Development and Validation of Machine Learning Classifiers for Predicting Treatment-Needed Retinopathy of Prematurity
Nasser Shoeibi1 *, Mehrdad Motamed Shariati1 , Mohammad-Reza Ansari-Astaneh1 , Mojtaba Abrishami1 , Mehdi Sakhai2 , Fatemeh Neghabi2
- Eye Research Center, Mashhad University of Medical Sciences, Mashhad, Iran
- Khatam Eye Hospital, Mashhad University of Medical Sciences, Mashhad, Iran
Abstract: To design and evaluate various supervised machine-learning models for identifying premature infants who require treatment based on demographic data and clinical findings from screening examinations.
Methods: We conducted a retrospective review of medical records for infants screened for retinopathy of prematurity (ROP) at our clinic over the past decade. We extracted demographic and clinical data, including eleven features: sex, maternal education, paternal education, birth weight, gestational age, ROP stage, zone of retinal involvement, age at examination, weight at examination, and CPR. We developed and assessed several classifiers: logistic regression (LR), decision tree (DT), support vector machine (SVM), naïve Bayes (NB), K-nearest neighbors (KNN), XGBoost, and artificial neural networks (ANN). The target variable was defined as whether the neonate received any treatment during the follow-up period.
Results: Our analysis included data from 9,692 infants. Among the machine learning models evaluated, the logistic regression model achieved the highest accuracy at 97%. The KNN, XGBoost, SVM, and ANN models also demonstrated strong performance, with accuracies of 96%. In terms of sensitivity (recall), the NB model exhibited the lowest false negative rate, indicating the highest sensitivity.
Conclusion: In the context of premature neonates, accurately diagnosing those who require treatment is crucial. Therefore, from a clinical perspective, prioritizing a model with the lowest false negative rate may be more beneficial than selecting one based solely on the highest accuracy. While AI can enhance decision-making processes by providing real-time risk assessments, these tools must be used to augment—not replace—clinical judgment. Clinicians must remain involved in interpreting model outputs and making final treatment decisions based on a holistic understanding of each patient’s unique circumstances.
