Purchase this article with an account.
D. R. Lally, G. Wollstein, D. Danks, H. Ishikawa, L. Kagemann, J. S. Schuman, for the Advanced Imaging in Glaucoma Study group,www.AIGStudy.net; Combining OCT, HRT, and GDx Through Machine Learning Classifiers for Glaucoma Detection. Invest. Ophthalmol. Vis. Sci. 2009;50(13):5817.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Machine learning classifiers are computational analyses trained to discover relationships between input parameters and a diagnosis. The purpose of this study was to determine if combining structural measurements from multiple imaging devices as inputs for machine classifiers enhances discrimination between healthy and glaucomatous eyes.
512 subjects with healthy (338 eyes, mean visual field mean deviation (MD) -0.12±1.03 dB), glaucoma suspect (458) and glaucomatous (228, mean MD of the last two groups -1.67±3.42 dB) were scanned with Stratus optical coherence tomography (OCT; Carl Zeiss Meditec (CZM), Dublin, CA), Heidelberg Retina Tomography (HRT3, Heidelberg Engineering, Heidelberg, Germany), and GDx-ECC (CZM) at the same visit. All output parameters from each device (91 total parameters) were used by machine classifiers (generalized linear models, probabilistic decision trees with multiple prediction methods, and support vector machines with multiple prediction methods) to predict the clinically defined diagnosis. Leave-one-out accuracy and 8-fold cross-validation area under the receiver operating characteristic curves (AUC) were calculated using outputs from each device separately and for all possible combinations of devices and outputs. Glaucoma suspect and glaucomatous eyes were grouped together for AUC calculations.
The best AUCs for a single device were obtained using the Recursive Partitioning and Regression Trees determining an expectation over the classes (RPART-exp) classifier: HRT=0.987, OCT=0.981, GDx=0.970. The best AUC for any possible combination of two devices was achieved using RPART-exp with HRT+OCT=0.984. Combining data from all 3 devices had an AUC=0.982 using the same classifier. No statistically significant differences were observed between these AUCs except for the comparison with the GDx AUC. The accuracy analysis demonstrated a slight improvement in diagnosis using combinations of various devices compared to a single device.
Machine classifiers provide good discriminating ability using data obtained from each of the imaging devices. Combining data from multiple devices does not significantly improve discriminating ability.
Clinical Trial: :
This PDF is available to Subscribers Only