Machine Learning Classifiers Can Detect Subtle Field Defects in Eyes of HIV Subjects

Abstract

Purpose:: HIV-positive subjects describe visual changes in eyes that appear normal and have normal appearing visual fields, suggesting subtle retinopathy. We applied machine learning classifiers to determine if visual fields in HIV subjects differ from normal visual fields.

Methods:: Standard 24-2 automated perimetry (SAP) with the Humphrey Field Analyzer from one eye of HIV subjects with high CD4 counts (n=79) and low CD4 counts (n=66) were compared with SAP’s from age-matched normal subjects (n=61). None of the eyes had visible retinopathy. Support vector machines (SVM) and relevance vector machines (RVM) perform well on SAPs. One advantage of RVM is that it can give the probability of HIV for each eye tested. These machine learning classifiers were trained with supervised learning on absolute sensitivity. Ten-fold cross validation separated the teaching cases from the test cases. The area under ROC curve (AUROC) was analyzed to determine if the fields of either low CD4 or high CD4 HIV subjects differed significantly from normal. Feature selection was applied to rank visual field locations for utility in diagnosis.

Results:: The age distributions of the three classes of subjects were 46.3±8.0 for normal, 46.4±8.9 for low CD4, and 48.7±8.2 for high CD4. The optimized AUROCs for SVM were 0.839±0.036 for the low CD4 group and 0.676±0.045 for the high CD4 group. The low CD4 and high CD4 groups each significantly differed from normal (p<0.00005 and p=0.006 respectively) with SVM. RVM results were comparable. Visual field locations most useful for classification of low CD4 eyes were in the superior temporal field, both nasal and temporal to the blind spot. The most useful field locations of the high CD4 groups were scattered.

Conclusions:: In both the low and high CD4 groups, in eyes that had no visible retinopathy, there were significant visual field abnormalities. The SAPs in all of the high CD4 and most of the low CD4 eyes appeared normal; hence, SVM and RVM found field defects hidden from human observers. Superior temporal field defects in low CD4 eyes suggest damage primarily in inferior nerve fibers, as our group has revealed with OCT and other imaging modalities. Machine learning classifiers are sensitive detectors of subtle field defects in otherwise normal appearing eyes in HIV-positive subjects.

Keywords: AIDS/HIV • perimetry • computational modeling