Abstract: PURPOSE: To analyze and predict the occurrence of postoperative complications after extraction of wisdom teeth by machine learning algorithms, to provide scientific basis for targeted treatment and early prevention of complications of pericoronitis. METHODS: From January 2018 to December 2018, 467 patients with pericoronitis of wisdom teeth who were treated in the Department of Stomatology of the First People's Hospital of Jiayuguan City, Gansu Province were selected. Among them, 373 patients served as training set and 94 patients served as test set. Feature selection and data were completed through detection of digital panoramic tomography combined with doctor’s treatment plan and patients’ follow-up results. By calculating Gini importance of each feature, all features were sorted by importance, then all features were analyzed and the selection was completed. Seven machine learning algorithms were used to establish a complication predicting model, which was trained on the training set until the result converged. The final algorithm was determined by the accuracy and F1 score of 10-fold cross-validation. RESULTS: A 467×16-dimensional dataset containing 15 feature attributes and 1 classification attribute were established. Random forest was finally selected as the core algorithm to complete the complication predicting model after testing. The accuracy of the final model was 89% and F1 weighted average score was 88%. CONCLUSIONS: Machine learning algorithm can effectively analyze the relationship between case features and curative effect after extraction of wisdom teeth as well as predicting complications, which has high clinical practicability.

Key words: Machine learning, Wisdom teeth, Extraction, Complication, Prediction