OphthaBERT: Automated Glaucoma Diagnosis from Clinical Notes.

Area of Science:

• Ophthalmology
• Natural Language Processing
• Medical Informatics

Background:

• Glaucoma is a primary cause of irreversible blindness globally, necessitating early detection.
• Electronic Health Records (EHRs) are vital for glaucoma research but current diagnostic methods lack accuracy and scalability.
• Existing methods rely on International Classification of Diseases (ICD) codes or manual reviews, which are often inefficient.

Purpose of the Study:

• To introduce OphthaBERT, a clinical language model for ophthalmology.
• To develop a fine-tuned OphthaBERT model for automated glaucoma diagnosis from clinical notes.
• To enhance the accuracy and efficiency of identifying glaucoma patients and subtypes from EHRs.

Main Methods:

• Trained OphthaBERT on over 2 million diverse clinical ophthalmology notes.
• Fine-tuned OphthaBERT for binary and subtype glaucoma diagnosis extraction.
• Evaluated OphthaBERT's performance against state-of-the-art models and ICD codes.

Main Results:

• OphthaBERT demonstrated superior performance in masked token prediction and binary glaucoma classification, especially in low-data scenarios.
• The fine-tuned model significantly improved subtype diagnosis classification (0.23-point increase in macro-F1 over ICD codes).
• OphthaBERT achieved strong external validation performance for binary classification at Wilmer Eye Institute.

Conclusions:

• OphthaBERT offers an interpretable and equitable framework for ophthalmology language modeling.
• Automated glaucoma diagnosis using OphthaBERT enhances accuracy and efficiency compared to traditional methods.
• This model facilitates earlier and more precise glaucoma detection, crucial for preventing blindness.