Abstract
Obstructive sleep apnea (OSA) and related hypoxia are well-established cardiovascular and neurocognitive risk factors. Current multi-sensor diagnostic approaches are intrusive and prone to misdiagnosis when simplified. This study introduces an enhanced single-sensor-based OSA screening method, leveraging novel signal processing and machine learning to ensure accurate detection across diverse populations. Wrist actigraphy is a widely-used and energy-efficient tool for respiratory rate estimation. The main challenge in OSA pattern recognition is handling various disturbances in real-world applications. We developed a novel approach combining apex-centric tokenization with a Multi-Head Causal Attention (MHCA) mechanism. Apex-centric tokenization enhances sensitivity to OSA events, while MHCA refines predictions and increases specificity in detecting oxygen desaturation. Our study involved 58 participants, with overnight bilateral wrist actigraphy and concurrent polysomnography used as a reference for thorough analysis. By focusing on the physiological causal relationship of the events, the algorithm excelled in detecting moderate to severe oxygen desaturation, achieving a sensitivity of 85.7% and a specificity of 98.1%, even in the presence of disturbances such as restless leg movements and snoring. The estimated oxygen desaturation index correlated strongly with clinical standards (r = 0.89), and the correlation with the apnea-hypopnea index was 0.87. Both apex-centric tokenization and MHCA were crucial for this performance. Our approach shows potential for analyzing apnea patterns and related oxygen desaturation in a broader population using only wrist actigraphy, reducing measurement burdens and improving understanding of complex sleep disorders.