An efficient deep model for children sleep apnea detection using snoring signals

Obstructive sleep apnea-hypopnea syndrome(OSAHS) is a risky disorder that has negative effects on individuals' sleep or health. Snoring signals are widely accepted as a reliable and practical alternative in detecting sleep apnea to diagnose OSAHS. Most of previous works paid attention to detecting snoring signals or classifying the places of obstruction. In this paper, diagnosis of OSAHS in children via snoring signals classification is taken into consideration. We build our dataset via gathering and labeling patients and normal children's nocturnal sound recordings. A convolutional neural network (CNN) in parallel with a Transformer encoder network is applied in our method to extract the temporal and frequency information from the acoustic feature sequences. In the experiment our method achieves an accuracy of 95.96% in classifying patients' abnormal snoring events and the snoring from normal children and an accuracy of 84.78% in identifying normal snoring, patients' normal snoring and patients' abnormal snoring. This is acceptable for clinical purposes and indicates that it is competent to serve as a practical tool for diagnosis of OSAHS in children.