• Predicting POAG Progression With Machine Learning

    By Lynda Seminara
    Selected By: Richard K. Parrish II, MD

    Journal Highlights

    American Journal of Ophthalmology, December 2019

    Download PDF

    Can data-trained machine learning be used to identify glaucoma cases at high risk of progression? In addressing this question, Baxter et al. used a discrete event captured in the electronic health record (EHR)—surgical intervention—as a marker for progressive disease in patients with primary open-angle glaucoma (POAG). They found that some details in the EHR may have predictive value even if eye-specific data are lacking; pertinent information included blood pressure findings and certain classes of medication.

    The authors collected EHR data for 385 patients with POAG who were treated at the same academic institu­tion. The data were integrated into three models: multivariable logistic regression, random forests, and arti­ficial neural networks. Leave-one-out cross-validation was applied. The performance of each model was tested by calculating mean area under the receiver operating characteristic curve (AUC) as well as sensitivity, specificity, accuracy, and the Youden index.

    The analysis showed that multivari­able logistic regression was the most effective model for predicting progres­sive disease that would require surgery (AUC, 0.67). The other models were close behind (AUC, 0.65 for both). In the logistic regression model, higher mean systolic blood pressure was found to significantly increase the likelihood of glaucoma surgery (odds ratio [OR], 1.09; p < .001). Conversely, lower like­lihood of surgery was linked to use of ophthalmic medications (OR, 0.28; p < .001), nonopioid analgesics (OR, 0.21; p = .002), antihyperlipidemic medica­tions (OR, 0.39; p = .004), macrolide antibiotics (OR, 0.40; p =.03), and cal­cium blockers (OR, 0.43; p = .03). The authors acknowledged that the favor­able findings for nonophthalmic drug classes may support the exploration of possible new therapeutic targets.

    Accuracy was similar for the three models, ranging from 0.60 (artificial neural networks) to 0.62 (logistic regression and random forests). The best Youden index was achieved with logistic regression (0.26). The random forests model had the lowest sensitivity and the greatest specificity.

    This type of machine learning provides additional groundwork for developing automated risk predictions from systemic EHR data, which could improve clinical decision-making, the researchers said.

    The original article can be found here.