Predicting Surgical Interventions for Glaucoma with Clinically Available Data

Abstract

Purpose : To evaluate the accuracy of using patient demographics, clinical measurements, optical coherence tomography (OCT), and visual field (VF) testing in machine learning (ML) models to predict the need for surgical intervention for glaucoma within the next 12 months, 2 years, and 3 years.

Methods : Data collected from participants of the Diagnostic Innovations in Glaucoma Study and the African Descent and Glaucoma Evaluation Study were used to build and evaluate ML models predicting glaucoma surgery. For training, 369 glaucoma surgery cases were identified based on self-reported glaucoma-related ocular surgeries. Sample sizes of non-surgery controls varied with time horizon between 587–592. For modeling, data consisted of demographics (age, sex, race), medication and systemic medical condition (hypertension, diabetes) history, ophthalmic measurements (IOP, CCT, spherical equivalent, axial length) along with 24-2 VF results, and Heidelberg Spectralis RNFL thickness measurements. To predict glaucoma surgery at each time horizon (12 months, 2 years, and 3 years), measurements prior to that time period were included in the modeling. Finally, models were evaluated using area under the ROC-curve, precision, recall, and F1 score on a holdout set of participants from an independent study site. The algorithms tested include random forests, gradient boosting machines, and generalized linear models.

Results : Results for our highest performing models in predicting glaucoma surgical intervention are shown in Table 1. The best performing model was a gradient boosting machine. These models achieved AUC scores of 0.88, 0.91, and 0.91 and precision scores of 0.85, 0.87, and 0.86 at 12 months, 2 years, and 3 years, respectively.

Conclusions : Machine learning models trained on clinical data were able to predict surgical interventions with high accuracy up to three years prior to surgery. Future work will also incorporate full OCT imaging data to improve model predictions. These approaches could provide clinicians with an important tool to predict the need for glaucoma interventions and preserve vision.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Table 1. Overview of models predicting surgical intervention at different time horizons.