Patient Information Retrieval Based on BERT Variants and Clinical Texts in Electronic Medical Records
Quang Ba Minh Le, Chau Thi Ngoc Vo · 2023
Information retrieval is a task related to search a database for the most relevant similar objects to a given query object. In medicine, patient information retrieval is important to get the patients that are the most similar to a patient being considered. The resulting data can be further used for physicians to produce adaptive treatment plans as well as for other applications such as disease classification, re-admission prediction, and stay-length prediction. Due to its significance, several traditional information retrieval approaches were applied on medical databases. However, there is still a growing need for an effective solution to patient information retrieval in the context where more and more electronic medical records and their clinical texts are captured. In this paper, we focus on this task and propose to perform local learning on the BERT-based embeddings from clinical texts of patients to achieve an effective solution. The advanced properties of BERT variants help better represent each patient using the clinical texts instead of other data types like medication codes and demographic data. The experimental results on MIMIC III database have confirmed the effectiveness of our proposed solution. Above all, the better differences between our solution and the others in F-measure are statistically significant in all the cases.