RESUMO
BACKGROUND/OBJECTIVES: Axial length, a key measurement in myopia management, is not accessible in many settings. We aimed to develop and assess machine learning models to estimate the axial length of young myopic eyes. SUBJECTS/METHODS: Linear regression, symbolic regression, gradient boosting and multilayer perceptron models were developed using age, sex, cycloplegic spherical equivalent refraction (SER) and corneal curvature. Training data were from 8135 (28% myopic) children and adolescents from Ireland, Northern Ireland and China. Model performance was tested on an additional 300 myopic individuals using traditional metrics alongside the estimated axial length vs age relationship. Linear regression and receiver operator characteristics (ROC) curves were used for statistical analysis. The contribution of the effective crystalline lens power to error in axial length estimation was calculated to define the latter's physiological limits. RESULTS: Axial length estimation models were applicable across all testing regions (p ≥ 0.96 for training by testing region interaction). The linear regression model performed best based on agreement metrics (mean absolute error [MAE] = 0.31 mm, coefficient of repeatability = 0.79 mm) and a smooth, monotonic estimated axial length vs age relationship. This model was better at identifying high-risk eyes (axial length >98th centile) than SER alone (area under the curve 0.89 vs 0.79, respectively). Without knowing lens power, the calculated limits of axial length estimation were 0.30 mm for MAE and 0.75 mm for coefficient of repeatability. CONCLUSIONS: In myopic eyes, we demonstrated superior axial length estimation with a linear regression model utilising age, sex and refractive metrics and showed its clinical utility as a risk stratification tool.
