목적: Obstructive sleep apnea (OSA) is a common condition after middle age, especially in men. It is known that an obstruction of the upper airway during sleep, causing hypoxia and increasing the risk of developing cardiovascular disease. Delays in diagnosis and treatment in patients with OSA increase long-term all-cause mortality. Previous studies on predicting obstructive sleep apnea have shown results separately using 2D facial images and clinical information data. In this study, we aim to improve performance through a multimodal method that utilizes both 2D facial data and clinical information. 방법:We reviewed data of patients who underwent polysomnography, facial photography, and cephalometry between January 2016 and July 2023. Clinical parameters such as sex, age, body mass index, neck circumference, abdomen circumference, and sleep questionnaires were assessed. A model that integrates EfficientNet-B7 and DNN is used to learn multimodal facial images and clinical data. We evaluated the diagnostic accuracy of the multimodal model for severity of OSA. 결과:The area under the curve (AUC) values from internal validation were 0.873 between mild and severe OSA, 0.775 between normal to mild and moderate to severe OSA, and 0.787 between mild and moderate to severe OSA. In holdout validation, the AUC were 0.933 between mild and severe OSA, 0.823 between normal to mild and moderate to severe OSA, and 0.835 between mild and moderate to severe OSA. 결론:This study shows that the multimodal learning method is superior to the existing single-modal artificial intelligence-based OSA diagnosis method and can be a pioneering indicator for multimodal research on OSA.