Introduction. Progressive myopia (nearsightedness) in children represents one of the most acute medical and social problems in modern ophthalmology.
Purpose. To develop a machine learning model for predicting the outcome of scleroplasty in children 12 months after surgery.
Material and methods. A dataset was formed comprising 128 eyes of 128 patients who underwent scleroplasty at the S.N. Fyodorov Eye Microsurgery Federal State Institution (Moscow). The development of machine learning models for binary classification was conducted using the Python 3 programming language and the PyCaret library.
A total of 19 models were developed: Extra Trees Classifier, Linear Discriminant Analysis, Gradient Boosting Classifier, Naive Bayes, Logistic Regression, CatBoost Classifier, K Neighbors Classifier, MLP Classifier, Decision Tree Classifier, Quadratic Discriminant Analysis, Random Forest Classifier, Ada Boost Classifier, Light Gradient Boosting Machine, Gaussian Process Classifier, Extreme Gradient Boosting, SVM – Radial Kernel, Dummy Classifier, Ridge Classifier, and SVM – Linear Kernel. The target variable was the prognosis of the scleroplasty result in the form of a binary feature: favorable (64 eyes) and unfavorable (64 eyes) outcome. An outcome was considered favorable if the annual progression gradient 12 months after scleroplasty was greater than -1.00 D, and unfavorable if it was -1.00 D or less. The independent variables used to develop the machine learning models were: age, gender, UCVA (uncorrected visual acuity) before, Sph (sphere) before, Cyl (cylinder) before, BCVA (best corrected visual acuity) before, SE (spherical equivalent) before, K min before, K max before, R (radius) before, and AL (axial length) before. For each machine learning model, hyperparameter tuning was performed using crossvalidation on 10 folds using the Optuna library; optimization was carried out based on the AUC metric. The following quality metrics were calculated: AUC, accuracy, precision, recall, and F1-score. For the development and testing of the machine learning models, the total dataset was divided into training and test sets in a 69:31 ratio, with stratification performed by the target variable. Feature importance assessment was conducted using the feature_importances_ method.
Results. 19 machine learning models were developed for the binary classification of scleroplasty outcomes in children 12 months after surgery (favorable/unfavorable outcome). Among them, the Extra Trees Classifier showed the best quality according to the AUC metric (AUC 0.79); all other quality metrics (Accuracy, Precision, Recall, F1) for this model were 0.70. The most important features for prediction were the following indicators: patient age, spherical component of refraction before surgery, AL (axial length) before surgery, and UCVA before surgery.
Conclusion. The developed model demonstrated acceptable quality for predicting the outcome of scleroplasty in children 12 months after surgery.
