The collapsibility of the upper airway has a known anatomical basis that is mediated by an interaction of obesity and craniofacial abnormalities. The pattern of these abnormalities, if detected in… Click to show full abstract
The collapsibility of the upper airway has a known anatomical basis that is mediated by an interaction of obesity and craniofacial abnormalities. The pattern of these abnormalities, if detected in a subject's facial image, can help predict the presence of obstructive sleep apnea (OSA). OBJECTIVE We utilized facial photographs (front and profile) from 376 patients who had undergone an overnight polysomnogram to identify those with and without OSA. Approach (i): Processing the images had three steps: landmark identification, feature generation, and automatic classification. Firstly, we used unaligned landmarks and investigated the impact of manually and automatically determined landmarks on the performance of the classification algorithms. Main results (i): Using a feature set calculated from manually determined landmarks ('constrained features') we obtained 67.0% accuracy in identifying OSA (AHI  >  10) and using automatically determined landmarks we obtained 65.5% accuracy for OSA. Approach (ii): Secondly, we aligned the facial landmark coordinates by using an image registration technique utilising linear transformations. Main results (ii): Directly using the aligned landmark coordinates ('unconstrained features') by the classification stage as the new feature set resulted in an accuracy of 69.7% in OSA detection using manually determined landmarks and 69.2% using automatically determined landmarks. SIGNIFICANCE The performance of our fully automatic system using the unconstrained feature set is comparable to other published systems requiring a manual landmarking process. In conclusion, we demonstrate the feasibility of using automatic landmarking as well as unconstrained features in the successful prediction of OSA using facial landmarks.
               
Click one of the above tabs to view related content.