Purchase this article with an account.
E. Bitrian, D. Mock, K. Nouri-Madhavi, A. Coleman, A. Afifi, F. Yu, J. Caprioli; Prediction of Regional Visual Field Outcomes in Glaucoma Through Rate Modeling. Invest. Ophthalmol. Vis. Sci. 2010;51(13):3995.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate the ability to predict regional visual field outcomes with rate modeling techniques.
389 eyes with treated primary open-angle glaucoma and a minimum of 6 years of follow-up were included. Data for the initial four years or first half of the follow-up time, whichever was longer, were used to predict the thresholds at 51 locations across the 24-2 Humphrey visual fields. Linear, quadratic and exponential regression models were explored for goodness of fit of pointwise threshold trends over time. An exponential model demonstrated the best overall fit and was therefore used to predict the thresholds at the end of follow-up. Threshold sensitivities and confidence intervals of the 51 locations at the end of the follow-up were predicted. The correlations of the predicted values with the actual values were determined. Two data-smoothing methods were then used to weight the thresholds at each location: nearest neighbor (8 adjacent locations), and nerve fiber layer (NFL) clusters. The adjacent locations were weighted inversely proportional to the square of the distance to the reference location in both methods. The 80% prediction intervals for predicted thresholds were calculated and plotted, together with gray scale representations.
The average (± SD) mean deviation of the baseline visual fields was -10.8 (± 5.4) dB. Mean follow-up was 8.2 years. The correlation R² between predicted and final measured values for individual threshold sensitivities was 0.32. After data smoothing with the nearest neighbor and NFL clusters, R² increased to 0.55 and 0.58, respectively. The average of the predicted minus observed thresholds also improved from 6.85 dB in the original model to 3.68 dB and 2.7 dB in the nearest neighbor and NFL cluster techniques, respectively. An example of the 80 % prediction interval for NFL clusters is shown in Figure 1.
Data smoothing methods improve the accuracy of predicting regional visual field outcomes. An exponential model with NFL cluster data smoothing performed best to predict visual field outcomes, and provides confidence intervals for clinically useful interpretations.
This PDF is available to Subscribers Only