SNN-BS: A Clinical Terminology Standardization Method Using Siamese Networks with Batch Sampling Strategy

Abstract

Clinical terminology standardization is important for effective integration and sharing of medical information. It aims to convert clinical colloquial descriptions into standard clinical terminologies. However, the accuracy and efficiency of this task are challenged by the gap between colloquial descriptions and standard terminologies, the slight discrepancy across standard terminologies, and the low efficiency of terminology retrieval. To address these challenges, we propose a novel method called SNN-BS for standardizing clinical terminology based on a Siamese network with a batch sampling strategy. SNN-BS enhances its discrimination ability by sampling a set of terminologies to form a retrieval set with the target terminology. By combing two kinds of similarities, we amplify the differences in features between colloquial descriptions and clinical terminologies while considering deeper semantic relationships. Moreover, we use the lighter Bert-tiny model to encode the terminologies and improve the efficiency of terminology retrieval by reducing comparison numbers through regarding it as a question-and-answer selection task. Finally, we conducted experiments on two datasets to evaluate the performance of our model. The experimental results demonstrate that our method achieves a high level of accuracy, reaching 91.30$\%$ and 90.24$\%$, respectively, which outperforms the baselines.