Author’s Disclosure Block: Raheem Remtulla, none; Guillermo Rocha, none

Abstract Body
Purpose: Visual field (VF) testing is crucial in monitoring glaucoma patients, providing key insights for clinical management. However, the process is often hindered by technician shortages and reliability issues. While attempts to reduce technician dependency have been explored, little effort has been focused on predicting visual field outcomes using retinal nerve fiber layer (RNFL) or clinical data. In this study, we leveraged machine learning to predict the pattern standard deviation (PSD) using clinical inputs. Study Design: Machine learning retrospective study Methods: Publicly accessible data from 499 eyes were collected between 2012 and 2015, including 297 glaucoma cases (primary open-angle glaucoma or normal-tension glaucoma) and 202 non-glaucoma controls. All patients underwent VF 30-2 testing, OCT RNFL Imaging, andstandard clinical evaluation. A neural network model was developed using Levenberg-Marquardt optimization, with a 50-layer hidden network trained on input factors: diagnosis, age, intraocular pressure (IOP), central corneal thickness (CCT), and mean RNFL. The network's target output was PSD. The dataset was split 85% for training, 5% for validation, and 10% for testing. Model performance was evaluated using root mean squared error (RMSE) and Pearson correlation coefficient (r). Additionally, feature importance was assessed by systematically removing individual input factors and observing the resulting change in error rate. Results: The neural network demonstrated efficient training across 12 epochs, with consistent error reduction in training, validation, and test sets. RMSE values were 1.87 for training, 2.25 for validation, and 2.19 for testing. The r was 0.89 for training, 0.85 for validation, and 0.89 for testing, indicating strong predictive accuracy with minimal overfitting. The feature importance analysis revealed that the primary contributors to PSD prediction were mean RNFL, diagnosis, CCT, IOP, and age, listed in order of significance. Conclusion: Our neural network successfully predicted PSD from RNFL and clinical data with strong performance metrics. While RNFL and diagnosis were the most influential factors, CCT's significant role in PSD prediction adds to the growing evidence linking thin CCT with visual field progression. This work demonstrates that neural networks hold the potential to predict or even generate VFs based solely on RNFL and clinical inputs, opening new avenues for glaucoma management with lower technician reliance.