Table 1_Prediction of myopia onset and shift in premyopic school-aged children: a machine learning-based algorithm.docx

PurposeThis study aimed to investigate longitudinal changes in ocular parameters and develop a machine learning-based model for predicting myopia onset and shift within 1 year in school-aged premyopic children.
MethodsThis prospective cohort study enrolled 320 premyopic children aged 6–12 years from the Ophthalmology Clinic of The Second Affiliated Hospital of Dalian Medical University. Uncorrected visual acuity (logMAR), cycloplegic spherical equivalent (SE), axial length (AL), average corneal curvature (CC), and subfoveal choroidal thickness (SFCT) were measured at baseline and 6-month intervals for 12 months. Premyopia was defined as - 0.50 D < SE ≤ + 0.75 D. A multivariable analysis evaluated predictive factors including age, gender, parental myopia, baseline SE, AL, CC, axial length/corneal radius (AL/CR), and SFCT. Machine learning algorithms were used to predict 1-year myopia onset and myopia shift, along with Shapley Additive exPlanations (SHAP) interpretation.
ResultsAmong 284 participants (88.8% retention rate), 141 children (49.3%) developed myopia. The cohort exhibited an annual SE progression of −0.695 ± 0.222 D and AL elongation of 0.356 ± 0.122 mm. The AL/CR increased from 2.986 ± 0.061 to 3.029 ± 0.072 (p < 0.001), while SFCT demonstrated a significant reduction of 21.535 ± 9.731 μm (p < 0.001). The optimal model achieved an AUC-ROC of 0.963 (95% CI: 0.930–0.997) for myopia onset prediction, with baseline SE emerging as the most significant predictor, followed by parental myopia, SFCT, and age. Meanwhile, our algorithm also achieved clinically acceptable 1-year predictions of SE.
ConclusionPremyopic children exhibited accelerated myopic progression. Our machine learning-based predictive models showed promising performance for myopia onset and myopia shift, providing clinically valuable risk stratification for targeted prevention strategies.