[Paper Review] Clinical Concept Embeddings Learned from Massive Sources of Medical Data.

Andrew L. Beam, Benjamin Kompa+5 more|arXiv (Cornell University)|Apr 4, 2018

Biomedical Text Mining and Ontologies28 citations

TL;DR

This paper introduces cui2vec, a method that leverages GloVe and word2vec algorithms on a massive fusion of 60 million insurance claims, 20 million clinical notes, and 1.7 million biomedical journal articles to learn clinical concept embeddings for 108,477 medical concepts. The approach achieves state-of-the-art performance on clinical benchmarks and provides pre-trained embeddings and an interactive tool for public use.

ABSTRACT

Word embeddings have emerged as a popular approach to unsupervised learning of word relationships in machine learning and natural language processing. In this article, we benchmark two of the most popular algorithms, GloVe and word2vec, to assess their suitability for capturing medical relationships in large sources of biomedical data. Leaning on recent theoretical insights, we provide a unified view of these algorithms and demonstrate how different sources of data can be combined to construct the largest ever set of embeddings for 108,477 medical concepts using an insurance claims database of 60 million members, 20 million clinical notes, and 1.7 million full text biomedical journal articles. We evaluate our approach, called cui2vec, on a set of clinically relevant benchmarks and in many instances demonstrate state of the art performance relative to previous results. Finally, we provide a downloadable set of pre-trained embeddings for other researchers to use, as well as an online tool for interactive exploration of the cui2vec embeddings.

Motivation & Objective

• To evaluate the suitability of GloVe and word2vec for capturing medical relationships in large-scale biomedical data.
• To construct the largest known set of clinical concept embeddings (108,477 concepts) by combining diverse data sources.
• To benchmark cui2vec on clinically relevant tasks and demonstrate superior performance compared to prior methods.
• To release pre-trained embeddings and an interactive tool to support reproducibility and broader research use.

Proposed method

• The authors apply GloVe and word2vec algorithms to a unified corpus combining insurance claims, clinical notes, and full-text biomedical articles.
• They leverage theoretical insights to unify the two algorithms, enabling consistent training across heterogeneous data sources.
• The model learns dense vector representations (embeddings) for 108,477 medical concepts, indexed by Clinical Universal Identifier (CUI).
• Data from 60 million members, 20 million clinical notes, and 1.7 million journal articles are aggregated and preprocessed for training.
• The embeddings are trained end-to-end using skip-gram and continuous bag-of-words architectures, optimized via negative sampling.
• The resulting cui2vec model is evaluated on clinical benchmarks and released with a downloadable dataset and interactive visualization tool.

Experimental results

Research questions

• RQ1 Can GloVe and word2vec effectively learn meaningful medical concept representations from heterogeneous biomedical data sources?
• RQ2 How does combining claims data, clinical notes, and full-text articles improve the quality and coverage of clinical embeddings?
• RQ3 Does the cui2vec model outperform existing methods on clinically relevant downstream tasks?
• RQ4 To what extent do the learned embeddings capture clinically relevant semantic and syntactic relationships?

Key findings

• The cui2vec model achieves state-of-the-art performance on multiple clinically relevant benchmarks, outperforming previous approaches.
• The integration of claims data, clinical notes, and full-text articles significantly enhances embedding coverage and quality for 108,477 medical concepts.
• The model demonstrates strong performance in capturing semantic relationships between medical concepts, as validated on benchmark tasks.
• The pre-trained cui2vec embeddings are publicly available for download, enabling reuse across diverse biomedical NLP applications.
• An interactive online tool is provided to explore the semantic relationships within the cui2vec embeddings in real time.
• The unified framework of GloVe and word2vec enables consistent and scalable training across large, heterogeneous medical text sources.