A random forest (RF) machine learning model was used to predict visual field sensitivity using a combination of OCT derived metrics. The RF model was selected due to its ability to find nonlinear relationships in a multidimensional space without requiring large data sets (compared to deep-learning approaches). It also enables the evaluation of contributing factors via feature importance analysis.
20 The model was implemented after including nine features as input: EZ Loss, TRT, ORT, TR MeanI, OR MeanI, TR MinI, OR MinI, x-coordinate, and y-coordinate relative to the fovea. Due to the small number of input features (
n = 9), an explicit feature reduction method was not implemented. These metrics were chosen due to established correlations with visual function. The model was trained to predict the visual field sensitivity at the corresponding locus with mean-squared-error loss criterion. Multiple models were trained in a leave-one-patient-out (LOPO) cross-validation approach where both eyes of the patients were left out when training a single model. Training was performed with bootstrapping where randomly selected data points were used to construct decisions trees to reduce overfitting. Five hundred random estimators
20 and 10,474 maximum samples during bootstrapping were selected empirically as parameters of the model during training. For each eye, predictions were made locus by locus, and were then compared to the ground truth visual field values. Root mean square error (RMSE) and mean absolute error (MAE) were calculated across all predictions for all eyes. Feature importance of the models were analyzed to understand the contribution of different metrics in predicting visual field sensitivity.