Abstract
Purpose :
This study aims to evaluate the impact of foveal curvature on Mean Deviation (MD) values in visual field sensitivity.
Methods :
We analyzed 52,902 instances of Optical Coherence Tomography images and corresponding visual field data measured by the Humphrey Field Analyzer (HFA) from 6,385 patients (12,593 eyes) tested at Tsukazaki Hospital between January 2017 and December 2022. We developed machine learning models to predict MD values using parameters such as foveal curvature, ganglion cell layer thickness, retinal nerve fiber layer thickness, and other variables. 80% of the data was randomly selected as training data based on subject ID. We created various machine learning models and selected the best performing one, Extra Trees Regressors, which was then used to analyze the prediction accuracy on the remaining 20% of the test data. A comparative model excluding foveal curvature features was also developed and analyzed. To compare the prediction accuracy of models with and without foveal curvature, a bootstrap method was used for the difference in Mean Absolute Error (MAE) between the two models. The analysis and model development were conducted separately for HFA test modes 24-2 and 10-2.
Results :
In HFA24-2, the model including foveal curvature showed a MAE of 3.49, Mean Squared Error (MSE) of 23.81, Root Mean Squared Error (RMSE) of 4.88, Coefficient of Determination (R^2) of 0.54, Root Mean Squared Logarithmic Error (RMSLE) of 0.68, and Mean Absolute Percentage Error (MAPE) of 3.29, compared to the model without curvature which showed an MAE of 3.49, MSE of 23.98, RMSE of 4.90, R^2 of 0.54, RMSLE of 0.68, and MAPE of 3.19. The 95% confidence interval (CI) for the difference in MAE for both models analyzed using the bootstrap method ranged from -0.17 to 0.03. In HFA10-2, the model with curvature had an MAE of 3.45, MSE of 24.30, RMSE of 4.93, R^2 of 0.77, RMSLE of 0.58, and MAPE of 2.45, as opposed to the model without curvature which had an MAE of 3.52, MSE of 25.59, RMSE of 5.06, R^2 of 0.75, RMSLE of 0.58, and MAPE of 2.40. The 95% CI for the difference in MAE for both models analyzed using the bootstrap method ranged from 0.02 to 0.12, a significant difference.
Conclusions :
Models predicting MD values that include foveal curvature demonstrated relatively high predictive accuracy, particularly in HFA10-2 tests. Foveal curvature may contribute to changes in the visual field.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.