Purchase this article with an account.
Hiroshi Murata, Yuri Fujino, Linda M Zangwill, David Garway-Heath, Ryo Asaoka; Verification of the variational Bayes linear regression using the DIGS and the UKGTS datasets. Invest. Ophthalmol. Vis. Sci. 2017;58(8):2868.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
It is clinically very important to predict visual field (VF) progression accurately. We reported the variational Bayes linear regression (VBLR: Murata H et al. IOVS 2014), which outperformed the ordinary least squared linear regression (OLSLR) for predicting future VF progression. One of the merits of the model was its capability to exploit large datasets to improve prediction accuracy. However, the study was performed with training and validation data obtained at a single institute (the University of Tokyo Hospital). The purpose of the current study is to validate the VBLR model using two external datasets: Diagnostic Innovations in Glaucoma Study (DIGS), and United Kingdom Glaucoma Treatment Study (UKGTS).
The TOKYO data included 7070 eyes of 4166 patients: all University of Tokyo Hospital institutional data with VFs ≧ five times (5.6 ± 2.7 times), was used for training VBLR in the current study, to obtain the prior distribution of VF loss rates. The Baseline MD was -6.8 ± 6.5 dB, and the follow up period was -6.5 ± 2.9 years. DIGS and UKGTS data consisted of 248 eyes of 173 primary open angle glaucoma (POAG) patients and 609 eyes of 313 POAG patients; all subjects had ≧ 10 VF records, excluding baseline VFs. TD values of the 11th VF were predicted using TD values of 52 test points from the second to the tenth VFs (VF2-10) in each eye. The prediction performance was summarized by the root mean squared error (RMSE) using the difference between predicted and actual point wise sensitivities. RMSE was also calculated using OLSLR.
RMSEs (mean + SD) with each dataset are shown in the following table.VBLR resulted in significantly smaller RMSEs than OLSLR (p < 0.05, linear mixed model), especially for shorter VF series, thus showing better prediction performance.
VBLR shows promise for improving our ability to predict progressive VF loss.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only