Purchase this article with an account.
D.C. Hoffman, G. Li, K. Nie, K. Nouri–Mahdavi, D.E. Gaasterland, J. Caprioli; Probability of Visual Field Progression With High Dimensional Analysis of Variance . Invest. Ophthalmol. Vis. Sci. 2005;46(13):4666.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose:To evaluate new statistical methods to determine the probability of visual field progression which take into account the spatial and temporal correlations within visual field series. Methods: Five hundred eyes from 396 patients from the Advanced Glaucoma Intervention Study (AGIS) were evaluated. All patients had > 7 Humphrey 24–2 visual field (VF) full threshold exams (mean 15.4 ± 3.7 SD, range 7 to 21) over > 3 years (7.3 ± 1.7, 3.0 to 10.7). Fourier and wavelet transforms were applied to the data to de–correlate spatial dependencies and to reduce dimensionality. To decrease high frequency noise in a series, 1/8th of the transformed data with the highest signal/noise information was analyzed by applying Fan’s adaptive Neyman (FAN) test and a wavelet thresholding test. The FAN test was modeled with and without a temporal correlation structure. Average slopes and pointwise linear regressions (PLR) with Bonferroni corrections were calculated and p values were obtained. Eyes with p values < 0.05 and with slopes < –0.4 (dB/year) were considered to be progressing. These were compared to eyes identified by AGIS scoring and by the Glaucoma Hemifield (GH) PLR method. The latter classifies series as progressing when two or more locations in a GH cluster have p < 0.01.and slopes < –1 (dB/year). Results: AGIS scoring found 28% and GH PLR found 30% of eyes as progressing with a kappa of 0.42. Compared to these criteria, FAN with time found 21% of eyes progressing, wavelet 23%, FAN without time 24%, and PLR 27%. The kappa for FAN and AGIS was 0.41, for FAN and GH PLR 0.42, for wavelet and AGIS 0.43, and for wavelet and GH PLR 0.58. Conclusions: PLR methods generally ignore spatial relationships in VF series while AGIS scoring was not designed to calculate p values for progression. Calculating the probability of VF progression with methods which reduce dimensionality and de–correlate spatial dependencies in VF data show reasonable agreement with conventional methods to detect progression. These techniques may identify progressing VF series undetected by other methods.
This PDF is available to Subscribers Only