Validity of Using Artificial Intelligence to Predict Skeletal Maturation in Orthodontic Treatment: A Systematic Review and Meta-analysis

Abstract

Abstract This systematic review aimed to evaluate the performance and clinical validation of artificial intelligence (AI) and rule-based methods for skeletal maturation assessment using cervical vertebral maturation (CVM) and hand–wrist radiography. The review was conducted in accordance with the PRISMA 2020 guidelines and registered in PROSPERO. The PICOS framework guided selection of experimental, randomized controlled, and observational studies that applied AI to radiographic images of the cervical vertebrae or hand–wrist region. Eligible studies compared AI-derived classifications with expert assessments. Studies based on non-radiographic indicators, non-English publications, and investigations not involving AI methodologies were excluded. A systematic search was performed in PubMed, Scopus, EBSCOhost, and SpringerLink for articles published between 2015 and 2025. The methodological quality and risk of bias were evaluated using the QUADAS-2 tool. For CVM-based studies reporting sufficient quantitative data, a random-effects meta-analysis with restricted maximum likelihood (REML) estimation was performed using logit-transformed accuracy proportions. Heterogeneity was assessed using the Q statistic and the I 2 index. Studies based on hand–wrist radiography were excluded from the quantitative synthesis due to heterogeneity in outcome measures and the limited number of studies. Seven studies met the eligibility criteria, including five that used lateral cephalograms for CVM analysis and two that used hand–wrist radiographs for skeletal age estimation. CVM-based models demonstrated moderate to high diagnostic performance, with reported accuracy or agreement ranging from 60.4 to 82.8%, and kappa values reaching 0.985. These findings corresponded to a pooled accuracy of 80%. When compared with human observer assessments, several CVM-based models achieved comparable levels of agreement in selected studies. However, their performance remained variable across validation settings. Hand–wrist-based models showed consistently high agreement with human observers, with correlation coefficients up to r  = 0.98 and mean absolute errors below 6 months, indicating stable reproducibility relative to expert readings. AI demonstrates promising potential for assisting skeletal maturation assessment, although accuracy varies across model architectures and validation settings. Current evidence supports the use of AI primarily as a diagnostic support tool, with further prospective multicenter studies required before routine clinical implementation. However, the current evidence remains limited due to the small number of studies and methodological heterogeneity.