Diagnostic accuracy of artificial intelligence models for orthodontic extraction decisions: A systematic review and meta-analysis

Background Artificial intelligence (AI) has been increasingly applied to orthodontic treatment planning, particularly to extraction versus non-extraction decisions. However, the available evidence is methodologically heterogeneous, and many studies do not provide sufficiently comparable diagnostic data for formal quantitative synthesis. Objective To systematically evaluate the evidence on AI models for orthodontic extraction decision-making and to provide a restricted diagnostic-accuracy synthesis of studies with directly comparable binary extraction/non-extraction outcomes. Methods This systematic review followed PRISMA principles, was registered in PROSPERO (CRD42024537540), and assessed risk of bias using QUADAS-2. Six databases were searched from inception to February 2025. Original studies evaluating AI or machine-learning models for orthodontic extraction decision-making were eligible. The evidence was structured into primary diagnostic-accuracy studies, secondary accuracy studies, and narrative-only studies. A restricted bivariate random-effects Reitsma model was used to pool sensitivity and specificity for studies with useable 2 × 2 data. A separate exploratory random-effects synthesis of reported accuracy was conducted for secondary studies lacking a fully comparable diagnostic table. Results Eighteen orthodontic studies, representing 9340 participants, met the inclusion criteria. Four studies were suitable for the restricted primary diagnostic-accuracy synthesis. In this subset, pooled sensitivity was 0.828 (95% CI 0.608–0.937) and pooled specificity was 0.931 (95% CI 0.857–0.968), with substantial heterogeneity for sensitivity and moderate heterogeneity for specificity. The summary receiver operating characteristic analysis yielded an AUC of 0.953. Twelve additional studies contributed to the exploratory secondary pooling of reported accuracy, which yielded a pooled estimate of 0.857 (95% CI 0.815–0.890). Sensitivity analyses restricted to studies with lower methodological uncertainty or hold-out validation showed higher pooled accuracy and lower heterogeneity. Conclusions AI models show promise for supporting orthodontic extraction decision-making, particularly in binary extraction/non-extraction tasks. However, only a small subset of the literature currently supports formal diagnostic-accuracy synthesis, and the broader evidence remains limited by heterogeneity in outcomes, reporting, and validation strategies. At present, AI should be interpreted primarily as a clinical decision-support tool rather than as a substitute for expert orthodontic judgment. Registration PROSPERO CRD42024537540.