April 2014
Volume 55, Issue 13
Free
ARVO Annual Meeting Abstract  |   April 2014
Predicting progressive glaucomatous optic neuropathy using random forests based on longitudinally collected standard automated perimetry data
Author Affiliations & Notes
  • Lucie Sharpsten
    Hamilton Glaucoma Center, Department of Ophthalmology, University of California, San Diego, La Jolla, CA
  • Juanjuan Fan
    Department of Mathematics and Statistics, San Diego State University, San Diego, CA
  • Shaban Demirel
    Devers Eye Institute, Legacy Health System, Portland, OR
  • Xiaogang Su
    Department of Mathematical Sciences, University of Texas at El Paso, El Paso, TX
  • Chris A Johnson
    Department of Ophthalmology and Visual Sciences, University of Iowa, Iowa City, IA
  • Richard A Levine
    Department of Mathematics and Statistics, San Diego State University, San Diego, CA
  • Footnotes
    Commercial Relationships Lucie Sharpsten, None; Juanjuan Fan, None; Shaban Demirel, None; Xiaogang Su, None; Chris Johnson, None; Richard Levine, None
  • Footnotes
    Support None
Investigative Ophthalmology & Visual Science April 2014, Vol.55, 5619. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Lucie Sharpsten, Juanjuan Fan, Shaban Demirel, Xiaogang Su, Chris A Johnson, Richard A Levine; Predicting progressive glaucomatous optic neuropathy using random forests based on longitudinally collected standard automated perimetry data. Invest. Ophthalmol. Vis. Sci. 2014;55(13):5619.

      Download citation file:


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

      ×
  • Supplements
Abstract
 
Purpose
 

Standard implementations of the Classification and Regression Tree (CART) approach do not allow clustered outcome data with longitudinally collected covariates to be analyzed. The purpose of this study is to test a decision tree approach that can handle correlated binary data using a marginal approach and then extend this using the random forests (RF) method to model longitudinally collected perimetry data. We determine whether the RF method offers improvements over standard CART and the newer marginal approach.

 
Methods
 

Age-adjusted standard automated perimetry (SAP) thresholds, along with other clinical variables gathered at the initial examination of 166 individuals with high-risk ocular hypertension or early glaucoma, were used as predictors in both the standard CART and marginal approaches. Baseline SAP thresholds as well as the slope and p-value from longitudinally collected SAP thresholds were used as predictors in the RF model. The classification variable was a determination of progressive glaucomatous optic neuropathy (pGON) based on longitudinally gathered stereo optic nerve head photographs. Models were compared using area under the receiver operating characteristic curves (AUC).

 
Results
 

The AUCs for the marginal (AUC=0.55) and standard CART (AUC=0.43) approaches were not statistically different (P=0.25). However, using baseline data from both eyes of each individual and their longitudinal SAP data collected prior to the determination of pGON, the proposed RF model had the best AUC (AUC=0.65) and was statistically different from the standard CART model (P=0.01).

 
Conclusions
 

Using visual field and other demographic data the proposed RF model provides better prediction of pGON than the standard CART or marginal models probably because it accounts for inter-eye correlations and the autocorrelations that can occur in longitudinally collected data. Furthermore, the RF model was computationally feasible using this dataset of readily available clinical data, and the results obtained are consistent with previous studies in the ophthalmic literature. Most importantly, the RF approach allows the importance of different visual field locations for predicting pGON to be ranked, which may provide new insight in to the complicated nature of glaucoma management.

 
Keywords: 473 computational modeling • 758 visual fields  
×
×

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.

×