cui2vec

cui2vec generates vector embeddings for 108,477 medical concepts by training unsupervised word-embedding models on multimodal clinical and biomedical corpora to represent relationships among medical concepts.

Key Features:

• R package: Implemented as an R package for generating and distributing embeddings.
• Multimodal corpus: Trained on 60 million insurance claims, 20 million clinical notes, and 1.7 million full-text biomedical journal articles.
• Vocabulary coverage: Produces embeddings for 108,477 medical concepts.
• Unsupervised word embeddings: Learns relationships among medical concepts using unsupervised word-embedding techniques.
• Co-occurrence fitting: Fits embeddings using co-occurrence data derived from the combined datasets.
• Data integration: Integrates disparate data types into a unified vector space representing medical concepts.
• Pre-trained embeddings: Provides pre-trained embeddings for downstream scientific analyses.
• Benchmark methodology: Includes a benchmark methodology specifically tailored to assess the statistical power of embeddings for medical concepts.
• Comparative performance: Demonstrates state-of-the-art performance in comparative analyses relative to prior methods.

Scientific Applications:

• Semantic characterization: Characterizing semantic relationships among medical concepts across claims, clinical notes, and biomedical literature.
• Clinical and biomedical NLP: Supporting natural language processing analyses of clinical notes and full-text biomedical journal articles.
• Statistical evaluation of embeddings: Assessing the statistical power of embedding-based analyses using the provided benchmark methodology.
• Method benchmarking: Benchmarking and comparing embedding approaches against previous methods.

Methodology:

Unsupervised learning of word embeddings fitted to co-occurrence data from 60 million insurance claims, 20 million clinical notes, and 1.7 million full-text biomedical journal articles, with multimodal data integration and validation via a benchmark assessing statistical power.