Convex Representations Using Deep Archetypal Analysis for Predicting Glaucoma.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early diagnosis is often missed due to asymptomatic disease progression.
• Current methods struggle to detect glaucoma before significant vision loss occurs.

Purpose of the Study:

• To identify clinically relevant patterns of glaucomatous vision loss using convex representation.
• To predict glaucoma onset several years prior to clinical manifestation.
• To develop an automated tool for early glaucoma detection.

Main Methods:

• Developed a deep archetypal analysis model to identify glaucomatous vision loss patterns.
• Projected visual fields onto identified patterns for enhanced representation.
• Implemented class-balanced bagging with neural networks to address class imbalance in prediction.

Main Results:

• Deep convex representation achieved an Area Under the Curve (AUC) of 0.71 for predicting glaucoma up to four years prior to onset.
• This improved upon original visual field features (AUC 0.55) and standard convex representation (AUC 0.61).
• Relevance Vector Machine (RVM) achieved an AUC of 0.64.

Conclusions:

• Deep archetypal analysis combined with balanced bagging classification offers a promising automated tool for glaucoma prediction.
• This method can identify at-risk individuals before irreversible vision loss, potentially impacting clinical practice.
• Early detection through advanced computational models is crucial for managing glaucoma and preventing blindness.