Machine-learning random forest algorithms predict post-cycloplegic myopic corrections from noncycloplegic clinical...

Area of Science:

• Ophthalmology
• Data Science
• Machine Learning

Background:

• Accurate refractive error assessment is crucial for myopia screening and correction.
• Cycloplegic refraction, while accurate, presents logistical challenges in clinical practice.
• Developing predictive models using noncycloplegic data can improve efficiency and accessibility.

Purpose of the Study:

• To develop and validate machine learning models for predicting post-cycloplegic myopia and refractive outcomes using noncycloplegic clinical data.
• To enhance the accuracy of myopia screening through a classification model.
• To provide an objective starting point for noncycloplegic subjective refractions using a regression model.

Main Methods:

• A cross-sectional study analyzed data from 2483 eyes.
• Random forest classification and regression models were built using pre-refraction measurements (e.g., axial length, corneal curvature) and uncorrected visual acuity.
• Model performance was evaluated using metrics like accuracy, precision, sensitivity, specificity, R-squared, and RMSE.

Main Results:

• The classification model achieved high accuracy (out-of-bag: 92%, cross-validation: 93%, external validation: 94%) and precision (95%).
• The regression model demonstrated strong predictive power with an external validation R-squared of 0.88 and RMSE of 0.63.
• Both models showed robust performance across internal and external validation datasets.

Conclusions:

• Machine learning models can effectively predict post-cycloplegic refractive outcomes from noncycloplegic data.
• The classification model aids in early myopia detection and screening.
• The regression model offers a reliable objective starting point for subjective refractions, improving efficiency in clinical settings, especially where cycloplegia is challenging.