Multicenter evaluation of machine and deep learning methods to predict glaucoma surgical outcomes

TL;DR

This study uses machine learning to predict outcomes of glaucoma surgery based on electronic health records, helping identify patients at risk of poor results.

Contribution

The study introduces a 1D-CNN model that effectively predicts glaucoma surgical outcomes using preoperative EHR data.

Findings

The best model achieved 76.4% AUROC and 71.6% accuracy in predicting surgical failure.

IOP-related failure had the highest prediction accuracy with 82% AUROC.

Model performance slightly declined on external test data by 2–4%.

Abstract

To develop machine learning (ML) and neural network (NN) models to predict glaucoma surgical outcomes, including intraocular pressure (IOP), use of ocular antihypertensive medications, and need for additional glaucoma surgery, using preoperative electronic health records (EHR) from a large multicenter cohort.

This cohort study included 9,386 patients who underwent glaucoma surgery across 10 institutions in the Sight Outcomes Research Collaborative (SOURCE). All patients had at least 1 year of follow-up and 2 postoperative visits with IOP measurements. Models were trained using preoperative EHR features to predict surgical failure, defined as any of the following: IOP remaining above 80% of preoperative value beyond the immediate postoperative period, increased postoperative glaucoma medications, or need for additional glaucoma surgery. Model performance was evaluated on two test sets: an internal holdout set from sites seen during training and an external holdout set.

Of 13,173 surgeries, 8,743 (66.4%) met failure criteria. The best-performing model for overall surgical failure prediction was a one-dimensional convolutional neural network (1D-CNN) with AUROC of 76.4% and accuracy of 71.6% on the internal test set. The top-performing classical ML model was random forest (AUROC 76.2%, accuracy 72.1%). Prediction performance was highest for IOP-related failure (AUROC 82%), followed by increased medication use (80%) and need for an additional surgery (68%). AUROC declined slightly (2–4%) on the external test set.

ML and DL models can predict glaucoma outcomes using preoperative EHR data. Translational relevance: prediction models may support clinical decision-making by identifying glaucoma patients at risk of poor postoperative outcomes.

Figures