(137) A Systematic Review to Summarize and Critically Appraise of Existing Phenotype Libraries Using Electronic Health Records- a FLOW study under the VAC4EU organisation

Background: Observational pharmacoepidemiological studies utilizing algorithmic phenotypes derived from electronic health records (EHRs) are commonly undertaken. Multiple unique phenotype libraries exist which vary by size, regions, coding system, and clinical context. We herein present results from a systematic literature review of existing phenotype libraries created to date.

Objectives: To summarize and appraise the attributes, accessibility, interoperability and portability of libraries created to date.

Methods: We systematically searched three literature databases (Scopus, PubMed, and Web of Science), to identify studies that detailed general phenotype libraries (libraries characterizing diverse phenotypes). Our search combined MeSH terms related to "electronic health records”, "phenotype algorithm" and "phenotype library". Articles were identified for full-text screening and subsequent inclusion via double-screening. Characteristics extracted from articles included size, Unified Medical Language Systems (UMLS) available, clinical/programming knowledge required to construct phenotypes, human and/or computer readable phenotypes, validation process, portability and library management tools.

Results: We identified 9 unique phenotype libraries built across the United States, United Kingdom and Canada. Sizes ranged from 92 to 92,367 phenotypes available per given library. UMLS libraries available included ICD-9, ICD-10, SNOMED-CT, OPCS, RxNorm, LOINC, amongst others. Libraries utilized combinations of code lists and advanced algorithms to build phenotypes; all required clinical and programming knowledge to this end. Four libraries utilized machine learning and NLP to identify phenotypes in addition to manual classification. Phenotype validation processes were diverse and not present in all libraries.

Conclusions: We highlight low convergence and high diversity across existing phenotype libraries. Commonalities include UMLS integration, use of complex algorithms, and both clinical and programming knowledge required to build phenotypes. Development of a large-scale library which can harmonize UMLS based libraries across countries and languages is warranted, both to strengthen phenotyping methodology and produce crucial knowledge for the sustainable conduct of conduct VAC4EU studies.