Real-time apnea-hypopnea event detection during sleep by convolutional neural networks.

Sleep apnea-hypopnea event detection has been widely studied using various biosignals and algorithms. However, most minute-by-minute analysis techniques have difficulty detecting accurate event start/end positions. Furthermore, they require hand-engineered feature extraction and selection processes. In this paper, we propose a new approach for real-time apnea-hypopnea event detection using convolutional neural networks and a single-channel nasal pressure signal. From 179 polysomnographic recordings, 50 were used for training, 25 for validation, and 104 for testing. Nasal pressure signals were adaptively normalized, and then segmented by sliding a 10-s window at 1-s intervals. The convolutional neural networks were trained with the data, which consisted of class-balanced segments, and were then tested to evaluate their event detection performance. According to a segment-by-segment analysis, the proposed method exhibited performance results with a Cohen's kappa coefficient of 0.82, a sensitivity of 81.1%, a specificity of 98.5%, and an accuracy of 96.6%. In addition, the Pearson's correlation coefficient between estimated apnea-hypopnea index (AHI) and reference AHI was 0.99, and the average accuracy of sleep apnea and hypopnea syndrome (SAHS) diagnosis was 94.9% for AHI cutoff values of ≥5, 15, and 30 events/h. Our approach could potentially be used as a supportive method to reduce event detection time in sleep laboratories. In addition, it can be applied to screen SAHS severity before polysomnography.