Core Concepts
A machine learning-based multistage system can accurately predict the progression of visual acuity in real-world ophthalmic patients with age-related macular degeneration, diabetic macular edema, and retinal vein occlusion.
Abstract
The authors present a workflow for developing a research-compatible data corpus by fusing data from different IT systems in an ophthalmology department of a German hospital. The extensive data corpus allows making predictive statements about the expected progression of visual acuity (VA) in patients with age-related macular degeneration (AMD), diabetic macular edema (DME), and retinal vein occlusion (RVO).
Key highlights:
The authors found a significant deterioration of visual acuity over time, especially in patients with AMD.
They propose a multistage system that classifies VA progression into "therapy winners", "stabilizers", and "losers" (WSL classification scheme).
An ensemble of deep neural networks achieves over 98% classification accuracy in completing incomplete optical coherence tomography (OCT) documentations, enabling more precise VA modeling.
The VA prediction requires at least four VA examinations and optionally OCT biomarkers from the same time period to predict VA progression within a forecasted time frame, restricted to intravitreal operative medication (IVOM) therapy or no therapy.
The final VA prediction accuracy reaches 69% in macro average F1-score, comparable to the performance of ophthalmologists.
Stats
Significant deterioration of visual acuity over time in patients with AMD (p ≤ 0.0001, large/medium effect size)
Weak significant effect for DME (p = 0.0016, medium effect size)
No significant effect for RVO (p = 0.0607)