Real-Time Sleep Apnea Detection and Prediction From Single-Channel Airflow

Sleep apnea (SA) is a widespread sleep disorder that causes oxygen desaturation and sleep fragmentation, affecting hundreds of millions worldwide. Real-time feedback on respiratory status is crucial for enabling timely intervention in treatment, improving therapeutic outcomes. This study introduces a novel method for real-time detection or prediction of SA events using airflow (AF) signals. By employing deterministic learning, a dynamical model of the respiratory system is constructed, from which two features are extracted: temporal dynamic trend features and current dynamical features. The former reflects airway vulnerability over time, though it struggles to pinpoint exact SA event timing. The latter indicates current airway patency but may misclassify brief AF restrictions as SA events. Combining these features with appropriate weighting balances their limitations, enabling both detection and prediction. In tests across multiple datasets, the method achieved a maximum detection accuracy of 91.1% and a prediction accuracy of 90%, demonstrating its potential for clinical application and superiority over retrospective diagnostic tools.