Artificial intelligence for the diagnosis of malocclusion in children and adolescents: a systematic review of diagnostic accuracy with an exploratory meta-analysis

Background Occlusal anomalies are very common in children and adolescents, and traditional diagnostic techniques are time-consuming and observer-dependent. While artificial intelligence (AI) is a promising tool for automated orthodontic diagnosis, its overall diagnostic performance remains unknown. Objective To evaluate the accuracy of AI-based models in the direct diagnosis and classification of malocclusion in children and adolescents (age ≤18 years). Applications supporting diagnosis (landmark detection, cervical vertebral maturation staging and growth prediction) were summarised as a secondary objective of diagnosis. Methodology PRISMA 2,020 and PRISMA-DTA guidelines were followed. Six databases were searched for studies comparing AI-based models with clinical/radiographic reference standards in participants aged ≤18 years. Study selection, data extraction, and QUADAS-2 assessment were performed independently by two reviewers. Bivariate random-effects meta-analysis was planned for ≥3 studies with 2 × 2 data, with univariate random-effects analysis as an exploratory alternative. Results After re-screening, 10 studies were excluded and 6 lacking full text were classified as awaiting classification. A primary meta-analysis was not performed because only two eligible studies provided 2 × 2 data. An exploratory analysis of four skeletal Class III studies yielded pooled sensitivity of 89.2% (95% CI: 73.9–96.0%) and specificity of 90.5% (95% CI: 82.2–95.2%), with DOR 120.9 (95% CI: 23.7–616.2) and AUC approximately 0.94. Substantial heterogeneity was observed (I 2 = 85–92%). Reported metrics were not considered equivalent. QUADAS-2 indicated mostly low risk of bias, but overall certainty was low. Conclusion AI-based diagnosis of malocclusion in children and adolescents appears to be promising, but currently, the evidence is limited, of low certainty, and insufficient for widespread use in clinical practice; for now, AI should be used in addition to clinicians’ judgement. Larger, standardised diagnostic accuracy studies reporting complete 2 × 2 data are required.