May 2004
Volume 45, Issue 13
Free
ARVO Annual Meeting Abstract  |   May 2004
Bayesian Networks to Classify Visual Field Data
Author Affiliations & Notes
  • A. Tucker
    Info Systems and Comp, Brunel University, Uxbridge, United Kingdom
  • V. Vinciotti
    Info Systems and Comp, Brunel University, Uxbridge, United Kingdom
  • X. Liu
    Info Systems and Comp, Brunel University, Uxbridge, United Kingdom
  • D. Garway–Heath
    Glaucoma Research, Moorfields Eye Hospital, London, United Kingdom
  • Footnotes
    Commercial Relationships  A. Tucker, None; V. Vinciotti, None; X. Liu, None; D. Garway–Heath, None.
  • Footnotes
    Support  EPSRC (UK) GR/R35018/01
Investigative Ophthalmology & Visual Science May 2004, Vol.45, 2131. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      A. Tucker, V. Vinciotti, X. Liu, D. Garway–Heath; Bayesian Networks to Classify Visual Field Data . Invest. Ophthalmol. Vis. Sci. 2004;45(13):2131.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Abstract: : Purpose: There is a vast amount of data gathered on patients suffering visual deterioration including Visual Field (VF), Optic Nerve Head (ONH), Retinal Nerve Fibre Layer (RNFL) image and demographic data. In this study, a number of statistical classifiers for the detection of glaucoma, based on VF data, is investigated. Methods:Bayesian Network Classifiers (BNC), Linear Regression (LR) and K–Nearest Neighbour (KNN) are investigated, with a focus on BNCs due to the explicit nature of the models. Data from 102 normal and 78 glaucomatous eyes were used to train the classifiers using 10–fold cross validation. Resulting ROC curves, and the effect of placing a cost function on false positives and false negatives, were explored. Each VF point is related to an RNFL bundle entry point at the ONH and the distance (in degrees at the ONH) between field points can be calculated (Garway–Heath et al. Ophthalmology 2001). The structure of the BNC was explored with respect to RNFL anatomy to assess the discovered relationships. Results:The area under the ROC for the BNC was 0.76 compared to 0.78 for KNN and 0.82 for LR. The BNC is probably at a disadvantage given its higher parameterisation relative to the amount of data available. When applying costs to the misclassification, BNC performed best when specificity is low, whereas LR performed best when specificity was high. The BNC structures clearly reflect expected relationships within the VF. For example, VF points in the nasal step are the most influential in classifying glaucoma. 72% of links are contained within the same RNFL bundle and the mean difference in ONH angle between VF points linked in the BN was 18 degrees (out of a maximum of 180 degrees). Links strongly correlated to the class node show that the nasal step area and the points above and below the horizontal meridian are important. The superior peripheral points, often ignored as having high variability and subject to eyelid artefact, were also highlighted as influential in classifying VFs as glaucomatous. This accords with the pattern of RNFL loss that also results in a nasal step. An unexpected finding was strong correlation to the class node of the superior nasal paracentral point. Change at this location is usually thought to represent more advanced disease. Conclusions: Our study reveals the potential of BNCs for knowledge discovery within ophthalmic databases. They make interactions within the VF explicit, are easily interpreted by non–statisticians and can be queried in order to discover interesting characteristics of visual field deterioration in different ophthalmic conditions.

Keywords: visual fields • computational modeling 
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×