Deep Learning in Sleep Medicine: Evaluating Resnet50 and Inceptionv3 for Apnea Detection

Obstructive sleep apnea (OSA) is a medical condition in which the airway becomes obstructed regularly, and resulting in sleep disruption. Sleep apnea is a breathing condition in which a person stops breathing repeatedly while sleeping. OSA is usually diagnosed by an expensive procedure that requires the patient to remain in the hospital overnight. However, this method is expensive, inconvenient, and time consuming. Snoring, poor night sleeps due to choking or gasping, and waking up unrefreshed are all common symptoms of OSA. In this paper, we proposed the two deep pre-trained convolutional neural network models ResNet50 and InceptionV3 for predicting obstructive sleep apnea. We consider the 2-D facial images of different subjects of various conditions such as sleepy, normal, happy and sad images. Even with the small amount of data using transfer learning, InceptionV3 had a good average accuracy of around 91% for OSA prediction and ResNet50 had a good average accuracy of around 82%.