Detecting Hypopnea and Obstructive Apnea Events Using Convolutional Neural Networks on Wavelet Spectrograms of Nasal Airflow.

We present a novel approach for detecting hypopnea and obstructive apnea events during sleep, using a single channel nasal airflow from polysomnography recordings, applying a Convolutional Neural Network (CNN) to a 2-D image wavelet spectrogram of the nasal signal. We compare this approach to directly training a 1-D CNN on the raw nasal airflow signal. The evaluation was conducted on a large dataset consisting of 69,264 examples from 1,507 subjects. Our results showed that both approaches achieved good accuracy, with the 2-D CNN outperforming the 1-D CNN. The higher accuracy and the less complex architecture of the 2-D CNN show that converting biological signals into spectrograms and using them in conjunction with CNNs is a promising method for sleep apnea recognition.