Development and Validation of Multinomial Logistic Regression and Machine Learning Algorithms for Identifying Refractive Status in Chinese School Aged Children.

purposeTo develop and validate models to identify pre-myopic children from their myopic and hyperopic peers using ocular biometry and demographic data.

methodsMultinomial logistic regression (MLR) and machine learning algorithms (random forest and extreme gradient boosting) were applied to data from a school vision screening programme conducted by a private eye hospital in China. The dataset included cycloplegic refraction, ocular biometry and basic demographic information from 36,925 children aged 6-15.99 years. Features were axial length (AL), keratometry, axial length/corneal radius of curvature ratio (AL/CR), gender and age. A penalised MLR with L1 (least absolute shrinkage and selection operator; LASSO) regularisation was used to identify the optimal features, followed by model refinement using likelihood ratio tests and Akaike Information Criterion comparisons to assess interaction terms. Refractive status was classified into three groups: myopic (spherical equivalent refractive error [SER] ≤ -0.50D), hyperopic (SER > +0.75D) and pre-myopic (remaining SER values). Class weighting was applied to the pre-myopic group to prioritise its detection. The dataset was split into 80% training and 20% testing. An additional subgroup of younger age children (6-9.99 years) was used to develop an age-specific model.

resultsOnly the right eye data were used. The optimal model included AL, AL/CR, age, gender and a significant three-way interaction between age, gender and AL/CR. MLR achieved area under the curve (AUC) values of 0.874 in the younger age model and 0.899 in the all-age model. For pre-myopic class detection, the younger age model had 80.6% sensitivity, 60.8% specificity and an AUC of 0.768. The all-age model had 81.5% sensitivity, 71.9% specificity and an AUC of 0.839. Overall, MLR performed similarly or better than the other models.

conclusionOcular biometry combined with demographic data enables detection of pre-myopia in school-aged children and may support its incorporation into population-based vision screening programmes.