Abstract
Purpose :
This study aims to propose a spatio-temporal model for predicting glaucomatous progression in RNFL and describing the geographical and temporal association among RNFL thickness maps.
Methods :
143 eyes from 95 patients with primary open-angle glaucoma were enrolled at the University Eye Center, The Chinese University of Hong Kong (Table 1). Totally, 2,199 RNFL thickness maps with signal strength ≥7 , imaged the optic nerve head (ONH) by Cirrus HD-OCT (Carl Zeiss Meditec) with the “optic disc cube” scan, were included in this study.
The sample was divided into training and testing data. Pixel-by-pixel correlations between RNFL thickness were evaluated in the training data and fitted to a space-time separable exponential correlation model:
correlation{RNFL thickness(location 1, time 1),RNFL thickness(location 2, time 2)}
= η × exp{-δs×(spatial distance) -δt×(time difference)}
for each ONH sector; where η represents the index of measurement reliability, δs and δt represent the exponential rates of decay of correlation across space and time.
Future RNFL thickness maps were predicted by kriging predictor and precision of prediction was evaluated by root mean square error (RMSE) in the testing data.
RMSE of the kriging prediction was compared with RMSE of the prediction by pointwise simple linear regression model based on Wald test under linear mixed effect model. A p-value < 0.005 was considered to be statistically significant.
Results :
Mean RMSEs for the whole RNFL thickness map by the proposed spatio-temporal model were generally smaller for predicting RNFL thickness maps 6 months onwards after 36 – 48 months of monitoring. Unlike the pointwise linear regression model, the mean RMSEs by the proposed model were generally stable across different prediction intervals (Table 2). These suggested the proposed model can provide a more precise prediction of changes in RNFL.
Conclusions :
The proposed spatio-temporal model takes into account the correlation of RNFL thickness across space and time. Rate of decay in correlation can be estimated across space and time to provide a better understanding of progression pattern for each patient. The empirical study demonstrated the model can provide better prediction than commonly used simple linear regression model.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.