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

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Predicting surgical outcomes in glaucoma is crucial for patient management and resource allocation.
• Current methods for outcome prediction often lack precision and rely on limited preoperative data.

Purpose of the Study:

• To develop and evaluate machine learning (ML) and deep learning (DL) models for predicting glaucoma surgical outcomes.
• To utilize preoperative electronic health records (EHR) for outcome prediction, including intraocular pressure (IOP), medication use, and need for further surgery.
• To assess model performance in a large, multicenter cohort with extensive follow-up.

Main Methods:

• A cohort of 9,386 patients undergoing glaucoma surgery across 10 institutions was analyzed.
• Preoperative EHR data were used to train ML/DL models to predict surgical failure.
• Surgical failure was defined by persistent elevated IOP, increased medication use, or need for re-operation.
• Model performance was validated on internal and external test sets.

Main Results:

• The best overall surgical failure prediction model was a 1D convolutional neural network (1D-CNN) achieving 76.4% AUROC.
• Random forest, a classical ML model, achieved comparable performance with 76.2% AUROC.
• Models showed highest predictive accuracy for IOP-related failure (82% AUROC), followed by medication use (80%) and need for additional surgery (68%).
• Performance slightly decreased on an external test set.

Conclusions:

• Machine learning and deep learning models demonstrate significant potential in predicting glaucoma surgical outcomes.
• Preoperative EHR data are valuable predictors of postoperative success.
• These predictive models can aid clinicians in identifying high-risk patients, potentially improving surgical decision-making and patient care.