Some "healthy" glaucoma patients are actually in more danger of going blind than those with severe vision loss.
Doctors traditionally prioritize patients based on how much vision they have already lost. A new deep learning model analyzed thousands of cases and found that current severity is a terrible predictor of how fast the disease will move. High-risk patients can have nearly perfect vision today but possess a trajectory that leads to total blindness in a fraction of the time. This system identifies these fast-progressors before they notice any symptoms at all. It means medical resources can be shifted to save the vision of people who look fine on a standard eye chart.
Deep Kernel Learning for Stratifying Glaucoma Trajectories
arXiv · 2605.00708
Effectively stratifying patient risk in chronic diseases like glaucoma is a major clinical challenge. Clinicians need tools to identify patients at high risk of progression from sparse and irregularly-sampled electronic health records (EHRs). We propose a novel deep kernel learning (DKL) architecture that leverages a Gaussian Process (GP) backend. The GP's kernel is defined by a transformer-based feature extractor applied to clinical-BERT embeddings to model glaucoma patient trajectories from mu