An interpretable prediction model for the risk of retinopathy of prematurity development based on machine learning and SHapley Additive exPlanations (SHAP)

Abstract

Abstract Purpose: Building a model to predict the occurrence of ROP for preterm infants based on machine learning method, expecting this model to be widely used in clinical practice. Method:The clinical data of 642 preterm infants (126 children with ROP and 516 preterm infants without ROP) in our hospital were extracted, divided into training and validation sets according to the ratio of 4:1, and the prediction models were constructed separately by six machine learning, and the model with the best prediction performance was screened, and the prediction results of the machine learning models were visualized and interpreted by SHAP method. Results: Among the models constructed by the six machine learning , the model constructed by XGBoost has the best AUC both in the training set (0.96) and in the validation set (0.949).severe pre-eclampsia, apgar 1 min, gestational age at birth, a very low birth weight, blood transfusion, and neonatal hyperglycemia were the candidate predictors for the XGBoost. SHAP showed that apgar 1 min, gestational age at birth, a very low birth weight, blood transfusion, and neonatal hyperglycemia were risk factors for the occurrence of ROP, and severe pre-eclampsia could contribute to the occurrence of ROP. Conclusion: The XGBoost created based on machine learning with the predictive features of severe pre-eclampsia, apgar 1 min, gestational age at birth, a very low birth weight, blood transfusion, and neonatal hyperglycemia showed a high predictive value for ROP. This model could be clinically applied to screen patients at high risk of ROP.