Predicting worsening visual field via masked dual stage attention recurrent neural network

Based on dual-stage attention recurrent neural network (DA-RNN), a masked DA-RNN (MDA-RNN) model was developed to predict future visual fields (VF). The training dataset comprised 154,105 VF examinations of 22,404 eyes. Performance was measured as the mean absolute error (MAE) between the predicted and actual values of mean deviation (MD), pattern standard deviation (PSD), Visual Field Index (VFI), 54 values of total deviation values (TDV) according to glaucoma severity and the prediction interval. Attention weights were analyzed to determine the characteristics of the MDA-RNN. Prediction errors of MDA-RNN for MD, PSD, VFI, TDV were 1.44 ± 2.05 dB, 1.03 ± 1.09 dB, 4.19 ± 7.07%, and 2.45 ± 1.97 dB respectively. Except for the VFI error, MDA-RNN exhibited a significantly lower prediction error than the masked bidirectional gated recurrent unit (BiGRU) model. The MDA-RNN assigned large weights to the TDV, PDV, MD, and false negative ratio (FN) in the encoder attention. In decoder attention, high scores were assigned to VF with lower FN and higher MD. MDA-RNN focused more on important features and could select more reliable VFs over all time steps, reducing noise in the VF input. The MDA-RNN was robust in predicting severe glaucoma, largely worsening glaucoma, and offered long-term prediction intervals.