Role of Artificial Intelligence in Predicting Visual Prognosis After Open Globe Trauma in Children

Abstract

Background and aims. Open globe eye injuries are a frequent cause of morbidity, amblyopia, and monocular blindness in children. In our study, we propose to establish deep learning models to predict the final visual acuity (VA) prognosis and the risk of amblyopia after an open globe injury in children. Methods. This study involves deep learning models based on a dataset containing 146 variables of 87 patients aged ≤16 years (87 eyes) who had an open globe injury between January 1, 2015, and December 31, 2021. We used the Knime software for predicting the final visual acuity prognosis and the risk of amblyopia. The methods used were the neural network system, the support vector machine (SVM), and the decision tree. Results. The deep learning system was able to predict the risk of having a poor final visual acuity prognosis with good accuracy for both the neural network system and SVM (76.9% and 88.9%, respectively). The identified prognostic factors for poor VA prognosis in the decision tree were low initial visual acuity, wound size >6mm and its shape, the presence of anterior chamber inflammation or abnormal ultrasound. Our study also accurately predicted the risk of amblyopia with good specificity (80.8% and 100%, respectively, for the neural network system and 78.4% and 74.1%, respectively, for SVM). Similarly, the decision tree identified children at high risk of subsequent amblyopia, namely initial visual acuity, presence of a limbal wound, absence of isolated corneal involvement, and presence of postoperative complications. Conclusion. Predicting VA prognosis and the risk of amblyopia after an open globe injury in children could play a major role in identifying high-risk groups to adjust the postoperative surveillance rate and reduce the optical disturbances caused by open globe injuries.