Prediction of Orthodontic Extraction Decisions Using Machine Learning Algorithms: A Retrospective Study.

Background: Orthodontic extraction decision-making remains difficult and highly subjective, particularly in marginal cases where the clinicalcues are ambiguous. Objectives: To design machine learning (ML) models for prediction of extraction vs. non-extraction decision-making and estimate the influenceof key clinical predictors on such decisions. Material and Methods: Retrospective analysis was performed on 120 patients with extraction and 80 patients without extraction from asample of pretreatment records over 2 years. Five ML models including Logistic Regression (LR), Random Forest (RF), Support VectorMachine (SVM), Decision Tree (DT) and XGBoost are employed in this research by applying Python's Scikit-learn. The datasetwas divided in two parts for training and testing at a ratio of 70:30. The sensitivity, specificity, accuracy and AUC-ROCwere used to evaluate and compare the performance of the models. In order to rank the most important features for decision-making, feature importance was calculated. Results: RF model provided the highest accuracy (93.5%) and AUC-ROC (0.95) values, whereas XGBoost was the second-bestmodel, with accuracy (90.2%) and AUC-ROC (0.92). Mandibular crowding (weight = 0.28) and IMPA (L1-MP angle,weight = 0.22) were the most influential predictors. Conclusions: Ensemble ML models, in particular RF, yield a promising objective methodology for clinical decision support in orthodontics topotentially lessen inter-clinician variation and enhance consistency in treatment planning.