Purchase this article with an account.
R Nagarajan, C Balachandran, D Gunaratnam, A Klistorner, S Graham; Neural Network Model for Early Detection of Glaucoma using Multi-focal Visual Evoked Potential (M-VEP) . Invest. Ophthalmol. Vis. Sci. 2002;43(13):3902.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Purpose: To demonstrate the use of an artificial neural network (NN) model for early detection of glaucoma, using multi-focal visual evoked potential (M-VEP) data from the ObjectiVisionTM perimetry system. Method: The population for this study consists of age-matched groups of 158 glaucoma patients with reproducible field defects on Humphrey Visual Field perimetry 24-2 (HVF), 92 normals and 149 pre-perimetric suspects (IOP≷21mm Hg and or increased cup disc ratio of ≷0.8). A pseudorandom cortically scaled pattern stimulus was presented at 58 locations (29 locations per hemifield, eccentricity 32deg) and the corresponding VEP recorded. A feed forward multilayer perceptron NN model has been developed using normal and glaucoma patients data. Input variables were eye (Left/Right), hemifield (Upper/Lower) and M-VEP amplitudes; the output variable was presence/ absence of a scotoma in a hemifield (HVF glaucoma hemifield test abnormal and a cluster of 3 or more no-rim abnormal points on the total deviation probability plot). Data for 711 hemifields (323 with scotoma and 388 normals) were used in the ratio of approximately 2:1:1 for training, verifying and testing the model, which was then used to forecast the presence of glaucoma in suspects. These predictions were compared with HVF and with Heidelberg Retinal Tomography (HRT) results on a subset of patients (glaucoma 84, suspects 83). Results: ROC area under the curve of 0.99, sensitivity of 0.95 and specificity of 0.94 (overall classification performance of 0.945) indicate a very good fit of the data to the model. At the eye level the model detected 96% of glaucoma cases diagnosed by HVF. This can be compared with HRT, which was abnormal or borderline (using Moorefield's criterion) in only 74% of the cases. While none of the 149 suspects were abnormal on HVF, running them through the model showed 115 (75%) as having early glaucoma. In 83 of suspects, HRT was abnormal or borderline in 41 (49%) compared to the model results of 57 (69%) abnormals. The model and HRT concurred in 30 (36%) cases as having glaucoma. Conclusion: A neural network model with M-VEP inputs can detect glaucoma accurately. Amongst suspects the model reports greater prevalence of glaucoma than HRT, making it a valuable early detection tool. This can be verified by following these cases over time. Further research can lead to development of NN models to detect glaucoma early and track its progression.
This PDF is available to Subscribers Only