Mendelian Randomization and Machine Learning for Anatomical Subtyping of Adolescent Sleep Apnea

Introduction and aims Obstructive sleep apnea (OSA) is associated with dentofacial skeletal and soft-tissue anomalies, but whether these anatomical pathways contribute independently, and whether they define clinically distinct subtypes, remain unresolved. We integrated genetic causal inference with data-driven anatomical phenotyping of adolescent OSA. Methods Bidirectional univariable and multivariable Mendelian randomization (UVMR, MVMR) were applied to European-ancestry genome-wide association study datasets, together with unsupervised and supervised machine learning of 35 cephalometric measures from a 259 Chinese adolescent OSA cohort. Results OSA and dentofacial anomalies were bidirectionally related; OSA additionally showed unidirectional effects on maxillary retrognathia (OR = 2.06, q = 0.006) and temporomandibular disorders (OR = 1.196, q = 0.016). For the soft-tissue pathway, evidence was strong in the OSA-to-adenotonsillar direction (OR = 1.055, q = 0.013) and suggestive in the reverse (OR = 1.037, q = 0.098). After mutual adjustment, MVMR indicated that dentofacial anomalies (OR = 1.037, p = .046) and chronic adenotonsillar disease (OR = 1.042, p = .007) remained suggestive evidence for OSA risk, but a conditional F < 10 for the soft-tissue exposure precluded estimation of its independent contribution. Unsupervised clustering identified 3 stable phenotypes (bootstrap Jaccard > 0.91), recovered within each cervical vertebral maturation stratum: skeletal (mandibular retrusion, vertical growth); soft-tissue (adenotonsillar hypertrophy, normal skeletal profile); and mixed, distinguished by imaging-confirmed degenerative temporomandibular joint (TMJ) changes (46.8%) and nasal turbinate hypertrophy (69.4%). Apnea-Hypopnea Index (AHI) did not differ across phenotypes ( p = .936). Internal cross-validated accuracy was 96.6%, with near-perfect discrimination for the skeletal and soft-tissue phenotypes (AUC > 0.996) with lower probabilistic confidence for the mixed phenotype. Conclusion Genetic and anatomical evidence converge qualitatively on skeletal and soft-tissue axes of risk, corresponding to 3 stable morphological subtypes that AHI does not distinguish. The mixed subtype, enriched for degenerative TMJ change, indicates an underrecognized structural correlate of OSA. Clinical Relevance Machine-learning analysis of routine cephalometric anatomy stratifies adolescent OSA into anatomy-based subtypes beyond AHI and flags patients warranting TMJ surveillance, although clinical translation awaits external and prospective testing.