Translating and evaluating historic phenotyping algorithms using SNOMED CT.

Area of Science:

• Medical Informatics
• Clinical Epidemiology
• Health Data Science

Background:

• Electronic health records (EHRs) require robust patient phenotype definitions for computational research.
• Traditional phenotyping in UK primary care used flat Read term lists.
• Systematized Nomenclature of Medicine-Clinical Terms (SNOMED CT) offers relational concept structures beneficial for phenotyping.

Purpose of the Study:

• To implement and evaluate SNOMED CT-based phenotyping approaches within the Clinical Practice Research Datalink (CPRD) Aurum database.
• To compare the performance of SNOMED CT codelists against gold-standard Read codelists for disease phenotype definition.

Main Methods:

• Developed SNOMED CT phenotype definitions for diabetes mellitus, asthma, and heart failure using 'primary', 'extended', and 'value set' methods.
• Derived SNOMED CT codelists for 276 disease phenotypes semi-automatically.
• Compared cohorts identified by SNOMED CT codelists against manually curated Read codelists in 500,000 patients.

Main Results:

• SNOMED CT codelists identified patient cohorts similar to Read codelists, with F1 scores exceeding 0.93.
• 'Value set' and 'extended' methods showed higher recall but lower precision than 'primary' methods.
• 257 out of 276 phenotypes were represented by a single SNOMED CT concept hierarchy, with 135 achieving an F1 score > 0.9.

Conclusions:

• SNOMED CT offers an efficient and effective method for defining disease phenotypes in primary care research databases.
• SNOMED CT-based phenotyping generates patient populations comparable to those derived from traditional, manually curated codelists.
• The relational structure of SNOMED CT facilitates improved and scalable phenotyping for EHR data analysis.