Purchase this article with an account.
Feng Gao, Philip Miller, Julia Beiser Huecker, Michael A Kass, Mae O Gordon; The Ocular Hypertension Treatment Study (OHTS): Improving prediction of which participants who developed glaucoma will/will not have visual field (VF) progression in 5 years. Invest. Ophthalmol. Vis. Sci. 2015;56(7 ):3694.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To improve accuracy of predicting which participants who developed glaucoma will/will not have VF progression by the use of joint model to incorporate longitudinal predictors
We analyzed VF data from 164 eyes of 164 participants with glaucoma who had > 7 post-diagnosis VFs (median f/up of 5 years) and > 3 pre-diagnosis VFs. VF progression was defined using post-diagnosis VFs and based on the 2-omitting point-wise linear regression (PLR) algorithm that flags a VF test point if the slope is statistically significant (p≤0.01) and clinically significant (≤-1.0 dB/year). An eye was classified “progression” if at least two adjacent points from the same hemifield met the above criteria. We then compared the accuracy of 3 models using pre-diagnosis mean deviation (MD) to predict which eyes had been classified as “progressed” : Model 1 - a logistic regression including only MD at diagnosis, Model 2 - a logistic regression adding the slope of pre-diagnosis MD estimated by a simple linear regression, and Model 3 - a joint model incorporating the whole series of pre-diagnosis MD. The joint model was fitted using Markov chain Monte Carlo (MCMC) method with WinBUGS software. All models were also compared using simulation studies where the parameters were chosen after the OHTS pre-diagnosis MD data.
32 eyes (20%) were classified as “progression”. Figure 1 shows the ROC curve and area under the curve (AUC) for each model. The joint model that uses all pre-diagnosis MD has the highest accuracy. Table 1 compares the estimated regression coefficients in each model to the “true value”. The greater the difference, the greater the bias in the effect estimated by each model. Estimates in Models 1&2 are badly biased towards null due to the presence of high variability in MD.
Although it is common practice to use the most recent measurements as covariates in prediction models, we show that the joint model which incorporates the whole series of longitudinal measurements (if available) uses information more efficiently and improves the predictive accuracy.
This PDF is available to Subscribers Only