Abstract
Purpose :
Machine learning is an application of artificial intelligence capable of identifying trends in data without being explicitly programmed. Glaucoma clinical encounters involve processing large quantities data to decide whether to escalate intraocular pressure (IOP) lowering therapy. Manual data interpretation is both time consuming and a potential source of medical error. We hypothesize that machine learning algorithms (MLAs) can be used to model individual or group provider practice patterns and serve as a tool to augment physicians' decision making ability.
Methods :
We perform a retrospective chart review of consecutive patients seen in glaucoma clinic with primary open angle glaucoma or normal tension glaucoma managed on one or two IOP lowering eye drops (N=50). Parameters investigated include demographics, IOP, central corneal thickness (CCT), cup-to-disc ratio (C:D), nerve fiber layer thickness (RNFL) and visual fields. We assign a binary clinical decision to each encounter: escalate therapy (new drop, SLT, surgery) vs. maintain therapy. We use Matlab to develop a simulated training dataset (N=500) and create a model using supervised classification MLAs such as decision trees, nearest neighbors and support vector machines. Model prediction accuracy is evaluated internally with 5-fold cross validation and externally with the retrospective clinical dataset.
Results :
Bootstrap-aggregated (bagged) decision tree MLAs were found to have the highest prediction accuracy of algorithms investigated. Accuracy increased as a function of the training set size and tree complexity. The optimized model accuracy is 89% by internal cross-validation with the training dataset and 72% by external validation with the clinical dataset. Parameters most frequently used by the MLA include current IOP, max IOP, C:D and RNFL thickness.
Conclusions :
MLAs based on decision trees provide an intuitive framework that most closely resembles the clinical decision making process. We believe larger training and clinical datasets will lead to increased prediction accuracy. Applying MLA based models to new encounters may provide physicians with an additional tool to guide decision making.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.