Abstract
Purpose: :
To compare the receiver operating characteristics of a novel test based on population statistics versus traditional t-test of the regression coefficient in visual field sequences.
Methods: :
We established the empirical distribution of regression coefficients when plotting visual field indices (MD, sLV, deviation in clusters and others) as obtained from Octopus G examinations of stable visual fields against time (population A). Such distributions are specific for given number of examinations and given interval between examinations. However, scaling regression coefficients with a suitable function of these parameters makes such distributions comparable, and allows for pooling in one single distribution. Population A consists of sets of 4 to 8 visual fields of 100 eyes which were judged to be stable according to HRT criteria (topographic change analysis and progression of stereometric parameters). The percentiles of this pooled distribution define our test for the individual regression coefficient to deviate from zero at given levels of significance. The method was then tested with a population B of glaucomatous visual fields not satisfying the clinical criteria of population A (52 eyes, each with a set of 4 up to 21 visual field examinations).The ROC curves obtained were compared with those obtained by judging the same data (populations A and B) with the traditional t-test of the regression coefficient to deviate from zero. For this comparison, a modified version of the publicly available software Octopus Field Analysis OFA V2.2 has been applied.
Results: :
Areas under curve increase by up to 50% with the novel test. Besides, a marked increase of sensitivity (at same specificities) can be observed when the number of examinations involved increases. This applies to both methods.To give an example, the sensitivity was analyzed for all sub segments of 4 consecutive visual fields at a specificity level of 95% (11 of 214 stable sequences judged as progressing). The sensitivity for MD was 16% for t-test based statistics and 43% for population based statistics. In sLV (PSD) the numbers were 10% and 48% respectively and in ARA (Abnormal Response Area) they were 18% and 51%.
Conclusions: :
Compared at the same specificities, the novel population based method is markedly more sensitive than traditional t-tests of the regression coefficient.
Keywords: visual fields • perimetry • development