Data-Efficient Deep Learning-Based Screening of Obstructive Sleep Apnea Using Lateral Cephalograms

Obstructive sleep apnea (OSA) is associated with substantial cardiometabolic and neurocognitive morbidity and increased mortality. However, a large proportion of moderate-severe cases remain undiagnosed in Asian populations. This gap is driven not only by limited access to polysomnography, but also by screening approaches that perform poorly in Asians and overlook craniofacial anatomy, a dominant contributor to OSA risk in non-obese individuals. As a result, many high-risk patients are not identified early within routine care pathways. This study aimed to develop a screening-oriented deep learning model to estimate patients' pretest probability of clinically significant OSA (moderate-severe disease; AHI ≥15) using lateral cephalograms routinely acquired in orthodontic practice. A retrospective dataset of 350 adult patients with PSG-confirmed diagnoses from Chang Gung Memorial Hospital was used, with images split into training, validation, and test sets (7:1:2). A DenseNet201 architecture pretrained on ImageNet was adapted for binary classification, with attention mechanisms and data augmentation incorporated to enhance robustness under real-world imaging variability. The attention-enhanced model achieved a test AUC of 0.755. Five-fold cross-validation demonstrated stable and reproducible performance (mean test AUC 0.7352 ± 0.0152; sensitivity 0.7667 ± 0.1733; specificity 0.6818 ± 0.1437), despite the modest dataset size. Gradient-weighted class activation mapping revealed anatomically plausible activation patterns, supporting interpretability and clinical credibility. These findings demonstrate that meaningful and generalizable screening performance can be achieved from routinely collected dental imaging using deliberately efficient training strategies, establishing a practical foundation for scalable AI-assisted OSA risk stratification and supporting further validation toward clinical deployment.