Accuracy of artificial intelligence in orthodontic extraction treatment planning: a systematic review and meta-analysis

Background

This systematic review and meta-analysis (BMC Oral Health, 2025) evaluated whether artificial intelligence (AI) can accurately predict the need for dental extractions in orthodontic treatment planning. The protocol was registered on PROSPERO (CRD42024582455). AI and machine learning are increasingly used in dentistry for diagnosis and treatment planning; orthodontic extraction versus non-extraction decisions depend on multiple factors and clinician experience, and AI may help standardise and support such decisions.

Methods

Searches were conducted in PubMed, Scopus, Web of Science, and Google Scholar up to June 2025 using terms related to AI, orthodontics, tooth extraction, and treatment planning. Eligible studies were cross-sectional, assessed AI-based models against clinical standards, and reported diagnostic accuracy (sensitivity and specificity). Data were extracted with a standardised form; quality was assessed with the JBI checklist for analytical cross-sectional studies. Pooled sensitivity and specificity were calculated using a random-effects model; heterogeneity was assessed with I²; subgroup analyses by AI model type and meta-regression (e.g. effect of disease prevalence) were performed; funnel plots were used to assess publication bias.

Results

Seven studies (2021–2024) from six countries were included, with 6,261 patients. Sample sizes ranged from 192 to 1,636. AI techniques included CNNs (ResNet-50/101, VGG16/19), Random Forest, MLP, Decision Trees, SVM, and Auto-WEKA. Pooled sensitivity was 70% (95% CI: 61–78) and pooled specificity 90% (95% CI: 87–92), with high heterogeneity (I² ≈ 97% and 94% respectively). In subgroup analysis, CNN-based models (ResNet and VGG) showed the highest and most consistent performance (e.g. ResNet sensitivity 75.8%, specificity 94.1%; VGG sensitivity 82.4%, specificity 93.1%) with I² = 0%. MLP and Random Forest had moderate performance with substantial heterogeneity. Meta-regression showed that higher extraction prevalence in studies was significantly associated with higher sensitivity (p = 0.050). Funnel plots suggested asymmetry and possible publication bias. Leave-one-out sensitivity analysis indicated that pooled sensitivity was robust.

Limitations and heterogeneity

Heterogeneity was high, likely due to differences in study design, populations, AI algorithms, and datasets. Demographic and clinical detail was often limited. Cross-sectional designs do not allow causal conclusions. The small number of studies and possible publication bias limit the strength of recommendations.

Conclusion and implications

AI models, particularly CNN-based ones, show moderate to high diagnostic accuracy for predicting orthodontic extractions and can be used to support evidence-based decisions. They should serve as complementary tools rather than replacing clinical judgment, especially in complex cases. Standardised training datasets, validation protocols, and large-scale multicentre studies are needed before broader clinical implementation.

Reference: Ziaei S et al. BMC Oral Health. 2025;25:1576. PMC12512631. https://pmc.ncbi.nlm.nih.gov/articles/PMC12512631/