Ensemble machine learning prediction model for clinical refraction using partial interferometry measurements in...

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Data Science

Background:

• Childhood refractive errors, including myopia, are a growing public health concern.
• Accurate prediction of refractive error is crucial for timely intervention and management.
• Existing prediction models may not fully capture the complex interplay of ocular biometric parameters.

Purpose of the Study:

• To develop and evaluate ensemble machine learning models for predicting clinical refraction in children.
• To compare the accuracy of these novel models against conventional regression techniques.
• To utilize the models for simulating relationships between ocular parameters and refractive error.

Main Methods:

• Retrospective analysis of age, sex, cycloplegic refraction, and partial interferometry data from 1965 patients (aged 5-16 years).
• Development of four ensemble regression models to predict spherical equivalents (SE).
• Comparison of model accuracy using Root Mean Squared Error (RMSE) against a multiple linear regression model.

Main Results:

• Ensemble models achieved superior accuracy with RMSE ranging from 0.800 to 0.829 diopters, compared to 1.213 diopters for the conventional model.
• Simulations indicated female sex is associated with higher myopia prevalence.
• Longer axial lengths showed a dampened increase in myopic refraction per unit length.

Conclusions:

• Ensemble machine learning models effectively predict childhood refractive errors using ocular biometric parameters.
• These models provide a more accurate and insightful tool for understanding refractive development.
• The models can simulate hypothetical scenarios to deepen the understanding of clinical refraction.