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

Purpose: To evaluate machine learning (ML) models for predicting treatment-requiring retinopathy of prematurity (TR-ROP) using image findings at 32 to 34 weeks of postmenstrual age, along with demographic and clinical characteristics.

Methods: This secondary analysis included 771 infants with a birth weight of less than 1251 g who had at least one imaging session by 34 weeks postmenstrual age and at least one subsequent ROP examination for determining TR-ROP by ophthalmologists in the Telemedicine Approaches to Evaluating Acute-Phase Retinopathy of Prematurity (e-ROP) Study. Six ML models (K-nearest neighbors, support vector machine, random forest, extreme gradient boosting, deep neural network [DNN], and transformer) were evaluated for predicting TR-ROP. Prediction performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity.

Results: Using image findings and demographic and clinical data, ML models achieved AUCs ranging from 0.777 (K-nearest neighbors) to 0.853 (DNN), sensitivity ranging from 0.765 (extreme gradient boosting) to 0.929 (DNN), and specificity ranging from 0.644 (DNN) to 0.698 (transformer). Using image findings alone, the DNN performed best with an AUC of 0.787, sensitivity of 0.729, and specificity of 0.725.

Conclusions: ML models using image findings, demographics and clinical characteristics moderately predict TR-ROP, with DNN achieving the highest AUC and sensitivity. Although ML models may provide tools for the early identification of high-risk infants for close monitoring and timely treatment of TR-ROP, future research is needed to improve their performance.

Translational relevance: ML has the potential to predict TR-ROP risk based on early image findings, demographics, and clinical characteristics.