Abstract
PURPOSE: To develop effective machine learning models that analyze pattern visual evoked potentials (PVEPs) to predict the stabilized visual acuity (VA) of patients with treated ocular trauma. METHODS: This experiment included 260 patients (220 males, average age 42.54 years) with unilateral ocular trauma. Four different machine learning algorithms, namely, support vector regression (SVR), Bayesian ridge (BYR), random forest regression (RFG), and extreme gradient boosting (XGBoost), were used to predict best corrected visual acuity (BCVA) values. Various ophthalmic parameters were input into the above algorithms for model training, and the performance of the algorithms was analyzed from the difference between the prediction value and the ground truth. Among the BCVA measured at least 6 months after injury was set as the ground truth. The best-performing model was further developed by tuning different parameter combinations. RESULTS: All models achieved high diagnostic performance, with accuracy values ranging from 0.7875 to 0.8133. The XGBoost model predicted BCVA values with the lowest mean absolute error (MAE), at 0.1598 logarithm of the minimum angle of resolution (logMAR); the lowest root mean square error (RMSE), at 0.2402 logMAR; and the highest accuracy, at 0.8959. CONCLUSION: Promising outcomes in BCVA prediction were achieved by the PVEP-trained machine learning models, which will be helpful in the clinical evaluation of patients after ocular trauma.