Abstract
Purpose :
To develop a machine learning framework that can automatically identify glaucoma using the peripapillary retinal nerve fiber layer thickness measurements.
Methods :
Retinal nerve fiber layer (RNFL) thickness measurements from 154 normal eyes and 97 eyes with glaucoma were acquired with Spectralis OCT. We developed four independent machine learning classifiers including extra-trees, K-Nearest Neighbors, logistic regression, and AdaBoost and integrated them in a hybrid model to generate a more informed decision. We used a total of 733 RNFL measurements as input to the machine learning classifiers consisting of RNFL thickness in six sectors, average RNFL thickness, and 726 peripapillary A-scans. We used 5-fold cross-validation to develop and test the models. We examined the accuracy of the hybrid model and each independent classifier using annotated instances and area under the receiver operating characteristics (ROC) curves. We selected the most discriminating subset of RNFL thickness measurements based on the Classifier Subset Evaluator (a simple tree classifier) using a Greedy Stepwise Forward Search.
Results :
Sensitivity was 92% for detecting glaucoma eyes correctly them in and specificity was 98% for detecting normal eyes correctly. The area under the ROC of the hybrid model was 0.97. The most discriminating RNFL thickness inputs to the classifier were average RNFL thickness, temporal-inferior RNFL thickness, A-scans 60, 231, and 296. We need to somehow define the topography of these A-scans.
Conclusions :
We developed a hybrid machine learning classifier that detected glaucoma with high specificity and sensitivity based on RNFL thickness measurements. This approach could be valuable in detecting glaucoma in clinical practice and for research purposes.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.