OCT-based diagnosis of glaucoma and glaucoma stages using explainable machine learning.

Abstract

Glaucoma poses a growing health challenge projected to escalate in the coming decades. However, current automated diagnostic approaches on Glaucoma diagnosis solely rely on black-box deep learning models, lacking explainability and trustworthiness. To address the issue, this study uses optical coherence tomography (OCT) images to develop an explainable artificial intelligence (XAI) tool for diagnosing and staging glaucoma, with a focus on its clinical applicability. A total of 334 normal and 268 glaucomatous eyes (86 early, 72 moderate, 110 advanced) were included, signal processing theory was employed, and model interpretability was rigorously evaluated. Leveraging SHapley Additive exPlanations (SHAP)-based global feature ranking and partial dependency analysis (PDA) estimated decision boundary cut-offs on machine learning (ML) models, a novel algorithm was developed to implement an XAI tool. Using the selected features, ML models produce an AUC of 0.96 (95% CI: 0.95-0.98), 0.98 (95% CI: 0.96-1.00) and 1.00 (95% CI: 1.00-1.00) respectively on differentiating early, moderate and advanced glaucoma patients. Overall, machine outperformed clinicians in the early stage and overall glaucoma diagnosis with 10.4 -11.2% higher accuracy. The developed user-friendly XAI software tool shows potential as a valuable tool for eye care practitioners, offering transparent and interpretable insights to improve decision-making.