Artificial Intelligence-Powered Craniofacial Photogrammetry Analysis of Pediatric Obstructive Sleep Apnea.

OBJECTIVES: To develop an artificial intelligence (AI) system for craniofacial morphology analysis in pediatric obstructive sleep apnea (OSA) using photogrammetry. STUDY DESIGN: Prospective, cross-sectional study. SETTING: Tertiary medical hospital. METHODS: Children aged 3 to 18 years with OSA-related symptoms were enrolled and underwent overnight polysomnography (PSG) and standardized craniofacial photogrammetry. Moderate-to-severe OSA in children was defined as an apnea-hypopnea index (AHI) ≥ 5 events/h in PSG. An AI model using the Dlib tool identified facial landmarks, and the Hough transform calculated variables from these coordinates. Measurements by humans, the AI model, and a manually adjusted AI model were compared. Random forest identified the top 10 variable importance for the OSA prediction model. RESULTS: Forty-three children with moderate-to-severe OSA and 43 age-, gender-, and obesity-matched controls were included. Shared predictors across models were mandibular plane angle, maxillary-mandibular relationship, lower facial length, and lower facial proportion. Additional predictors for the AI model included lower- and mid-face projection, while the adjusted AI model added mid-face projection, retrusive mandible, and cervicomental angle. The area under the curve (AUC) values for moderate-to-severe OSA prediction were similar in human, AI model, and adjusted AI model (0.74 vs 0.71 vs 0.70, P for ΔAUC > 0.05). The AI model significantly reduced measurement time (human vs AI vs adjusted AI = 511.4 vs 0.85 vs 15.8 seconds, P < .001). CONCLUSION: The AI-powered photogrammetry analysis system is a rapid and reliable tool with comparable performance to human measurements in evaluating pediatric OSA.