The aim of this study was to develop a machine learning model to assist in treatment decision-making for surgery, camouflage, and growth modulation in Class II malocclusion patients and to evaluate its validity and reliability.

A total of 506 Class II malocclusion patients were included in the study, with patients randomly assigned to a training set (405) and a test set (101). Four machine learning (ML) models — logistic regression (LR), decision tree (DT), random forest (RF), and support vector machine (SVM) — were trained to predict the most suitable treatment approach: camouflage, growth modulation (GM), or surgery. During the evaluation phase, the model was validated using an external dataset obtained from the AAOF Craniofacial Growth Legacy Collection. The accuracy of treatment decisions was evaluated for each model, along with 95% confidence intervals (CIs). Additionally, the chi-square test was used to assess the statistical significance of model performance.

The AUC-PR values indicate that SVM and RF are the best-performing models, both achieving 1.00 for GM, 0.92 for camouflage, and 0.82 for surgery, demonstrating strong classification capabilities across all classes. LR performs well for GM (0.97 ) but struggles with camouflage and surgery (both 0.66), indicating inconsistencies. The DT has the lowest overall performance, with 0.62 for GM and camouflage, and 0.55 for surgery, suggesting weaker classification reliability. Given these results, SVM and RF emerge as the most effective models, offering the best balance of precision and recall across all classes.

Support vector machine and random forest demonstrate strong classification for growth modulation with high precision and recall, while camouflage remains stable until 80% recall before precision declines. Surgery involves greater trade-offs between precision and recall. This study further supports that ANB, Nasolabial angle, SNA, H angle, Age, Mandibular plane angle can be used as strong predictors in assessing patient's treatment needs.