Baxter and colleagues hypothesize that non-ophthalmic data in the electronic health records (EHR) of POAG subjects may be analyzed using machine learning to predict the likelihood of undergoing any glaucoma surgery, including laser surgery, within six months. To do this, the investigators used data from All of Us, a federally funded research program with the goal of enrolling over one million Americans and to use their clinical and other data to enhance innovation in biomedical research, particularly precision medicine. This report represents the first application in ophthalmology of EHR data from this large cohort, currently comprised of over 350,000 subjects.
This investigation expands upon earlier work in which machine learning models were developed using non-ophthalmic EHR data to predict the occurrence of glaucoma laser or incisional surgery in a single-center study of 385 POAG subjects.1 In that study, the best discriminating ability was demonstrated with multivariable logistic regression, with an AUC of 0.67. Higher mean systolic blood pressure and the use of various classes of systemic medications were found to have been associated with subsequent surgery.
In the current study, the investigators used the best-performing model from the singlecenter study and applied it to the All of Us study cohort of 1231 POAG subjects, 286 (23.2%) of whom underwent laser or incisional glaucoma surgery. The previously developed model performed poorly, with an AUC of 0.49. This is unsurprising because decision-making strategies and practice patterns regarding recommendations for surgery are likely to vary by center and surgeon.
All of Us data were used to train new multivariable logistic regression, artificial neural networks, and random forests models using 56 predictors and 5-fold cross-validation, reserving 20% of the data for testing. Accuracy ranged from 0.87 with logistic regression to an astounding 0.97 with random forests.
The random forests method is a machine learning model that utilizes multiple decision trees and often improves predictive performance by allowing for the assessment of multiple variables with an outcome of interest. Though the authors present the top predictor variables, this form of machine learning can be a black box, and it can be unclear in what manner the factors were predictive of subsequent surgery.
The authors suggest the occurrence of laser or incisional glaucoma surgery is a surrogate for glaucoma progression. However, this is not entirely true. In the current practice environment, SLT is a well-accepted first-line therapy for POAG, and minimally invasive glaucoma surgeries (MIGS) are often utilized along with cataract surgery with the aim of reducing the need for ocular hypotensive medications.
When one considers the widely varying practice patterns that exist in the US with respect to the use of laser and incisional glaucoma surgery, particularly MIGS, it is hard to understand how non-ophthalmic EHR data can be used to predict the occurrence of surgery in patients with such a high degree of accuracy as observed in this study. Moreover, one must also remember a substantial proportion of patients decline surgical intervention when it is offered, further confusing the issues.. If the findings of this study are confirmed, they suggest that the development of AI models to identify patients at greatest risk of progression and vision loss is well within reach, particularly if non-ophthalmic EHR data are incorporated in the models.