Beyond the black box: Explainable AI for orthodontic extraction decisions

This is the authors' abstract. We don't add key points for this paper.

Introduction and aims: This study aimed to evaluate whether modern “black-box” AI models align with biological orthodontic principles by benchmarking advanced gradient boosting frameworks (XGBoost, LightGBM, CatBoost) against traditional baselines and interpreting their decision-making processes using Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Methods: A consensus-labeled retrospective dataset of 500 orthodontic patients (334 non-extraction, 166 extraction) sourced from three dental centers in Ho Chi Minh City was analyzed. A 5 × 3 nested cross-validation framework (5-fold outer, 3-fold inner loop) was applied within each training partition. Class imbalance was addressed using Synthetic Minority Over-sampling Technique for Nominal and Continuous (SMOTENC). Hold-out performance was reported with stratified bootstrap 95% confidence intervals (CIs). Exploratory pairwise comparisons used DeLong tests for ROC-AUC and McNemar tests for classification discordance, the latter with Bonferroni and Holm corrections (10 comparisons per protocol). A targeted CatBoost encoding ablation (numeric, native categorical, one-hot) was performed with hyperparameters frozen at the published GridSearchCV values. Results: Gradient boosting models showed high internal discrimination (hold-out ROC-AUC > 0.91), but 95% CIs overlapped across all algorithms and DeLong tests found no significant AUC differences. In the primary 70/30 protocol, CatBoost had the highest point-estimate balanced accuracy (0.875, 95% CI 0.815-0.930) and specificity (0.930, 95% CI 0.880-0.970); ranking reversed in the 80/20 sensitivity analysis, where XGBoost led (balanced accuracy 0.857, 95% CI 0.774-0.925). No McNemar comparison survived Bonferroni or Holm adjustment. Native CatBoost categorical encoding did not improve hold-out balanced accuracy relative to numeric scaling. SHAP and LIME indicated that arch-length discrepancy and soft-tissue profile contributed strongly to predictions, in patterns consistent with—but not proof of—expert orthodontic reasoning. Conclusions: Advanced gradient boosting demonstrated competitive internal performance for formulating pretreatment extraction plans in this retrospective cohort. Apparent model ranking was descriptive rather than confirmatory. SHAP and LIME improved transparency but do not establish biological causality. External validation and prospective evaluation remain necessary. Clinical relevance: This exploratory decision-support framework may help clinicians inspect borderline extraction cases by making model contributions visible. It predicts an expert-formulated pretreatment plan, not an executed clinical outcome, and is not ready for unsupervised deployment.