Automated Obstructive Sleep Apnea Screening from 2D Cephalograms with Modified ResNet-18: A Pilot Study

Background: Obstructive sleep apnea (OSA) remains significantly underdiagnosed despite its substantial health impacts. The limitations of polysomnography have created an urgent need for accessible screening tools, particularly in dental settings where early anatomical risk factors may be identified through routine imaging. Methods: This multicenter retrospective study analyzed 3,427 lateral cephalometric radiographs. A modified ResNet-18 architecture, adapted for airway analysis, was developed and validated against polysomnography data from 527 cases. Comprehensive interpretability methods including SHAP analysis and class activation mapping (heat maps) were implemented, with all performance metrics including 95% confidence intervals and model calibration assessed using Brier scores. Results: The model achieved 85% sensitivity (95% CI 83-87%) and 83% specificity (95% CI 81-85%) for airway obstruction detection, demonstrating strong agreement with polysomnography (κ=0.73, 95% CI 0.69-0.77). Model calibration was notable (Brier score = 0.12). Processing times averaged 0.038 seconds per image. SHAP analysis identified hyoid position as the most influential predictive feature, showing 89% concordance with clinical landmarks. Class activation maps correctly highlighted retropalatal and retroglossal regions in 89% of positive cases. Performance remained consistent across BMI and age subgroups. Conclusion: Automated analysis of routine cephalograms can provide efficient OSA screening with clinically relevant accuracy. The model's performance and interpretability, including heat map visualizations, support its potential integration into dental practice workflows.