Machine Learning Models for Predicting Treatment-Requiring Retinopathy of Prematurity in the e-ROP Study.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Retinopathy of prematurity (ROP) is a leading cause of vision loss in premature infants.
• Early detection and treatment of treatment-requiring ROP (TR-ROP) are crucial to prevent severe visual impairment.
• Current screening methods rely on manual examination, which can be resource-intensive.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting TR-ROP.
• To assess the utility of image findings at 32-34 weeks postmenstrual age, along with demographic and clinical data, for TR-ROP prediction.

Main Methods:

• A secondary analysis of 771 infants (birth weight < 1251 g) from the e-ROP study was performed.
• Six ML models (KNN, SVM, Random Forest, XGBoost, DNN, Transformer) were trained and evaluated.
• Performance metrics included AUC, accuracy, sensitivity, and specificity.

Main Results:

• ML models incorporating image findings and clinical data achieved AUCs from 0.777 to 0.853.
• The Deep Neural Network (DNN) model demonstrated the highest performance, with an AUC of 0.853, sensitivity of 0.929, and specificity of 0.644.
• Using image findings alone, the DNN achieved an AUC of 0.787.

Conclusions:

• ML models show moderate predictive capability for TR-ROP.
• DNN models offer promising results for early TR-ROP risk identification.
• Further research is needed to enhance ML model performance for clinical application.