Abstract
Purpose::
To determine the impact of combining results from structural and functional tests to improve discrimination between healthy and glaucomatous eyes compared to each type of testing results alone.To assess the improvement in diagnosis classification by using nonparametric Bayesian statistical modeling.
Methods::
We considered 123 eyes of 123 patients enrolled in the Diagnostic Innovations in Glaucoma Study. Two definitions of glaucoma are used: glaucomatous visual field damage based on repeatable abnormal SAP results, and glaucomatous optic disc based on masked assessment of optic disc stereophotographs. Participants had GDx VCC, Stratus OCT, HRT II imaging, FDT N-30, and SWAP testing completed within a 6-month interval. Diagnostic data are modeled as a latent class model in which one parameter from each test is selected. The Dirichlet process mixture model (DPMM), a nonparametric Bayesian model, is applied. Markov chain Monte Carlo (MCMC) computational techniques are used to estimate these models. The primary outcome measures are posterior distribution estimates of the DPMM. Sensitivity at fixed specificity based on MCMC samples of the DPMM posterior distribution are computed.
Results::
At a 90% specificity cut-off, sensitivity was computed for individual and combined test parameters based on two glaucoma definitions. For individual parameters, the sensitivities are 38% for SWAP/PSD, 46% for FDT/PSD, 40% for HRT/MRC, 42% for GDx/NFI, and 60% for OCT inferior average. Based on the SAP definition of glaucoma, the sensitivity was 60% for the two functional parameters combined (SWAP and FDT), 45% for three structural parameters combined (GDx, OCT, and HRT), and 65% for the five functional and structural parameters combined. Based on the stereophotograph definition, sensitivities for the two functional parameters combined were 55%, 62% for the three structural parameters, and 68% for five parameters combined. The nonparametric Bayesian modeling recovered the patient groups and identified potential subgroups of glaucoma patients.
Conclusions::
Combining parameters from structural and functional tests can improve the sensitivity of glaucoma detection compared to individual tests alone. The Bayesian modeling approach provides a novel and flexible statistical framework for combining tests. Results from the combined Bayesian models may be used to select individual or combinations of testing parameters with the best discrimination properties.
Keywords: clinical (human) or epidemiologic studies: biostatistics/epidemiology methodology • imaging/image analysis: clinical • visual fields