Predicting postoperative visual acuity in epiretinal membrane patients and visualization of the contribution of explanatory variables in a machine learning model

Abstract

Background The purpose of this study was to develop a model that can predict the postoperative visual acuity in eyes that had undergone vitrectomy for an epiretinal membrane (ERM). The Light Gradient Boosting Machine (LightGBM) was used to evaluate the accuracy of the prediction and the contribution of the explanatory variables. Two models were designed to predict the postoperative visual acuity in 67 ERM patients. Model 1 used the age, sex, affected eye, axial length, preoperative visual acuity, Govetto’s classification stage, and OCT-derived vector information as features to predict the visual acuity at 1, 3, and 6 months postoperatively. Model 2 incorporated the early postoperative visual acuity as an additional variable to predict the visual acuity at 3, and 6 months postoperatively. LightGBM with 100 iterations of 5-fold cross-validation was used to tune the hyperparameters and train the model. This involved addressing multicollinearity and selecting the explanatory variables. The generalized performance of these models was evaluated using the root mean squared error (RMSE) in a 5-fold cross-validation, and the contributions of the explanatory variables were visualized using the average Shapley Additive exPlanations (SHAP) values. Results The RMSEs for the predicted visual acuity of Model 1 were 0.14 ± 0.02 logMAR units at 1 month, 0.12 ± 0.03 logMAR units at 3 months, and 0.13 ± 0.04 logMAR units at 6 months. High SHAP values were observed for the preoperative visual acuity and the ectopic inner foveal layer (EIFL) area with significant and positive correlations across all models. Model 2 that incorporated the postoperative visual acuity was used to predict the visual acuity at 3 and 6 months, and it had superior accuracy with RMSEs of 0.10 ± 0.02 logMAR units at 3 months and 0.10 ± 0.04 logMAR units at 6 months. High SHAP values were observed for the postoperative visual acuity in Model 2. Conclusion Predicting the postoperative visual acuity in ERM patients is possible using the preoperative clinical data and OCT images with LightGBM. The contribution of the explanatory variables can be visualized using the SHAP values, and the accuracy of the prediction models improved when the postoperative visual acuity is included as an explanatory variable. Our data-driven machine learning models reveal that preoperative visual acuity and the size of the EIFL significantly influence postoperative visual acuity. Early intervention may be crucial for achieving favorable visual outcomes in eyes with an ERM.