歯科矯正治療後の肺胞骨欠損リスク予測のための機械学習ベースの予測モデル

Alveolar bone defects, such as fenestration and dehiscence, induced by orthodontic treatment represent significant complications that can impact treatment outcomes and long-term health. This study aims to develop a machine learning-based predictive model for classifying the risk grade of alveolar bone defects following orthodontic treatment.

This study utilised a retrospective cohort design, involving 354 patients who had completed orthodontic treatment. Participants were divided into a training set and a validation set in an 8:2 ratio. The outcome indicator was a three-category risk grade (low, medium, high) based on the length of bone defects. From an initial pool of 100 predictor variables, 23 key features were selected using LASSO regression. Five machine learning models were then constructed: Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Multilayer Perceptron (MLP), Multivariate Logistic Regression, and Decision Tree. Model performance was comprehensively evaluated using metrics including accuracy, precision, recall, specificity, F1-score, and the area under the receiver operating characteristic curve (AUC). The best-performing model, XGBoost, was compared with conventional methods and further interpreted using Shapley Additive exPlanations (SHAP) to assess feature importance.

LASSO regression identified 23 critical predictors. Among all models, XGBoost, which ultimately incorporated 20 variables, demonstrated the most superior predictive performance, achieving an accuracy of 0.9429 and an AUC of 0.9955. SHAP analysis indicated that the primary risk factors influencing the model's decisions included the pre-treatment bone defect score, the width of the basal bone arch in the middle maxilla, and the length of the posterior mandible, among others. The XGBoost model demonstrated significantly superior performance compared to conventional evaluation methods.

This study successfully developed an XGBoost machine learning model incorporating 20 variables to predict the risk grade of alveolar bone defects following orthodontic treatment. The model demonstrated excellent predictive performance, significantly outperforming conventional assessment methods and other commonly used machine learning models. Core variables such as the pre score, lower posterior length, and upper mid base width were found to substantially enhance its predictive capability. This model shows potential as a clinical decision-support tool to assist orthodontists in pre-treatment risk assessment and personalised treatment planning.