Handling missing data and measurement error for early-onset myopia risk prediction models.

Area of Science:

• Ophthalmology
• Biostatistics
• Data Science

Background:

• Early identification of children at high risk of myopia is essential for timely intervention and prevention of progression.
• Missing data and measurement error (ME) pose significant challenges in myopia risk prediction modeling, potentially introducing bias.

Purpose of the Study:

• To explore and compare various imputation methods (single imputation, MI-MAR, MI-ME, MI-MNAR) for addressing missing data and ME in myopia prediction.
• To evaluate the performance of different machine-learning and statistical models in myopia risk prediction using these imputation techniques.

Main Methods:

• Four imputation methods were explored: single imputation (SI), multiple imputation under missing at random (MI-MAR), multiple imputation with calibration (MI-ME), and multiple imputation under missing not at random (MI-MNAR).
• Four machine-learning models (Decision Tree, Naive Bayes, Random Forest, Xgboost) and three statistical models (logistic regression, stepwise logistic regression, LASSO logistic regression) were compared.
• Model performance was assessed using AUROC and AUPRC on the Shanghai Jinshan Myopia Cohort Study data and through a simulation study.

Main Results:

• MI-ME combined with logistic regression provided the best prediction results when both missing data and ME were present.
• MI-MAR outperformed SI in handling missing data when ME was absent.
• Statistical models demonstrated superior prediction performance compared to machine-learning models.

Conclusions:

• MI-ME is a reliable method for managing missing data and ME in key predictors for early-onset myopia risk prediction.
• The choice of imputation method significantly impacts the accuracy of myopia risk prediction models.