Diagnostic accuracy of AI in orthodontic extraction decisions: “Are we ready to let Mr. Data run our Enterprise?” A commentary on a systematic review

Abstract

Objective

To collect evidence on the ability of artificial intelligence programs to accurately make extraction decisions in orthodontic treatment planning.

Data sources

Authors electronically searched the following databases: PubMed/ MEDLINE, EMBASE, LILACS, Web of Science, Scopus, LIVIVO, Computers & Applied Science, ACM Digital Library, and Compendex, Open Grey, Google Scholar, and ProQuest Dissertation and Thesis.

Study selection

Three independent reviewers collected the following data: number of cases of extraction and non-extraction, number of experts in orthodontics and their years of experience, number of variables used in the index model test, type of artificial intelligence and algorithms, accuracy outcomes, the three highest variable ranks weighted in the computational model, and the main conclusion.

Data extraction and synthesis

Risk of bias was assessed using Quadas 2 checklist for AI, and certainty of evidence was evaluated by GRADE.

Results

After 2 phases of screening by 3 independent reviewers, 6 studies met the inclusion criteria for the final review. The AI programs used by the included studies were as follows: ensemble learning/random forest, artificial neural network/multilayer perceptron, machine learning/back propagation and machine learning/feature vectors. All studies showed an unclear risk of bias for patient selection. Two studies had high risk of bias in the index test, while two others presented an unclear risk of bias in the diagnostic test. Meta-analysis of the pooled data resulted in 0.87 accuracy value for all studies.

Conclusions

The authors conclude that AI’s ability to predict extractions is promising but should be interpreted with caution.