0882 Local Deformation Analysis of Lateral Cephalogram for Childhood OSA Classification

Abstract

Abstract Introduction Craniofacial profile is one of the anatomical causes of obstructive sleep apnea (OSA). Cephalometry provides information on patients’ skeletal structures and soft tissues. Traditional cephalometric analysis focuses on linear distances, angles, ratios and area of specific variables. Its classification power is often disappointed. In this study, a novel approach to cephalometric analysis using local deformation information was carried out to assess its efficacy in OSA classification. Methods This study was a retrospective analysis based on 60 case-control pairs who were Chinese children recruited for sleep studies in the Prince of Wales Hospital, with accessible lateral cephalometry and polysomnography (PSG) data. Local deformation technique was adopted to derive 1215 deformations from 15 manual landmarking on each cephalogram. In addition, three linear distances (hyoid bone to mandibular plane, hyoid bone to posterior pharyngeal wall, and minimal distance between tongue base and posterior pharyngeal wall) were measured from each cephalogram. A total of 1218 information features were obtained per subject. Classification models were built with an equal ratio between OSA and non-OSA groups (defined by OAHI≥1 and OAHI<1 respectively). Forty pairs were used as training data and twenty pairs were used as testing data. Results Three model settings which used all 1218 cephalometric features, 800 features, and 500 features were tested. The accuracy for the three settings were 67.5% (sensitivity: 70%, specificity: 65%), 87.5% (sensitivity: 90%, specificity: 85%), and 92.5% (sensitivity: 95%, specificity: 90%) respectively. Apart from the three distances, the 500 topmost discriminative features were predominantly landmarks around the nasal cavity. Conclusion A new approach to cephalometric analysis using local deformation information can provide additional details on each cephalogram, hence, achieving better classification. The classification models using 500 features yielded the highest accuracy among the three settings. This setting could benefit most from the comprehensive comparison while avoiding overfitting. Support -