3D volumetric analysis outperforms 2D linear measurements in predicting vision after surgery for epiretinal membranes. An exploratory machine learning study.

purposeTo analyse pre-operative biomarkers, incorporating and comparing two- and three-dimensional (2D/3D) measurements with high multicollinearity, to predict functional outcomes after surgery in eyes affected by epiretinal membranes (ERMs).

methodsThrough the application of machine learning (XGBoost), proficient in handling feature selection and managing multicollinearity, a predictive model for post-surgical best corrected visual acuity (BCVA) was developed using clinical and anatomical pre-operative parameters including both 2D (linear) and 3D (volumetric) measurements of the retina. Association between pre-operative parameters, including clinical features and OCT-based linear and volumetric measurements of the retina, and post-operative BCVA in eyes affected by ERMs.

resultsEighty-six eyes were included in the study. In the model, foveal retinal volume was the strongest predictor of post-operative BCVA (SHAP 0.0623logMAR). Furthermore, volumes of the outer and inner retinal layers (SHAP 0.0342 and 0.0221logMAR) demonstrated greater predictive value for final vision than 2D measurements, as well as pre-operative BCVA and presence of ectopic inner foveal layers. Features dependence plots were drawn for the most important features, highlighting the non-linear relationships.

conclusionsUsing machine learning algorithms, we demonstrate 3D models have greater utility than traditional 2D parameters, when predicting visual outcome following ERM surgery and we question previous assumptions about the influence of certain linear parameters on visual outcomes post ERM surgery.