Purpose: To develop artificial intelligence (AI) to automatically extract the itemized medical data of patients with glaucoma from an electronic medical record (EMR) system by utilizing Bidirectional Encoder Representations from Transformers (BERT).
Subjects and methods: Between August 2020 and August 2021,1, 122 patients with glaucoma visited the Shimane University Hospital. In this study, 1,091 patients with glaucoma who had visited the hospital more than 5 times were included. Of these, EMRs of 250 patients were extracted as text data. The text data were divided into single words and items corresponding to each word were manually categorized as ground truth labeling to prepare the dataset. Japanese pre-trained BERT generated by Kyoto University was utilized to create an AI model (glaucoma BERT) by using the dataset of 150 patients by transfer learning with 5-fold cross-validation. The performance of glaucoma BERT was evaluated by using the remaining dataset of 100 patients.
Results: Glaucoma BERT automatically extracted data including intraocular pressure, uncorrected visual acuity, corrected visual acuity, subjective refraction, parameters of Humphrey automated perimetry, and prescribed eye drops from EMRs. The performances of glaucoma BERT were 99.9%, 92.8%, 93.3%, and 93.0% in terms of accuracy, precision, recall, and F-measure, respectively.
Conclusion: BERT-based transfer learning enabled development of AI for extracting medical data from EMRs using a limited dataset.
Nippon Ganka Gakkai Zasshi (J Jpn Ophthalmol Soc) 128: 21-29,2024.