Purchase this article with an account.
Leonardo Shigueoka, Edson Gomi, Marcelo Dias, Vital P Costa; Sensitivity and specificity of Machine Learning Classifiers using SD-OCT and Standard Automated Perimetry compared to Glaucoma Specialists, General Ophthalmologists, the OCT database and Anderson's Visual Field (VF) criteria. Invest. Ophthalmol. Vis. Sci. 2016;57(12):384.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To compare the sensitivity and specificity of machine learning classifiers (MLCs) for glaucoma detection using SD-OCT and standard automated perimetry (SAP) to those obtained with the OCT database, Anderson's VF criteria, general ophthalmologists and glaucoma specialists.
In a previous study, 10 MLCs were developed after being tested in 48 healthy and 62 glaucoma patients. The same MLCs were tested in a different population, consisting of 66 healthy individuals and 63 glaucoma patients with early damage. All patients underwent frequency-doubling technology (FDT), SAP and RNFL imaging with the Cirrus SD-OCT. Glaucoma was defined as the presence of a VF defect observed with FDT, along with optic nerve changes compatible with glaucoma and two IOP measurements > 21 mmHg. The sensitivity and specificity of all MLCs were compared with those obtained using Anderson's VF criteria (GHT outside normal limits, PSD<5% or 3 adjacent points in the Pattern Deviation plot with p<5%, 1 of them with p<1%), and the OCT database (defining glaucoma when 2 or more clock hours had RNFL measurements below the 99% CIs). We also measured the sensitivity and specificity obtained by 3 glaucoma specialists and 3 general ophthalmologists who were allowed to view both SAP and OCT exams. Exams were considered glaucomatous when at least 2 of 3 ophthalmologists indicated the presence of the disease.
Mean ages of the glaucoma and healthy populations were 58.2 + 7.2 and 56.2 + 7.1 years (p=0.11). Mean MD of the glaucoma group was -5.33 + 3.99 dB. The best MLC was Naïve Bayes (AUROC=0.918), which resulted in a sensitivity of 81% and a specificity of 87.9% (Table 1). The highest sensitivity was obtained with the MLC (93.8%), and the lowest sensitivity with general ophthalmologists (69.8%). The highest specificity was obtained by glaucoma specialists (95.4%),and the lowest with general ophthalmologists (68.1%) (Table 2).
The MLC we developed using OCT and VF data performed better than general ophthalmologists in differentiating glaucoma from normals. The MLC sensitivity was better than those achieved by OCT or VF data alone. MLCs integrating structural and functional data may be a useful tool for the diagnosis of glaucoma, especially when used by general ophthalmologists.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.
This PDF is available to Subscribers Only