Explainable deep learning applied to cranial X-rays has identified the forehead and frontal sinus as the single most predictive craniofacial region for obstructive sleep apnea — a finding that challenges decades of clinical teaching. Traditional cephalometric analysis has focused on lower-face structures such as the hyoid bone, mandibular plane, and retroglossal airway. The new results suggest that upper-face features may matter far more than previously recognized.
Obstructive sleep apnea (OSA) affects an estimated one billion adults worldwide. Craniofacial anatomy is a well-established contributor to OSA risk, especially in Asian populations where bony restriction of the airway is more prevalent than obesity-driven disease. Until now, cephalometric research has concentrated on the lower facial skeleton.
A team led by Luo et al. from Peking University, Beijing University of Technology, and the Cross-strait Tsinghua Research Institute used deep learning to uncover overlooked craniofacial predictors. Their results, published June 18 in Nature and Science of Sleep, challenge the lower-face orthodoxy.
The investigators built a multimodal deep learning model processing three data types simultaneously: frontal facial photographs, lateral facial photographs, and lateral cephalograms (standardized skull X-rays), together with demographic and clinical variables.
The architecture paired a pre-trained GoogLeNet convolutional neural network for image feature extraction with a fully connected network for tabular clinical data. The two streams merged into a final classification layer predicting OSA status (apnea-hypopnea index >= 5).
The cohort included 130 Chinese adults: 65 OSA patients and 65 matched controls. BMI was capped at 32 kg/m^2 to isolate craniofacial contributions from obesity-related airway collapse.
The model achieved an AUC of 0.87, accuracy of 87%, sensitivity of 93.3%, and specificity of 75%. The high sensitivity is especially relevant for screening, where missing a diagnosis carries greater risk.
The researchers deployed three complementary XAI techniques: Grad-CAM heatmaps to visualize which image regions drove model decisions, feature importance scoring to rank anatomical measurements, and average face analysis comparing composite facial morphologies between groups.
The Forehead Surprise
When XAI tools analyzed the lateral cephalograms, one region dominated: the upper third of the face, centered on the forehead and frontal sinus. The frontal sinus area received a feature importance score of 1.43, far exceeding traditional measures such as hyoid position, mandibular length, or posterior airway space.
Quantitative measurements confirmed the pattern. Frontal sinus area was significantly larger in OSA patients (249.49 vs 189.02 mm^2, p < 0.001). Forehead protrusion was also greater in the OSA group (p < 0.001).
The finding is striking because frontal sinus size and forehead morphology are not part of any standard OSA cephalometric analysis. Clinical protocols emphasize lower facial landmarks: hyoid-to-mandibular-plane distance, soft palate length, and retropalatal airway width. The upper face has been largely ignored.
Other Novel Craniofacial Features
The XAI analysis identified additional non-traditional features associated with OSA status:
- Inter-eyebrow distance ratio: The proportional distance between the eyebrows relative to facial width, suggesting mid-face width contributes predictive information.
- Lower lip-to-chin distance: A vertical proportion of the lower face distinct from standard mandibular length.
- Mid-facial length ratio: The proportional height of the mid-face relative to total facial height.
These features are absent from conventional cephalometric analyses, which measure linear distances and angles between fixed bony landmarks. The deep learning model learned directly from raw pixel data and identified predictive patterns never systematically annotated by human experts.
Why This Matters for OSA Screening
The findings could meaningfully improve OSA screening, particularly in Asian populations. Bony structural restriction of the airway is a more prominent OSA contributor in Asian cohorts than obesity, making craniofacial screening tools especially relevant. Yet current instruments such as STOP-Bang emphasize BMI, age, neck circumference, and symptoms, with limited anatomical assessment.
A screening approach incorporating forehead and frontal sinus morphology — detectable from a lateral cephalogram or potentially from facial photographs — could improve pre-diagnostic risk stratification. The 93.3% sensitivity is encouraging for a screening tool, where a missed diagnosis carries higher cost than an unnecessary sleep study referral.
Limitations
The study has important limitations. The sample of 130 participants is modest for deep learning, and the model was trained on a single ethnic group (Chinese Han adults). Whether upper-face patterns generalize to other populations is unknown. The BMI cap of 32 kg/m^2, while methodologically sound, limits applicability to the large proportion of OSA patients with obesity. The model also requires lateral cephalograms, which involve radiation and are not routine in OSA workup, though the authors suggest facial photographs alone may carry some predictive value.
Further validation in larger, multi-ethnic cohorts with broader BMI ranges is needed before translation to clinical screening tools. Prospective studies testing whether forehead and frontal sinus measurements improve diagnostic yield over standard questionnaires are the logical next step.
Source: Luo L, Yang R, Pei Z, Yu M, Gong X, Lei Y, Wang Q, Gao X. Explainable deep learning identifies craniofacial features for obstructive sleep apnea screening from multimodal data. Nature and Science of Sleep. 2026;18:609113. DOI: 10.2147/NSS.S609113.