Abstract
Purpose :
Although open-angle glaucoma (OAG) is often a slow insidiously progressive disease, occasionally patients experience rapid progression (RP), which can lead to irreversible blindness if not detected and treated promptly. We created novel supervised machine learning models using principal component analysis followed by a Soft Voting Ensemble classifier to identify the visits at which patients with OAG are about to experience RP over the next 3 years and may be at risk for blindness.
Methods :
We studied 571 patients with OAG from the AGIS and CIGTS trials data sets. Patients underwent tonometry and perimetry every 6 months for 6 years duration. Our supervised machine learning algorithm processed the patient’s glaucoma trajectory during at least 2 years and then predicted at each visit if they would experience RP over the next 3 years of follow-up. We defined RP as a statistically significant (one sided α=0.05) rate of decrease in mean deviation (MD) of ≥1 dB/year, obtaining the slope from linear regression. We built a simple model using patient demographic characteristics and then a more complex model that, in addition, incorporated past IOP, MD, and PSD data. Model training and hyper-parameter selection were done using 5-fold cross validation, then performance was evaluated on a held out (20%) test set.
Results :
As shown in Fig. 1, demographics alone have some predictive value, with 0.58 AUC at identifying visits where patients exhibit RP over 3 years of follow-up. Our full model that also incorporated tonometric and perimetric data achieved an AUC of 0.83 for predicting visits with RP.
Conclusions :
Supervised learning algorithms that consider both a patient’s demographics and longitudinal results from tonometry and perimetry are capable of identifying with reasonable accuracy the times when patients experience RP of OAG. Incorporating additional parameters such as data from OCT are likely to enhance such predictions. With further enhancement of these models, they may soon be sufficiently accurate for clinicians to incorporate their results into clinical decision-making.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.