Refining Sleep-Disordered Breathing Annotations Across Multiple Public Sleep Study Datasets.

Area of Science:

• Sleep Medicine
• Biomedical Data Science
• Artificial Intelligence in Healthcare

Background:

• Existing polysomnography annotations in major sleep studies (SHHS, MrOS, MESA) used simplified criteria for apneas/hypopneas, deviating from American Academy of Sleep Medicine (AASM) guidelines.
• This inconsistency leads to inaccurate apnea-hypopnea indices (AHIs) and potentially flawed research conclusions.

Purpose of the Study:

• To develop and validate a standardized annotation pipeline for polysomnography data integrating sleep staging, oxygen desaturation, and arousal events per AASM criteria.
• To re-evaluate AHI and obstructive sleep apnea (OSA) severity in large cohorts using refined annotations.
• To assess the impact of improved annotations on deep learning model performance for OSA classification.

Main Methods:

• Developed a standardized annotation pipeline incorporating AASM guidelines for sleep staging, oxygen desaturation, and arousal events.
• Retrospectively analyzed polysomnography data from SHHS1, SHHS2, MrOS1, MrOS2, MESA, and KISS cohorts.
• Compared original AHI values with recalculated AHIs using the refined pipeline and trained deep learning models on both original and refined annotations.

Main Results:

• Original annotations in SHHS, MrOS, and MESA significantly overestimated AHIs (MAE: 10.3–23.6 events/h).
• Refined annotations drastically reduced MAE to 0.56–1.29 events/h, demonstrating improved accuracy.
• Deep learning model performance for OSA severity classification improved, with the F1 score increasing from 0.5 to 0.69 after annotation refinement.

Conclusions:

• The standardized annotation pipeline enhances cross-cohort consistency and accuracy in polysomnography analysis.
• This approach enables more reliable utilization of existing sleep study datasets for clinical research and AI-driven applications.
• Adherence to AASM guidelines through standardized annotation is crucial for accurate sleep apnea assessment and research.