Background: The integration of machine learning (ML) into orthodontic practice represents a paradigm shift in how clinicians predict, plan, and evaluate treatment outcomes. This study investigates the comparative performance of multiple ML algorithms — including convolutional neural networks (CNNs), random forests, support vector machines (SVMs), gradient boosting, and deep neural networks (DNNs) — applied to a curated clinical orthodontic dataset.
Methods: A retrospective cohort of 1,000 patients was used. Six ML models were trained and validated on variables including skeletal class, dental measurements, patient age, compliance scores, and radiographic indices. Models were benchmarked on accuracy, sensitivity, precision, recall, and area under the receiver operating characteristic curve (AUC).
Results: The DNN achieved the highest accuracy (94.2%) and AUC (0.968), followed closely by the CNN (93.8%, AUC 0.961). Random forest and XGBoost offered competitive performance with greater interpretability. Dental alignment prediction yielded the highest precision at 96.5%.
Conclusion: Machine learning provides robust and clinically meaningful prediction of orthodontic outcomes. Deep learning models show superior predictive performance, while ensemble methods balance accuracy with explainability, making them practical for clinical deployment.