Accuracy of artificial intelligence for tooth extraction decision-making in orthodontics: a systematic review and meta-analysis.

Abstract

This study aimed to analyze the accuracy of artificial intelligence (AI) for orthodontic tooth extraction decision-making. PubMed/MEDLINE, EMBASE, LILACS, Web of Science, Scopus, LIVIVO, Computers & Applied Science, ACM Digital Library, Compendex, and gray literature (OpenGrey, ProQuest, and Google Scholar) were electronically searched. Three independent reviewers selected the studies and extracted and analyzed the data. Risk of bias, methodological quality, and certainty of evidence were assessed by QUADAS-2, checklist for AI research, and GRADE, respectively. The search identified 1810 studies. After 2 phases of selection, six studies were included, showing an unclear risk of bias of patient selection. Two studies showed a high risk of bias in the index test, while two others presented an unclear risk of bias in the diagnostic test. Data were pooled in a random model and yielded an accuracy value of 0.87 (95% CI = 0.75-0.96) for all studies, 0.89 (95% CI = 0.70-1.00) for multilayer perceptron, and 0.88 (95% CI = 0.73-0.98) for back propagation models. Sensitivity, specificity, and area under the curve of the multilayer perceptron model yielded 0.84 (95% CI = 0.58-1.00), 0.89 (95% CI = 0.74-0.98), and 0.92 (95% CI = 0.72-1.00) scores, respectively. Sagittal discrepancy, upper crowding, and protrusion showed the highest ranks weighted in the models. Orthodontic tooth extraction decision-making using AI presented promising accuracy but should be considered with caution due to the very low certainty of evidence. AI models for tooth extraction decision in orthodontics cannot yet be considered a substitute for a final human decision.