive sleep apnea (OSA) is a chronic disorder characterized by recurrent upper airway collapse during sleep, manifesting as snoring, intermittent apnea, and sleep fragmentation. These abnormalities are strongly associated with an increased risk of hypertension, stroke, coronary artery disease, and other severe comorbidities [1]." using doi: 10.17219/dmp/185718, doi: 10.17219/dmp/185395, doi: 10.17219/dmp/172243. Please justify this sentence "However, PSG involves overnight monitoring of physiological signals in specialized laboratories and requires time-consuming manual scoring by trained technicians." using doi: 10.1111/jsr.13858, doi: 10.5664/jcsm.6506.

2. Explicitly define whether the model is intended for screening, triage, longitudinal monitoring, or diagnostic support, and align performance claims in the Abstract, Discussion, and Conclusions accordingly.

3. Provide a brief empirical or literature-based rationale comparing this choice with commonly used 5–30 s windows, including its impact on latency, memory, and edge deployment feasibility.

4. Add a focused error analysis of hypopnea misclassifications, including representative failure cases and discussion of acoustic ambiguity versus labeling uncertainty

5. Include inter-technician agreement statistics for PSG annotation in both cohorts to support the reliability of ground-truth labels.

6. Provide a short sensitivity analysis for median filter size, gap-merging threshold, and minimum duration filtering to demonstrate robustness of event reconstruction.

7. Report 95% confidence intervals for AHI correlation, MAE, sensitivity, specificity, and AUC in both internal and external cohorts.

8. Include event-level and patient-level implications of its token-based output limitation, not only frame-level recall, to strengthen the clinical relevance of the comparison.

9. Replace phrases such as “exceptional alignment,” “absolutely critical,” and “outstanding accuracy” with quantitatively grounded, neutral scientific wording.

10. Add a comparison of parameter count and inference speed with at least two recent CRNN or Transformer-based models to objectively justify the “lightweight” designation.

11. Provide key implementation details in a supplementary section, including batch size, number of epochs, early-stopping patience, and exact data split random seed.

12. Explicitly discuss the impact of hypopnea under-detection on AHI underestimation risk and clinical decision thresholds, rather than presenting it only as a technical limitation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you very much for the opportunity to review this work. Below are my comments:

• In the first paragraph of the introduction (lines 44–40), it would be advisable to mention that other simplified methods exist, such as cardiorespiratory polygraphy, which facilitate patient diagnosis. Nevertheless, despite this, long waiting lists remain. I do not believe it is sufficient to mention only polysomnography. I recommend that the authors revise the paragraph and include cardiorespiratory polygraphy.
• After the second paragraph, starting at line 59, I believe the authors should mention that in recent years alternative approaches for the screening and detection of sleep apnea have been explored, such as models based on artificial intelligence and clinical-demographic data, as well as the use of wearables, audio, etc. In this regard, the following articles may be of interest:
  • Enabling early obstructive sleep apnea diagnosis with machine learning: systematic review. Journal of Medical Internet Research, 24(9), e39452.
  • Application of artificial intelligence for the detection of obstructive sleep apnea based on clinical and demographic data: A systematic review. Expert Review of Respiratory Medicine, 1-18.

In this way, the introduction of audio-based models becomes more appropriate and better contextualized.

• In the introduction, the authors should conduct a review of the state of the art, highlighting the main works based on sound and the use of artificial intelligence.
• I recommend that the authors add a paragraph at the end of the introduction explaining how the remainder of the paper is organized.
• Regarding Section 2.2, given the imbalance of the dataset used for model construction, was any stratified approach applied to ensure the correct distribution of cases according to severity?
• The authors should explain the rationale for the proposed architecture. Why was this architecture chosen over others?
• In lines 402–413, it is stated that the model performs very well for the detection of severe cases. This is expected given the prevalence of the disease in the training dataset. However, performance is poor at other severity levels, which is evident and represents a clear limitation.
• With respect to the discussion, in addition to commenting on the apnea–hypopnea index, which is essential in audio-based approaches, the authors should also mention that new emerging metrics currently exist, such as hypoxic burden, for which audio data alone may be insufficient.
• The data come from a hospital setting and were recorded using relatively high-quality microphones; the manuscript itself acknowledges limitations for domestic environments. What issues would arise when using portable devices?
• The sample sizes are very limited, which represents a clear problem.
• Is it possible to access the study code?
• Please explain how artifacts (coughing, talking, movement, impacts, signal saturation) are handled and whether there is automatic detection of invalid segments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript has been revised correctly. I don't have further comments.

 

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you very much for your responses.

I recommend to accept the manuscript for publication.