Answers are grounded in this paper and cite their sources.
Objective To develop and evaluate a machine learning framework for predicting orthodontic extraction patterns using features extracted from pretreatment intraoral scans (IOS) and lateral cephalometric radiographs (LCRs). Materials and Methods This retrospective study included 420 subjects (mean age, 17.1 years; 61% female) with treatment plans involving dental extractions. Expert consensus‐derived reference labels were established using a validated Delphi process involving four orthodontic experts. The dataset was randomly split into two mutually exclusive subsets: 30% for hyperparameter optimization using grid search with five‐fold cross‐validation and 70% ( n = 294) for model development and evaluation using leave‐one‐out cross‐validation (LOOCV). Performance was assessed using accuracy, sensitivity, specificity, positive and negative predictive values, likelihood ratios and F1 score. Results The Gradient Boosting classifier trained exclusively on canine‐molar relationship features achieved the highest overall classification accuracy (79.9%), whereas the Random Forest multimodal model incorporating all IOS‐ and LCR‐derived features achieved 78.2% accuracy and did not provide a significant improvement over canine‐molar relationship features alone. SHAP analysis identified canine and molar relationships as the most influential predictors. For the Random Forest multimodal model, class‐specific sensitivity (recall) was 84% for Class I, 77% for Class II, and 48% for Class III. Conclusions Sagittal canine‐molar relationship features contained the greatest discriminative information for extraction pattern classification in this dataset, and incorporating the full multimodal feature set did not improve predictive performance. The substantially lower sensitivity for the underrepresented Class III pattern highlights the need for larger, more balanced datasets.
No takes yet. Share an insight, caveat, or question.
Roederer et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: