Computational methods for classifying glaucomatous visual field measurements

Abstract

Glaucoma is a common eye disease that affects the optic nerve. It is the second leading cause of visual loss globally and while it can occur in all age groups, it is most common in the elderly. The main symptom of glaucoma is the progressive deterioration of the visual field. Management of glaucoma involves careful monitoring of the progress of disease with regular visual field tests. Accurate identification and early intervention can potentially prevent advanced vision loss. A number of mathematical, statistical, and data mining methods have been proposed to identify glaucomatous progression. However, all criteria used to assess change are hampered by noise that arises from individual visual field measurement. In addition, different clinical trials use different definitions of “progressing”. Currently there is no standard method for classifying changes in visual field measurements. The purpose of this thesis is to improve existing methods and to develop new methods for classification of glaucoma.The thesis first describes a glaucoma modeling software according to a patient’s clinical behaviour. The software can handle age-related visual decline, different types and rates of deterioration, and noise. Simulated data is a good resource for testing the efficiency of different methods in detecting progression, and for developing new methods with minimal cost.The thesis then investigates four classification techniques, including Event Analysis( EA), sequence matching, point-wise linear regression (PLR) and machine learning. For EA methods, the thesis proposed an algorithm “baseline-follow-up” for building a 95% (or 99%) confidence interval using a database of repeated Standard Automated Perimetry (SAP) tests of stable visual fields. Experimental results show that the proposed algorithm can improve the sensitivity compared with other EA methods. A major novel contribution is our introduction of sequence matching techniques to the application of glaucomatous visual field data. Sequence matching techniques typically rely on similarity measure. However, visual field measurements are very noisy, particularly in people with glaucoma, and there is lack of a standard definition of progression. It is therefore difficult to establish a reference dataset including both stable and progressive visual fields. We describe two different matching methods,Weighted Sequence Matching (SM) and Baseline Matching Stable Sequences (BMS). SM uses either the Euclidean or Manhattan distance function to select matches in a stable database R for a given query sequence. BMS uses a baseline computed from a query sequence to match means of stable sequences in R. Matches are used to determine whether a query sequence is progressive or not. For PLR methods, the thesis explores the influence of updating a PLR method by adding or deleting an observation, and discusses the application of Kappa statistic for agreement between methods.We finally investigate the application of machine learning methods for the classification of visual field data. Various input features are defined. The feature datasets are extracted from visual field data, in which each patient has been classified by experts. For this study we used the WEKA package, which provides implementations for Decision Tree, Decision Stump, Naive Bayes, and Bayes Network classifiers, as well as Bagging and Boosting methods for applying the classifiers. The accuracy of classification is presented to illustrate the ability of machine learning for classifying visual field change.