Subspace Sampling with Mixture of Gaussian Using SAP Results for Discriminating Between Healthy and Glaucoma Eyes

Abstract

 

Purpose:

 

To introduce a new probabilistic machine learning classifier, Subspace Sampling with Mixture of Gaussian (SSMOG), and compare it with Relevance Vector Machines (RVM) on multiple datasets using standard automated perimetry (SAP) for classifying healthy and glaucomatous eyes.

 

Methods:

 

66 eyes of 66 healthy controls and 156 eyes of 156 glaucoma patients with glaucomatous appearing optic discs and/or repeatable abnormal visual fields recruited from the Diagnostic Innovations in Glaucoma Study (DIGS) were included. SAP measurements were made using HFA II with 24–2 SITA, and several datasets were evaluated: 52 threshold values (24–2 test points) plus age (Thr), 52 pattern deviation plot values (PD) and 52 total deviation plot values (TD). SSMOG is a dimension–reducing method that allows the use of MOG classifiers on high dimension data by sampling muliple subsets of the full dimensional data set with many repetitions with replacement. In this study, for each dataset, 9 features were randomly chosen from the full feature set and SSMOG was trained on these subsets of features from all eyes. This sampling and training procedure was repeated 50 times. The final output of SSMOG was the combination of 50 MOG by averaging. Areas under receiver operating characteristic curves (AUROC) for classifying eyes as healthy or glaucomatous were computed to compare SSMOG results to RVM results (Bowd, IOVS, 2005). Ten–fold cross validation was used for both SSMOG and RVM for unbiased ROC areas.

 

Results:

 

Results are presented in the Table below. No significant differences between SSMOG and RVM performance were observed.

 

Conclusions:

 

SSMOG allows the use of MOG classifiers on high dimensional data and performs similarly to RVM. SSMOG is useful for classifying eyes as healthy or glaucomatous using SAP data. Like RVM, SSMOG explicitly models probability density distributions and offers a probabilistic output that is desirable for clinical use.  

 

Keywords: perimetry • visual fields • computational modeling