목적: To evaluate the current status of facial nerve palsy (FNP) and predict prognosis, various evaluation tools have been developed. Recently, deep learning algorithms have emerged as promising tools for predicting clinical outcomes of various disease. This study aims to predict the early prognosis of FNP by leveraging deep learning model. 방법:We retrospectively reviewed electronic medical records of FNP patients with data on age, gender, hypertension, diabetes, initial FNP grades, and follow-up information. Among them, 279 patients with at least three FNP grade evaluations within 30 days of the first visit and one evaluation between 30 and 300 days were included. Subjects were divided into a training set to develop deep-learning models and a test set for validation, using January 1, 2024, as the cut-off. A long short-term memory neural network predicted FNP grades at the fourth visit, while a support vector machine predicted acceptable recovery. 결과:The classification models, considering H-B grades in serial 2 visits until 30 days after first visit, age, gender, hypertension, diabetes, and delayed treatment, showed an accuracy of 0.903, a recall of 0.991, and an f1-score of 0.944. The accuracy, recall, and f1-score were 0.8, 0.882, and 0.882, respectively in validation data. Additionally, the classification model achieved an accuracy of 0.848, a recall of 0.946, and an f1-score of 0.898 for individuals who did not recover within 30 days of the first visit. Furthermore, regression model was established using age and sex, along with two of hypertension, diabetes, and delayed treatment (treatment starting more than 7 days after onset), were used as features. Model (hypertension, diabetes) 1 had a standard error of 0.4575 and a validation prediction error of 0.4629, while Model 2 (hypertension, delayed treatment) showed a worse standard error (0.5225) and validation prediction error (0.6034) than Model 1. Model 3 (diabetes and delayed treatment) had a standard error of 0.4525 and a validation prediction error of 0.6095. For individuals who did not recover within 30 days of the first visit, the standard errors of Model 1, Model 2, and Model 3 were 0.6524, 0.6384, and 0.6411 in the origin data. 결론:This study demonstrated that changes in FNP grades can be successfully estimated using information from serial FNP visits and clinically important features. Given that FNP results in significant social and emotional costs, appropriate intervention is necessary based on the estimated prognosis. Our deep learning model may assist clinicians in predicting prognosis and providing consultation and treatment for FNP patients.