Abstract
Purpose: :
To establish criteria for significant change in optic nerve head topography by random permutation of a dataset of images obtained with the Heidelberg Retina Tomograph II.
Methods: :
Data were obtained from a previous study (Nicolela et al., 2006) in which 23 glaucoma patients (mean MD [better eye] -2.3 dB, range -13.0, +1.1 dB) had been followed over 5 sessions over an 8-week period. At each session, three HRT2 mean topographies had been obtained for both eyes. For the analysis presented here, the data were treated as a "no-change" dataset with normal as well as elevated variability characteristics (depending on fluctuation of IOP during the 8-week period). Random- and systematic permutations of the original dataseries were performed by duplicating the images and manipulation of the date stamp. Because alignment artifacts are an important source of variability, each of the 5 sessions was selected to serve as a baselines for a family of permuted series. Permutations were generated a) by deriving all possible in-time combinations of the 5 sessions (respecting the within-session arrangement of the 3 mean topographies) and b) by deriving a subset of 5000 different combinations for which the mean topographies were randomly reassigned across different sessions. Analyses of change were carried out based on regression over time of global and sectoral stereometric indices (rim area) as well as on the probability maps of the Topographical Change Analysis (TCA) of the Heidelberg Eye Explorer.
Results: :
The permutation distributions of rim area trend estimates were often characterized by discontinuities, outliers and large tails. In the randomly permuted datasets, statistically significant (p<5%) trends of rim area (false-positives) were observed more often with linear regression compared to rank-based statistics. In the control eyes, between-session variability appeared slightly (3%, p<0.01) larger than variability observed within a single session, and permutation distributions were systematically larger when a) re-alignment to different baselines were performed (p<0.001) and b) when the within-session order was disregarded.
Conclusions: :
Permutation analysis of serial observation provides a robust method for interpreting change at the optic disc within individual patients. However, because image-alignment issue may mimic apparent change, such analyses need to perform re-alignment to different baselines within the image series.
Keywords: optic disc • imaging/image analysis: clinical • clinical (human) or epidemiologic studies: biostatistics/epidemiology methodology