This study aimed to develop a deep learning-based model to detect obstructive sleep apnea (OSA) using craniofacial photographs.Participants referred for polysomnography (PSG) were recruited consecutively and randomly divided into the training, validation, and test groups for model development and evaluation. Craniofacial photographs were taken from five different angles (front, right 90° profile, left 90° profile, right 45° profile, and left 45° profile) and inputted to the convolutional neural networks. The neural networks extracted features from photographs and outputted the probabilities of the presence of the disease. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were calculated using PSG diagnosis as the reference standard. These analyses were repeated using two apnea-hypopnea index thresholds (≥ 5 and ≥ 15events/h).A total of 393 participants were enrolled. Using the operating point with maximum sum of sensitivity and specificity, the model of the photographs exhibited an AUC of 0.916 (95% confidence interval [CI], 0.847-0.960) with a sensitivity of 0.95 and a specificity of 0.80 at an AHI threshold of 5 events/h; an AUC of 0.812 (95% CI, 0.729-0.878) with a sensitivity of 0.91 and a specificity of 0.73 at an AHI threshold of 15 events/h.The results suggest that combining craniofacial photographs and deep learning techniques can help detect OSA automatically. The model may have potential utility as a tool to assess OSA probability in clinics or screen for OSA in the community.© 2022. The Author(s), under exclusive licence to Springer Nature Switzerland AG.