OSA-NET : An Efficient Convolutional Neural Network for OSA Diagnosis Screening Tool

Abstract

Obstructive sleep apnea (OSA) has a high prevalence worldwide, particularly among adults and the population with comorbidities. Adult men have a higher incidence rate and approximately 15-30% prevalence, and OSA is associated with other diseases such as obesity, hypertension, and diabetes. Polysomnography is essential for diagnosing OSA, but discomfort and overnight testing are practical difficulties. Therefore, several questionnaires, such as the STOP-BANG and Berlin questionnaire, have been developed for initial screening of OSA, although their use is limited due to their low accuracy. One of the pathophysiology of OSA is related to craniofacial anatomy, and several previous studies have investigated facial anatomy using 2D or 3D photographs with a small number of patients, but practical application for screening and diagnosing OSA has not been attempted yet. Therefore, our research aims to screen OSA and stratify its severity with feasible 2D photographs and a big dataset using a novel deep learning algorithm. We developed a new CNN-based algorithm called OSA-Net, which diagnoses OSA patients using information on facial anatomical abnormalities by ready-to-use photographs. Facial pictures of 900 patients who underwent polysomnography were used as a dataset for model training. The results of our algorithm showed a high diagnostic accuracy over 85% for classifying OSA severity into normal, mild, moderate, and severe degrees. Furthermore, we used Grad-CAM to verify the accuracy of our model, and the results showed that our model accurately recognizes facial contours. This study demonstrates the potential of using artificial intelligence with image big data as a new OSA screening tool.