A cohort definition workflow is the sequence that turns a clinical question into a reproducible, executable rule set: which patients enter, on what event, under which code set, with what exclusions, and when they leave. The documented platforms differ less in interface than in what they treat as the durable artifact. OHDSI's Book of OHDSI states plainly that a cohort is not defined by its code set, because a cohort requires specific logic for how the code set is used, such as whether the first or any occurrence of a diagnosis code qualifies [3]. GARDE, a standards-based clinical decision support platform, instead frames cohort identification as a screening population step followed by fact retrieval and evaluation, exchanging data through HL7 FHIR and CDS Hooks [1]. Those two framings produce different answers to the practical questions buyers ask about ownership, inputs, outputs and version boundaries.

What a Cohort Definition Contains Beyond a Code List

The Book of OHDSI defines a cohort as a set of persons who satisfy one or more inclusion criteria for a duration of time, and notes that cohorts are typically defined independently from other cohorts in a study so they can be reused [3]. Its anatomy separates a cohort entry event, inclusion criteria with temporal logic, and a cohort exit, with observation-period boundaries anchoring the start and end [7]. A training deck from the same community lists constraints that read like design rules: one person may not belong to the same cohort multiple times during the same period, a cohort may have zero or more members, and a code set alone is not a cohort [7].

That distinction matters operationally. If the code set is the artifact, then changing a vocabulary release silently changes membership. If the logic is the artifact, the code set becomes a referenced input that can be versioned separately. OHDSI's concept set expressions are described as logical expressions representing a list of concepts in the OHDSI vocabularies, executed as JSON via webAPI into standard SQL [7]. The practical consequence for a population health team is that review should happen at the level of entry event, temporal inclusion and exit strategy, not only at the level of the diagnosis list.

GARDE: Screening Population, Fact Retrieval and Evaluation

GARDE's published architecture describes bidirectional data exchanges between heterogeneous electronic health record sources, population health management systems and clinical decision support components, using HL7 FHIR and CDS Hooks for interoperability and dissemination [1]. Its workflow begins with a screening population step, which the authors describe as an opportunity to preprocess and filter unnecessary data using organization-specific database query or logic before more complex decision support operations are computed [1]. A later step retrieves patient facts, where the issued request specifies which facts to retrieve rather than which screening population, and data transforms to FHIR STU3 or R4 depend on the facts requested [1].

The deployment description is unusually concrete about record ownership. At the Utah site, patient data were accessed directly from the enterprise data warehouse, processed on a local CDS Hooks server, and results were written directly to the EHR's population health management system; GARDE identified 135,817 screening patients, of whom 5,155 (3.8%) met NCCN criteria for genetic testing [1]. Those counts are the authors' reported single-site results for a familial cancer use case, not a benchmark for other cohorts or vendors. The architecture also shows a commercial system in the loop: a figure depicts a population health management application based on Epic's Healthy Planet, with outreach messages sent through the patient portal [1]. GARDE is described as open-source and system-agnostic [1].

Practical detail: separating the definition from the data extract

A workable review habit is to keep three artifacts distinct: the cohort definition, the concept set versions it references, and the data extract it was run against. OHDSI's documentation supports this separation by treating concept sets as expressions executed into SQL rather than as the cohort itself [7], and by describing cohort entry, inclusion and exit as separate components [7]. GARDE's screening-population step shows the same separation at the data layer, where organization-specific filtering happens before decision support operations are computed [1]. When a definition changes, the team can then state whether membership moved because of logic, vocabulary or source data, which is the minimum needed for any downstream quality or research use.

Where Commercial Cohort Builders Put the Definition

John Snow Labs documents a Cohort Builder, and the vendor’s description places definition inside a dashboard that lists cohorts with category, patient and document counts, inclusion logic, creation date and whether new data has arrived since creation [6]. The detail view shows definition tags identifying which ICD, SNOMED or OMOP codes define membership, alongside a clinical description and hierarchical links to related cohorts [6]. Filtering supports diagnoses by those code systems, age range, demographics including gender, race and ethnicity, visit frequency patterns, document volume thresholds and free-text search, with nested AND/OR logic [6].

Two documented steps sit outside pure rule evaluation. A manual exclusion step lets users select multiple patients and remove them, with counts and metadata updating immediately [6]. A documentation step captures a descriptive name, an organizational category, a clinical description and optional parent-child links, which the vendor states supports reproducibility in research workflows [6]. These are vendor claims about its own interface, not independently verified behavior. The editorial point is that manual exclusion creates a membership change with no stated rule behind it, so a governance process needs to record who excluded whom and why, or the cohort's reproducibility claim rests on documentation discipline rather than on the definition alone.

Versioning, Deployment Boundaries and Unresolved Questions

Version boundaries appear in different places across these systems. In GARDE, the FHIR version used for transforms is STU3 or R4 depending on the facts requested, and the FactDB storage model follows the FHIR type of the requested facts [1]. That means a change in FHIR version is an interface-level event, not a cohort-logic event. In OHDSI tooling, the documented execution path runs SQL through SqlRender and DatabaseConnector so code can be translated and executed against a variety of database platforms [3], which places portability in the query layer rather than in the cohort definition. Neither source documents a formal cohort versioning scheme, approval workflow or audit trail, and readers should treat those as open questions to raise with any vendor.

Ownership is similarly split. GARDE writes results into the EHR's population health management system at the Utah deployment [1], so the downstream record of cohort membership lives in a commercial system the research team does not control. John Snow Labs states that every cohort created is immediately available across its platform for analytics, quality improvement and research [6], which implies a single-tenant definition store but does not describe export formats or external reuse. A buyer comparing these approaches should ask where the canonical definition resides, what happens to historical membership when a definition changes, and whether membership can be reconstructed from stored logic alone. The supplied documentation does not answer those questions for either product.

A Suggested Validation Scenario for Cohort Logic Review

Consider a hypothetical, clearly labeled as a suggestion rather than a tested method: a health system wants a cohort of patients eligible for genetic evaluation of familial cancer, mirroring GARDE's documented use case [1]. The team writes the entry event as a qualifying family history observation, adds inclusion criteria for age and encounter type, defines exit at the end of the observation period, and binds each criterion to a versioned concept set. Before release, they run the definition against a frozen data extract, then re-run it after a vocabulary update to see which patients moved and why. This mirrors OHDSI's guidance to think through how a cohort exists across healthcare settings and data sources [3].

The scenario exposes the questions that matter more than interface speed. Does the platform store the logic separately from the code set, as OHDSI's framing implies [3]? Does it record manual exclusions with attribution, given the documented exclusion step [6]? Does it write membership into an external system of record, as in the Utah deployment [1]? And does it expose the FHIR version dependency that governs fact retrieval [1]? A team that can answer those four questions has a defensible cohort workflow. A team that cannot has a saved search whose membership will drift without anyone noticing, which is a governance failure rather than a software defect.