R. Kemp Massengill

Glaucoma detection and evaluation through pattern recognition in standard automated perimetry data

BackgroundPerimetry remains one of the main diagnostic tools in glaucoma, and it is usually used in conjunction with evaluation of the optic nerve. This study assesses the capability of automatic pattern recognition methods, and in particular the support vector machines (SVM), to provide a valid clinical diagnosis classification of glaucoma based solely upon perimetry data.MethodsOver 2,200 patient records were reviewed to produce an annotated database of 2,017 eyes. Visual field (VF) data were obtained with HFA II perimeter using the 24-2 algorithm. Ancillary information included treated and untreated intraocular pressure, cup-to-disk ratio, age, sex, central corneal thickness and family history. Ophthalmic diagnosis and classification of visual fields were provided by a consensus of at least two glaucoma experts. The database includes normal eyes, cases of suspect glaucoma, pre-perimetric glaucoma, and glaucoma with different levels of severity, as well as 189 eyes with neurologic or neuro-ophthalmologic defects. Support vector machines were trained to provide multi-level classifications into visual field and glaucoma diagnosis classes.ResultsNumerical validation indicates 70–90% expected agreement between multi-stage classifications provided by the automated system, using a hierarchy of SVM models, and glaucoma experts. Approximately 75% accuracy for the classification of glaucoma suspect and pre-perimetric glaucoma (which by definition do not exhibit glaucomatous defects) indicates the ability of the numerical model to discern subtle changes in the VF associated with early stages of glaucoma. The Glaucoma Likelihood Index provides a single number summary of classification results.ConclusionsAutomatic classification of perimetry data may be useful for glaucoma screening, staging and follow-up.