Prediction of obstructive sleep apnea using deep learning in 3D craniofacial reconstruction

BACKGROUND: Obstructive sleep apnea (OSA) is a common sleep disorder. However, current diagnostic methods are labor-intensive and require professionally trained personnel. We aimed to develop a deep learning model using upper airway computed tomography (CT) to predict OSA and to warn the medical technician if a patient has OSA while the patient is undergoing any head and neck CT scan, even for other diseases. METHODS: A total of 219 patients with OSA [apnea-hypopnea index (AHI) ≥10/h] and 81 controls (AHI <10/h) were enrolled. We reconstructed each patient’s CT into 3 types (skeletal structures, external skin structures, and airway structures) and captured reconstructed models in 6 directions (front, back, top, bottom, left profile, and right profile). The 6 images from each patient were imported into the ResNet-18 network to extract features and output the probability of OSA using two fusion methods: Add and Concat. Five-fold cross-validation was used to reduce bias. Finally, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were calculated. RESULTS: All 18 views with Add as the feature fusion performed better than did the other reconstruction and fusion methods. This gave the best performance for this prediction method with an AUC of 0.882. CONCLUSIONS: We present a model for predicting OSA using upper airway CT and deep learning. The model has satisfactory performance and enables CT to accurately identify patients with moderate to severe OSA.