Fusion of Whole Night Features and Desaturation Segments Combined with Feature Extraction for Event-Level Screening of Sleep-Disordered Breathing

PURPOSE: Misdiagnosis and missed diagnosis of sleep-disordered breathing (SDB) is common because polysomnography (PSG) is time-consuming, expensive, and uncomfortable. The use of recording methods based on the oxygen saturation (SpO2) signals detected by wearable devices is impractical and inaccurate for extracting signal features and detecting apnoeic events. We propose a method to automatically detect the apnoea-based SpO(2) signal segments and compute the apnoea–hypopnea index (AHI) for SDB screening and grading. PATIENTS AND METHODS: First, apnoea-related desaturation segments in raw SpO(2) signals were detected; global features were extracted from whole night signals. Then, the SpO(2) signal segments and global features were fed into a bi-directional long short-term memory convolutional neural network model to identify apnoea-related and non-apnoea-related events. The apnoea-related segments were used to assess the AHI. RESULTS: The model was trained on 500 individuals and tested on 8131 individuals from two public hospitals and one private centre. In the testing data, the classification accuracy for apnoea-related segments was 84.3%. Individuals with SDB (AHI [Image: see text] 15) were identified with a mean accuracy of 88.95%. CONCLUSION: Using automatic SDB detection based on SpO(2) signals can accurately screen for SDB.