Predicting polysomnographic severity thresholds in children using machine learning.

Area of Science:

• Pediatric Sleep Medicine
• Artificial Intelligence in Healthcare
• Wearable Technology

Background:

• Tonsillectomy and adenoidectomy (T&A) are common procedures for pediatric obstructive sleep disordered breathing (oSDB).
• Polysomnography is effective for risk stratification but is resource-intensive.
• A need exists for cost-effective methods to identify children requiring postoperative monitoring.

Purpose of the Study:

• To develop and validate machine learning models for identifying children with severe oSDB.
• To assess the utility of wearable sensor data (actigraphy and oximetry) for predicting polysomnography-derived severity.
• To establish a resource-conscious screening pathway for children undergoing T&A.

Main Methods:

• Machine learning models were developed using clinical data and sensor data (actigraphy, oximetry).
• Children aged 2-17 years undergoing polysomnography were included in the study.
• Model performance was evaluated based on predicting apnea-hypopnea index (AHI) severity.

Main Results:

• Clinical parameters alone showed poor predictive accuracy for AHI severity (AHI >2: 48-56%; AHI >10: 50-61%).
• Combining oximetry and actigraphy data significantly improved prediction accuracy.
• Accuracies reached 87-89% for AHI >2 and 95-96% for AHI >10 when using combined sensor data.

Conclusions:

• Machine learning utilizing oximetry and actigraphy effectively identifies children needing overnight monitoring based on oSDB severity.
• This approach supports a resource-conscious screening pathway for children undergoing T&A.
• The findings demonstrate the potential for a lower-cost, patient-friendly screening tool for severe obstructive sleep apnea syndrome in children.