Sequential Machine Learning Approach to Support Premolar Extraction Decisions in Orthodontics.

Abstract

Objective(s): To develop and validate an innovative sequential machine learning framework for premolar extraction decision-making in orthodontics, addressing critical gaps in existing machine learning models and providing transparent clinical reasoning through advanced interpretability techniques.

Materials and methods: Five hundred adult patients from Cali, Colombia (88.4% mestizo, 11.6% Afro-Colombian) were analysed using 36 orthodontist-selected variables and evaluated eight supervised learning algorithms. Implemented a novel two-stage sequential architecture, where mandibular extraction predictions informed maxillary decisions, similar to human clinical reasoning. This approach utilized a comprehensive interpretability analysis, including SHAP, permutation importance and partial dependence plots, to ensure clinical transparency and educational value.

Results: The sequential framework achieved high performance, with 92% mandibular accuracy (Gradient Boosting), 93% maxillary accuracy (XGBoost) and an overall accuracy of 92.5%, surpassing the best previous result by 8.3%. The machine learning model successfully classified all nine symmetric extraction combinations, compared to only five in prior studies. Interpretability analysis revealed clinically meaningful thresholds for L1-APog (4.88-6.37 mm) and lower crowding, with mandibular decisions having a strong influence on maxillary predictions.

Conclusion: This new sequential architecture not only mimics clinical human reasoning but also provides transparent, evidence-based recommendations for mestizo populations. By bridging machine learning with practical clinical application, this tool establishes a new way for AI-assisted orthodontic treatment planning.