Purchase this article with an account.
Esteban Morales, Parham Azarbod, Abdelmonem Afifi, Fei Yu, Anne Coleman, Kouros Nouri-Mahdavi, Joseph Caprioli; Smoothing Algorithms for Pointwise Visual Field Predictions in Glaucoma. Invest. Ophthalmol. Vis. Sci. 2013;54(15):3928.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The study was conducted to compare four threshold smoothing techniques for predicting the rates of visual field (VF) worsening in glaucoma.
798 patients with primary open-angle glaucoma and 6 or more years of follow-up were included. Thresholds at each VF location for the first four years or first half of the follow-up time (whichever was greater) were smoothed with clusters defined by the Glaucoma Hemifield Test (GHT), the Garway-Heath clusters, nearest neighbor weighting (NN), and weighting by the correlation of rates of all other VF locations. Thresholds for smoothed locations were regressed with a pointwise exponential model. VF predictions derived from the exponential regression were compared with average of the actual VF thresholds at the final two follow-up visits. Since there is no formal statistical test to assess the statistical significance of the difference in root mean square (RMS) values among four approaches for model prediction, the approaches were simply ranked by the RMS values with smaller RMS value indicating better model prediction.
The average (±SD) follow-up time and number of visual field exams used for the smoothing and prediction were 4.1 (±1.3) years and 8.1 (±2.4), respectively. The average (±SD) follow-up time for the final two visits was 8.3 (±2.2) years. The RMS values of the differences between the observed and the predicted thresholds at last follow up for each VF location was 11.8 dB for the GHT clusters, 12.4 dB for the correlation weighting, 12.4 dB for the Garway-Heath clusters, 12.5 dB for the NN weighting, and 12.6 for the raw (unsmoothed) values.
Although the differences among the pointwise VF predictions with the four smoothing techniques and the unsmoothed predictions is not large, the GHT method improved the predictions of final threshold values more than other methods, while the use of the raw values provided the worst predictions. It is worth considering using GHT clustering to improve the accuracy of pointwise predictions of VF outcomes in glaucoma.
This PDF is available to Subscribers Only