What is Epic Cosmos, and when should you use it?

A research team may need to follow recorded care across several health systems. An operations team may need to explain activity within one hospital. Both work with Epic data, but the records needed to answer their questions can differ substantially.

The choice of data determines which events a study can observe. It also determines which relationships the team can examine. Establish those boundaries early, because they affect the conclusions that can be used in a publication or an operational discussion.

Research across participating health systems

A researcher can use Cosmos to examine recorded care across participating health systems. Epic describes longitudinal records that combine inpatient and outpatient charts from different organizations. They can extend the history available to a study beyond one institution's extract.

Diagnoses, medications and specialty-specific information are among the data described in the Epic Cosmos overview. Those fields help a team assess whether the proposed population and events are available before developing the full analysis.

Coverage remains part of interpreting a broader record. Care outside the contributing sources may be unobserved. If the conclusion depends on whether a follow-up event occurred, an absent record needs to be understood within that coverage.

Compare the observed population with the population the study intends to describe. More records do not by themselves establish that the two are equivalent. Explain relevant differences when applying the findings beyond the people represented in the analysis.

A question the records can answer

Begin with the population and the event the study intends to measure. An initial cohort count helps assess feasibility, but its definition needs examination before it supports an analysis. Check that the recorded event represents the event named in the research question.

Specify when observation starts and what counts as the outcome. Encounter dates, diagnosis dates and medication order dates describe different events. Substituting one for another can change the study's meaning even when the same people remain in the cohort.

Observation time also affects interpretation. Someone with a long recorded history has more opportunity to show a follow-up event than someone whose record begins recently. The analysis should explain how it handles different observation periods and incomplete histories.

Examine missing values before treating a field as evidence of absence. A blank observation can reflect incomplete documentation, an unavailable source or a workflow that never collected the information. Those possibilities may require different treatment in the study.

This work is easier to review when the definition is written alongside the query. The analyst and subject specialist can then inspect the same choices: who enters the cohort, which records count and what the result can reasonably establish.

Access and analytical tools

Access follows the participating organization's process. Its Cosmos or research informatics team can explain the tools, approval and training required for the proposed work. That establishes the analytical route available to the researcher.

Aggregate exploration helps refine the question and estimate a cohort. To implement detailed logic on line-level records, a researcher may need SQL and a data science environment. Those are different kinds of work, with corresponding access and training requirements.

Stanford's guidance sets out training for its Data Science Virtual Machine access, while Yale publishes its own access process. A researcher should use the requirements supplied by their participating organization for the intended tools.

Learning the tool includes learning what its filters do to the population. An early query review with someone who understands the fields can reveal an unintended selection. That review is useful before the same logic is applied throughout the study.

Reproducible findings and controlled exports

Save the definition, query version and analysis date with each finding. Together, they provide a basis for investigating a different result when the study is repeated. The team can compare its method and the data available on each occasion.

Older records can change as documentation is added or corrected, as Stanford's guidance explains. The same query may therefore return a different answer later. Retaining the method and analysis date helps distinguish a source update from a revision to the study.

Establish the permitted output before arranging to share a dataset. Epic's Cosmos governance rules retain line-level data in the secure portal. Stanford also describes review of aggregate files and images. These conditions affect what a collaborator can receive.

Include the population, period, exclusions and relevant gaps with a finding. A collaborator can use that description to assess the method, including when they cannot inspect the underlying line-level records. The result then carries the context needed to interpret it.

Questions that need local records

An institutional operating question may depend on local records beyond the clinical history. Reviewing clinical activity against staffing requires workforce assignments for the same periods. A study-operations question may also need research records maintained outside Epic.

Approved Clarity or Caboodle access can supply the Epic part of that analysis. The implementation connects those records to other sources and applies shared definitions and permissions. Local staffing allocations or department assignments can be as important to the result as the clinical events.

Databasin implements this local analysis: source connections, business views and the reports or AI tools people use to explore them. Cosmos work follows the participating organization's process. The two routes address different evidence needs within an institution.

A program may use Cosmos for one study and institutional data for an operating review. When discussing either finding, identify its source and purpose. This keeps the interpretation tied to the population and process the analysis examined.

Example: defining a follow-up cohort

This illustrative study looks for a recorded follow-up within 30 days of a qualifying encounter. The team needs rules for entry, observation and missing evidence before it can interpret the count.

The count describes events found under these rules. An absent follow-up record remains a question about the available evidence; the definition explains how the study treats it.

Before starting an analysis

1. Define the observed eventIdentify the population, starting event and outcome. For each, establish which available field represents the event the study intends to measure.
2. Examine observation and gapsExamine the available history and missing values. Explain how they affect who enters the cohort and what can be concluded from the result.
3. Establish the access routeEstablish access through the participating institution for the intended tools and purpose. Include the output and its export conditions in that discussion.
4. Preserve the methodSave query versions, definitions and analysis dates with each finding. They give the research team a basis for investigating differences on a later run.

Sources