Visual acuity prediction on real-life patient data using a machine learning based multistage system.

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Intravitreal operative medication therapy (IVOM) is standard for age-related macular degeneration (AMD), diabetic macular edema, and retinal vein occlusion.
• Predicting visual acuity (VA) and detecting vision loss is challenging due to heterogeneous and incomplete real-world data.
• Existing treatments often fail to prevent long-term vision decline.

Purpose of the Study:

• To develop a research-compatible data corpus by fusing data from different IT systems.
• To create a predictive model for visual acuity (VA) progression in patients undergoing IVOM.
• To classify patient responses to therapy into 'winners', 'stabilizers', and 'losers' (WSL).

Main Methods:

• A data corpus was created by integrating ophthalmology IT systems.
• Deep neural networks were used for Optical Coherence Tomography (OCT) biomarker classification, achieving >98% F1-score.
• A multistage system predicted VA progression using at least four VA examinations and optional OCT biomarkers.

Main Results:

• The OCT biomarker classification achieved an F1-score exceeding 98%, improving data completeness.
• The VA prediction model achieved a 69% macro average F1-score, outperforming ophthalmologists (57.8% F1-score).
• Significant vision deterioration was observed over time in AMD patients.

Conclusions:

• The developed data corpus and AI workflow enable predictive modeling of VA progression.
• The WSL classification scheme effectively categorizes patient responses to IVOM.
• The AI model shows promise in predicting vision outcomes, aiding clinical decision-making and patient management.