Purchase this article with an account.
M. H. Goldbaum, I. Falkenstein, D.-U. Bartsch, J. Hao, I. Kozak, F. Mojana, N. Nigam, T.-W. Lee, T. J. Sejnowski, W. R. Freeman; Relevance Vector Machine Analysis of Multifocal ERGs of Eyes in HIV-Positive Subjects Without Infectious Retinitis Shows Deficiencies Compared to Normal Eyes. Invest. Ophthalmol. Vis. Sci. 2008;49(13):947.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Studies have demonstrated structural and functional inner retinal damage in HIV-positive subjects without infectious retinitis, suggesting subtle retinopathy. We applied multifocal electroretinography (mfERG) of the posterior pole in these patients and HIV negative controls. We used machine learning classifiers (MLCs) to determine if electrophysiologic recordings in HIV subjects differ from those in non-HIV individuals.
Standard mfERG was used in 51 patients with high CD4 counts (86 eyes), 54 patients with a history of low CD4 counts (85 eyes) and 41 age-matched non-HIV subjects (82 eyes). None of the eyes had visible retinopathy. A thread electrode (DTL Plus Electrode, Diagnosis LLC, Lowell, MA) was used for recording. The stimuli were displayed on a CRT monitor with RETIscan software Version 3.20.15 (Roland Consult Elektrophysiologische Diagnostik Systeme, Wiesbaden, Germany). The mean simultaneous response component for the second order kernel was recorded. Implicit times (a and b latencies) and response densities (a and b amplitudes) were measured. Eight cycles were averaged for each subject from 103 hexagonal locations and analyzed with the RETIscan software. We trained relevance vector machines (RVM), which are probabilistic MLCs, with supervised learning on latencies or amplitudes. Ten-fold cross validation separated the teaching cases from the test cases. The area under the ROC curve (AUROC) was analyzed to determine if mfERG patterns of either low CD4 or high CD4 HIV subjects differed significantly from normals.
The age distribution of the three classes of subjects was 38.4±5.6 for normal, 41.4±8.1 for low CD4, and 44.6±8.5 for high CD4. For example the AUROCs for the high CD4 versus non-HIV group for b amplitude and b latency using RVM were 0.630±0.044 (p=.048 compared to chance) and 0.705±0.041 (p=.001), respectively. For low CD4, the b amplitude and latency were 0.531±0.047 (p=.62) and 0.701±0.042 (p=.002). Feature selection ranked mfERG locations for utility in differentiating classes.
mfERG objectively shows cumulative functional damage in HIV retinopathy in both high and low CD4 subjects compared to non-HIV normal retinas. MLCs can ease interpretation of mfERGs. It can be difficult for human observers to learn to interpret mfERGs. In contrast, RVM learns from the data how to detect differences in non-infectious HIV retinopathy that are not apparent to human experts assessing mfERG recordings.
This PDF is available to Subscribers Only