MVCNN model for pediatric obstructive sleep apnea prediction via facial photos

Abstract

Introduction Polysomnography (PSG) is the gold standard for pediatric obstructive sleep apnea (OSA) diagnosis but is limited for large-scale screening and telemedicine by high costs and specialized requirements. Given the close correlation between craniofacial features and OSA, this study developed a non-invasive risk stratification model (non-OSA, mild OSA, and moderate-to-severe OSA) using two-dimensional multi-view facial photographs based on a Multi-View Convolutional Neural Network (MVCNN). Materials and methods This multicenter retrospective study included 212 children with frontal and lateral facial photographs linked to polysomnography. Models were evaluated at the patient level after aggregating each participant's available image records. The MVCNN was internally evaluated in 149 patients and externally evaluated in 63 patients. Patient-level discrimination and classification metrics were estimated with bootstrap confidence intervals, and paired bootstrap analysis compared models evaluated on common patients. Results In the internal patient-level evaluation, MVCNN achieved a macro-AUC of 0.898 (95% CI, 0.863-0.930), macro-F1 of 0.839, balanced accuracy of 0.840, and overall accuracy of 0.846. Its macro-AUC exceeded those of ResNet-50 by 0.061 (95% CI, 0.022-0.106; P = 0.004) and VGG-16 by 0.104 (95% CI, 0.057-0.164; P < 0.001). In the external cohort, MVCNN achieved a macro-AUC of 0.876 (95% CI, 0.827-0.919), macro-F1 of 0.833, balanced accuracy of 0.826, and overall accuracy of 0.841. Conclusions Patient-level analyses support the discriminatory performance of multi-view facial photography for pediatric OSA risk stratification in the studied dental cohorts. Prospective validation, direct view-ablation studies, and calibration in broader clinical populations are required before clinical deployment.