Objective: Code sets play a central role in analytic work with clinical data warehouses, as components of phenotype, cohort, or analytic variable algorithms representing specific clinical phenomena. Code set quality has received critical attention and repositories for sharing and reusing code sets have been seen as a way to improve quality and reduce redundant effort. Nonetheless, concerns regarding code set quality persist. In order to better understand ongoing challenges in code set quality and reuse, and address them with software and infrastructure recommendations, we determined it was necessary to learn how code sets are constructed and validated in real-world settings.
Methods: Survey and field study using semi-structured interviews of a purposive sample of code set practitioners. Open coding and thematic analysis on interview transcripts, interview notes, and answers to open-ended survey questions.
Results: Thirty-six respondents completed the survey, of whom 15 participated in follow-up interviews. We found great variability in the methods, degree of formality, tools, expertise, and data used in code set construction and validation. We found universal agreement that crafting high-quality code sets is difficult, but very different ideas about how this can be achieved and validated. A primary divide exists between those who rely on empirical techniques using patient-level data and those who only rely on expertise and semantic data. We formulated a method- and process-based model able to account for observed variability in formality, thoroughness, resources, and techniques.
Conclusion: Our model provides a structure for organizing a set of recommendations to facilitate reuse based on metadata capture during the code set development process. It classifies validation methods by the data they depend on — semantic, empirical, and derived — as they are applied over a sequence of phases: (1) code collection; (2) code evaluation; (3) code set evaluation; (4) code set acceptance; and, optionally, (5) reporting of methods used and validation results. This schematization of real-world practices informs our analysis of and response to persistent challenges in code set development. Potential re-users of existing code sets can find little evidence to support trust in their quality and fitness for use, particularly when reusing a code set in a new study or database context. Rather than allowing code set sharing and reuse to remain separate activities, occurring before and after the main action of code set development, sharing and reuse must permeate every step of the process in order to produce reliable evidence of quality and fitness for use.
Lazy Set-Sharing Analysis
