Abstract
Diabetic retinopathy (DR) is a leading cause of blindness in American adults. If detected early, DR can be treated to preventing further damage causing blindness, therefore, early detection is very important for the treatment of DR. There is an increasing interest in developing Al technologies to help early detection of DR using electronic health records (EHR). The detailed diagnoses information documented in image reports is a valuable resource that could help detect lesions from the medical image, thus helping early detection of DR. In this study, we examined two state-of-the-art transformer-based natural language processing models, including BERT and RoBERTa, to extract DR-related concepts from clinical narratives. We identified four different categories of DR-related clinical concepts including lesions, eye parts, laterality, and severity, developed annotation guidelines, annotated a DR-corpus of 536 image reports, and trained four transformer-based NLP models for clinical concept extraction. The experimental results show that the BERT model pretrained with the MIMIC III dataset achieved the best strict/lenient F1-score of 0.9503 and 0.9645, respectively.