Interpretable machine-learning model for cataract associated factors identifying in patients with high myopia
Journal:
medRxiv
Published Date:
Feb 27, 2026
Abstract
Purpose: To systematically evaluate ocular biometric and systemic laboratory factors associated with cataract in highly myopic eyes and to characterize potential nonlinear associations using an interpretable machine learning approach. Methods: This cross-sectional study included 770 eyes of 594 patients with high myopia. Demographic traits, ocular biometric and systemic laboratory factors were collected. Multiple machine learning models were compared, and the random forest (RF) model was selected and fine-tuned. Feature selection and nonlinear relationship analyses were performed, which were further confirmed with partial dependence plots. Results A simplified fine-tuned RF model with 17 features reached stable discriminative performance, with a mean AUC of 0.762 (95%CI: [0.731, 0.794]) among 10 independent testing sets. Age and axial length (AL) turned out to be the most influential features which had non-linear relationships highly myopic cataract, with an inflection point seen around 65.75 (95%CI: [63.72, 67.79]) years for age and 30.55 (95% CI: [29.22, 31.88]) mm for axial length respectively, while the ratio of anterior chamber depth to axial length (ACD/AL) was associated with highly-myopic cataract in a U-shape. Ocular biometric factors were more strongly related to highly myopic cataract than systemic laboratory factors. Conclusions Ocular biometric factors, especially age, AL, and composite indices, have strong and non-linear connections with highly myopic cataract. These results emphasize the significance of ocular structural arrangement in cataract within highly myopic eyes and indicate that interpretable data-driven methods could offer clinically relevant understandings regarding its phenotypic description.
