The Comprehensive Systematic Review of Association of Bone Age Assessment on Radiographs to Prediction of Final Adult Height Screening

Abstract

Introduction: Accurate prediction of adult height is essential in pediatric endocrinology for diagnosis, treatment decisions, and monitoring of growth disorders. Bone age assessment from hand-wrist radiographs remains the cornerstone of height prediction, yet the accuracy of various methods across different populations and clinical conditions remains incompletely characterized. This systematic review evaluates the association between radiographic bone age assessment and the accuracy of final adult height prediction.

Methods: A systematic review of 80 studies published between 2000-2026 was conducted. Studies were included if they involved children aged 2-18 years, utilized radiographic bone age assessment (Greulich-Pyle, Tanner-Whitehouse, automated systems), reported final adult height measurements, and provided quantitative accuracy measures. Data extraction encompassed bone age methods, prediction algorithms, population characteristics, and statistical measures of prediction accuracy.

Results: In healthy populations, modern automated systems achieve root mean square errors of 2.7-3.3 cm, with deep learning models reaching mean absolute errors of 1.75 cm. The Bayley-Pinneau method remains the most widely used algorithm (>40 studies). Prediction accuracy varies substantially by clinical condition: in growth hormone deficiency, baseline underestimation of 4-8 cm reverses to overestimation of 2-4 cm after treatment; in constitutional delay, overestimation of 1-5 cm occurs; in precocious puberty, underestimation of 3-15 cm at diagnosis improves after gonadotropin-releasing hormone agonist therapy. Automated systems (BoneXpert) reduce inter-rater variability but show condition-specific limitations. Integrative models incorporating parental height, pubertal staging, and novel radiomic markers achieve superior accuracy (R² up to 0.94).

Discussion: The accuracy of bone age-based height prediction is moderated by clinical condition, bone age status relative to chronological age, timing of assessment, and sex. Condition-specific biases are systematic and predictable. Integrative approaches combining multiple skeletal maturity indicators and clinical variables consistently outperform single-method predictions.

Conclusion: Bone age assessment provides moderately accurate adult height prediction in healthy children (RMSE 2.5-3.5 cm) but shows significant condition-specific biases that must be considered clinically. The most reliable predictions require method selection matched to clinical condition, incorporation of parental height and pubertal status, and use of automated systems to minimize rater variability. Future directions include population-specific deep learning models and multi-modal integrative algorithms.