Purpose: To develop a diagnostic setup with classification rules for combined analysis of morphology [Heidelberg Retina Tomograph (HRT)] and function [frequency doubling technology (FDT) perimetry] measurements.
Methods: We used 2 independent case-control studies from the Erlangen eye department as learning and test data for automated classification using random forests. One eye of 334 open angle glaucoma patients and 254 controls entered the study. All individuals underwent HRT scanning tomography of the optic disc, FDT screening, conventional perimetry, and evaluation of fundus photographs. Random forests were learned on individuals of the Erlangen glaucoma registry (102 preperimetric patients, 130 perimetric patients, 161 controls). The classification performances of random forests and built-in classifiers were examined by receiver operator characteristic analysis on an independent second cohort of individuals (47 preperimetric patients, 55 perimetric patients, 93 controls).
Results: HRT measurements had a higher diagnostic power for early glaucomas and FDT perimetry for glaucoma patients with visual field loss. A combination of all parameters using automated classification was superior to single tests in comparison to the diagnostic instrument with the higher diagnostic power in the respective group. Highest sensitivities at a fixed specificity (95%) in the patients of the present test population were: HRT=32%, FDT=19%, combined analysis=47% in preperimetric patients and HRT=76%, FDT=89%, combined analysis=96% in perimetric patients.
Conclusions: The feasibility of machine learning for medical diagnostic assistance could be demonstrated in patients from 2 independent study populations. A predictive model using automated classification is able to combine the advantages of morphology and function, resulting in a higher diagnostic power for glaucoma detection.