Real-world evidence · Pharma observational teams
Stop rebuilding the dataset for every study question
Most real-world evidence time goes into the same work, repeated: agree the definitions, rebuild the cohort, re-derive the variables, explain the attrition, then do it all again when the source refreshes. CuRE is designed to turn that into one approved study specification that produces a traceable, analysis-ready dataset, on data that can stay in your own warehouse
From question to dataset
Define once, defend every number after
The hard part of real-world evidence is rarely the statistics. It is agreeing what a patient, an exposure and an outcome mean, applying that the same way every time, and proving it afterwards
01
Analysis-ready datasets from one specification
CuRE is designed to build each analysis-ready dataset from a single declarative study specification: the cohort, index date, windows, variables, outcome flags and time-to-event endpoints, picked from a fixed menu rather than written as free-form SQL. Each specification is designed to be approved through a governed rule registry before it runs, and to produce one row per subject, or per subject and endpoint, delivered as CDISC Dataset-JSON, with every variable traced back to the source rows it came from
One spec in, a traceable dataset out
02
Attrition you can defend, refreshes you can explain
Each dataset release is designed to record its source snapshot, policy versions and specification, and how many people each cohort step removed. A release is designed never to change once made: a source refresh produces a new release beside the old one, which must reproduce on a re-run before it goes live. A drift report compares two releases, showing how many people entered or left and which variables moved, with small cells suppressed
Immutable releases and drift reports
03
A semantic layer, not a folder of code lists
Value sets, operational definitions and the vocabulary they draw on live in one governed layer. Concept sets and cohort definitions are versioned by content, so a version never changes once made and an analysis can pin the exact one it used. Terminology edits move from draft through steward review to approval, definitions are shared across teams and partner organizations under governance, and a set of value sets and definitions exports as one content-hashed package
Governed value sets and definitions
04
The measures every claims study rebuilds
CuRE is designed to derive lines of therapy from drug exposures, using governed rules your organization chooses and approves for its disease area. Healthcare resource use is counted per patient in a window around the index date: inpatient admissions and days, emergency and outpatient visits, pharmacy fills and, where the source carries them, costs. Vaccine doses are derived from product and administration codes and de-duplicated across EHR and claims records
Line of therapy, resource use, vaccine exposure
Before the analysis
Know what you are analyzing
A dataset is only as good as the mapping underneath it. CuRE makes that work visible instead of leaving it in someone's ETL scripts
01
Profile the raw source first
Scan a source before mapping it: tables, row counts, missingness and distinct values per column, kept aggregate so no record values leave it
02
Keep a register of what did not map
Every source value that failed to map is kept verbatim, grouped by table and field, and sealed into a register with its disposition, including values no reviewer was ever shown
03
Harmonize units and codes
Measured values are normalized to UCUM units, and local codes map to standard OMOP concepts through governed, reviewable mappings
04
Review mappings by confidence
AI proposes mappings with a confidence tier on every row. Stewards review them, and acceptance is audited over spot-checked rows rather than assumed
05
Check quality the OHDSI way
OHDSI Data Quality Dashboard checks run against the mapped data and are summarized beside the pipeline that produced it
Where your data lives
Connectors to the platforms you already run
Your data is already spread across a warehouse, EHR feeds, claims files, EDC exports and registries. CuRE connects to them rather than asking you to move everything first. For a lakehouse such as Databricks or Snowflake, CuRE is designed to run cohorts and analyses inside your own account, so person-level data stays where it is and only governed results come back
Every connector, by categoryData lakes and warehouses
Designed to run inside Databricks or Snowflake; Amazon Redshift
EHR and EMR
FHIR R4, SMART on FHIR, HL7v2, C-CDA
Claims
X12 837 claims, 835 remittance
EDC
Medidata Rave, Veeva Vault EDC
Labs and registries
HL7v2 results, lab reports, SAS, Stata, SPSS
OHDSI tools
ATLAS, HADES
Your own AI agents
Let your agents ask the data, inside the rules
Many RWE teams are building their own AI agents. CuRE gives them a governed way in: an MCP server through which an agent can search the vocabulary and run and characterize cohorts, with each call limited to the scopes the agent was granted and recorded in an audit trail alongside the person who started it
Results that carry tenant data pass the same release check every other export passes, so an agent cannot see more than an analyst could
Bring us a study question
Tell us the question your team keeps rebuilding a dataset for. We will show you what one specification, one semantic layer and your own data can do with it