DeepSuggest: Using Neural Networks to Suggest Related Keywords for a Comprehensive Search of Clinical Notes

Abstract Objective A large amount of clinical data are stored in notes that frequently contain spelling variations, typos, local practice-generated acronyms, synonyms, and informal words. Instead relying on established but infrequently updated ontologies with keywords limited to formal language, we developed an artificial intelligence (AI) assistant (named “DeepSuggest”) interactively offers suggestions expand or pivot queries help overcome these challenges. Methods We applied unsupervised neural network (Word2Vec) the build keyword contextual similarity matrix. With a user's input query, DeepSuggest generates list relevant keywords, including word variations (e.g., forms, abbreviations, misspellings) other words related diagnosis, medications, procedures). Human is then used further refine their query. Results learns semantic linguistic relationships between from collection notes. Although only able recall 0.54 Systematized Nomenclature Medicine (SNOMED) synonyms average among top 60 suggested terms, it covers relationship our corpus for larger number raw concepts (6.3 million) than SNOMED ontology (24,921) retrieve terms not existing ontologies. The precision averages at 0.72. Usability test resulted achieve almost twice compared Epic (average 5.6 retrieved by 2.6 Epic). Conclusion showed ability improve retrieval when implemented suggesting semantically It promising tool helping users higher rate note searches thus boosting productivity practice research. can supplement query expansion.