Beyond the algorithm potential: orthodontic tooth-extraction decisions in the age of AI

Abstract

A Commentary on

Ziaei, S., Samani, D., Behjati, M. et al.

Accuracy of artificial intelligence in orthodontic extraction treatment planning: a systematic review and meta-analysis. BMC Oral Health 2025;25:1576. https://doi.org/10.1186/s12903-025-06880-9¹

Question

Can machine learning (ML) accurately predict the need for extraction in the context of orthodontic treatment?

Design

Systematic review for observational studies.

Study selection

Eligibility criteria: Cross-sectional studies compare AI-based models against orthodontists’ opinions in terms of extraction decision in the context of orthodontic treatment planning. Information sources: Four electronic databases (PubMed, Scopus, Web of Science, and Google Scholar) were searched up to June 2025, without language or date restrictions during the search. The search strategy combined both keywords and Medical Subject Headings (MeSH) terms, and it was supplemented by searching the reference lists of related articles. Study selection & Data extraction: The study selection was conducted by two reviewers, followed by the extraction of relevant data. Risk of bias and applicability: The reviewers assessed the quality of the studies using the JBI Critical Appraisal Checklist for Analytical Cross-Sectional Studies.

Data Synthesis

Meta-analysis of sensitivity and specificity using a random-effects model was performed in Python. This was followed by meta-regression using a mixed-effects model and subgroup analysis.

Results

Seven studies were included in this review. The included studies were classified as cross-sectional and varied in sample size, ranging from 192 to 1636 patients, with a wide age range. The AI models used were mainly convolutional neural networks (CNNs), including ResNet variants (ResNet-50, ResNet-101) and VGG networks (VGG16, VGG19), as well as other machine learning algorithms such as Random Forests and Decision Trees. The methodological quality of the three included studies was assessed as high, and four were moderate. The diagnostic performance of AI models for sensitivity estimates ranged from 0.31 (95% CI: 0.22–0.42) to 0.94 (95%CI: 0.90–0.96), with an overall pooled sensitivity of 0.70 (95% CI: 0.61–0.78). The specificity estimates ranged from 0.44 (95% CI: 0.30–0.59) to 0.97 (95% CI: 0.95–0.98), with an overall pooled specificity of 0.90 (95% CI: 0.87–0.92). Subgroup analysis revealed that sensitivity was 0.76 and 0.82 in ResNet (2 studies) and VGG (2 studies) models, respectively, with higher specificity of 0.94 and 0.93 for ResNet and VGG, respectively. However, these differences were not deemed statistically significant. Meta-regression found a statistically significant association between prevalence and sensitivity (β = 0.99, p = 0.05).

Conclusions

With a low certainty level, this synthesis suggested that AI models, CNN-based models, show moderate to high diagnostic accuracy (sensitivity: 70%; specificity: 90%) in predicting dental extractions for orthodontic treatment planning.