Efficient Biomedical Entity Linking: Clinical Text Standardization with Low-Resource Techniques

TL;DR

The paper tackles efficient biomedical entity linking by building a low-resource, zero-shot framework that maps clinical mentions to UMLS concepts via a prototype space of entity encodings built from canonical names and synonyms. It couples high-recall candidate generation with multiple disambiguation strategies, notably a parametric reranking and the incorporation of UMLS semantic information, achieving retrieval performance competitive with zero-shot and distant-supervised baselines on MedMentions while maintaining low training and inference costs. Comprehensive evaluation combines top-k retrieval metrics with article-level semantic analyses to reveal strengths and granularity-related limitations, guiding future improvements in context usage and abbreviation handling. The work demonstrates practical impact for standardizing clinical text with reduced computational resources and provides a nuanced view of evaluation beyond exact-match metrics.

Abstract

Clinical text is rich in information, with mentions of treatment, medication and anatomy among many other clinical terms. Multiple terms can refer to the same core concepts which can be referred as a clinical entity. Ontologies like the Unified Medical Language System (UMLS) are developed and maintained to store millions of clinical entities including the definitions, relations and other corresponding information. These ontologies are used for standardization of clinical text by normalizing varying surface forms of a clinical term through Biomedical entity linking. With the introduction of transformer-based language models, there has been significant progress in Biomedical entity linking. In this work, we focus on learning through synonym pairs associated with the entities. As compared to the existing approaches, our approach significantly reduces the training data and resource consumption. Moreover, we propose a suite of context-based and context-less reranking techniques for performing the entity disambiguation. Overall, we achieve similar performance to the state-of-the-art zero-shot and distant supervised entity linking techniques on the Medmentions dataset, the largest annotated dataset on UMLS, without any domain-based training. Finally, we show that retrieval performance alone might not be sufficient as an evaluation metric and introduce an article level quantitative and qualitative analysis to reveal further insights on the performance of entity linking methods.

Figures (6)

• Figure 1: This figure illustrates the sequential flow of our proposed approaches. Starting from the left, we begin with leveraging a neural embedding model to create a prototype space on the UMLS entities. The cosine similarity metric is used to perform semantic search on the queries given the input mentions. The resultant top-$k$ candidates are reranked using the listed methods for disambiguation and finally a comprehensive evaluation comprising of the retrieval performance and semantic similarity is performed.
• Figure 2: This figure highlights the trends associated with the retrieval performance improvement over varying top-k candidates using MiniEL$_0$, MinEL and MiniEL$_{1000}$ models. The improvement in R@1 is more significant as compared to that in R@5 for all the models and reranking methods. It can be observed that the retrieval performance of PARAMETRIC reranking decreases with increase in the top-$k$ (k>15) whereas the performance of SEMANTIC GROUP and SEMANTIC TYPE reranking is consistent across the top-$k$.
• Figure 3: This heatmap illustrates the percentage changes in the number of initial exact, related and missed matches for the MiniEL$_0$ model. The performance preceding the changes is labeled 'FROM' for the rows, while the subsequent performance is denoted by 'TO' for the columns. The experiments are performed on the st21pv version of Medmentions.
• Figure 4: This figure illustrates the disparity in similarity scores ($S_G - S_P$) at the article level (4392 articles), alongside the smoothed retrieval performance (R@1) per article using a moving average with a window size of 200. The region $A$ consists of semantically closer predictions and $B$ consists of semantically farther predictions.
• Figure 5: This figure shows the grid search on the parameters $a$, $b$ and $c$ for optimizing the R@1 performance of the MiniEL$_0$ model using the parametric approach discussed in the section \ref{['rerankingmethodology']}. The optimal combination of $a$, $b$ and $c$ is found to be $5$, $0.1$ and $0.05$, respectively.