Development, comparison, and internal validation of prediction models to determine the visual prognosis of patients with open globe injuries using machine learning approaches

Abstract Introduction Open globe injuries (OGI) represent a main preventable reason for blindness and visual impairment, particularly in developing countries. The goal of this study is evaluating key variables affecting the prognosis of open globe injuries and validating internally and comparing different machine learning models to estimate final visual acuity. Materials and methods We reviewed three hundred patients with open globe injuries receiving treatment at Khatam-Al-Anbia Hospital in Iran from 2020 to 2022. Age, sex, type of trauma, initial VA grade, relative afferent pupillary defect (RAPD), zone of trauma, traumatic cataract, traumatic optic neuropathy (TON), intraocular foreign body (IOFB), retinal detachment (RD), endophthalmitis, and ocular trauma score (OTS) grade were the input features. We calculated univariate and multivariate regression models to assess the association of different features with visual acuity (VA) outcomes. We predicted visual acuity using ten supervised machine learning algorithms including multinomial logistic regression (MLR), support vector machines (SVM), K-nearest neighbors (KNN), naïve bayes (NB), decision tree (DT), random forest (RF), bagging (BG), adaptive boosting (ADA), artificial neural networks (ANN), and extreme gradient boosting (XGB). Accuracy, positive predictive value (PPV), recall, F-score, brier score (BS), Matthew correlation coefficient (MCC), receiver operating characteristic (AUC-ROC), and calibration plot were used to assess how well machine learning algorithms performed in predicting the final VA. Results The artificial neural network (ANN) model had the best accuracy to predict the final VA. The sensitivity, F1 score, PPV, accuracy, and MCC of the ANN model were 0.81, 0.85, 0.89, 0.93, and 0.81, respectively. In addition, the estimated AUC-ROC and AUR-PRC of the ANN model for OGI patients were 0.96 and 0.91, respectively. The brier score and calibration log-loss for the ANN model was 0.201 and 0.232, respectively. Conclusion As classic and ensemble ML models were compared, results shows that the ANN model was the best. As a result, the framework that has been presented may be regarded as a good substitute for predicting the final VA in OGI patients. Excellent predictive accuracy was shown by the open globe injury model developed in this study, which should be helpful to provide clinical advice to patients and making clinical decisions concerning the management of open globe injuries.