We stood up the full Parthenon analyses pipeline end-to-end: 11 comparative effectiveness studies across 10 disease areas, 46 generated cohorts, and 26 executed analyses including R-based CohortMethod propensity score matching on populations up to 68,000 patients. Along the way, we found and fixed every null-safety bug that only surfaces when real analysis results hit the frontend.
We ran every analysis type in Parthenon — estimation, prediction, SCCS, evidence synthesis, pathways, characterization, and incidence rates — against our full Acumenus CDM with 1 million synthetic patients. Thirteen seeded analyses. Seven different OHDSI methodologies. One session.
This post covers what happened when we moved from the 2,694-patient Eunomia demo dataset to production scale, the bugs that only surface at a million patients, and the hard-won lessons about propensity score modeling on synthetic data.
The OHDSI HADES ecosystem is remarkable. CohortMethod, PatientLevelPrediction, SelfControlledCaseSeries — these R packages encode decades of pharmacoepidemiology methodology into reusable software. In theory, you point them at an OMOP CDM database, call a few functions, and get publication-ready causal inference results.
In practice, getting these packages to run correctly in a modern production environment required solving problems that no documentation warned us about.
This is the story of what we encountered building Parthenon's R runtime — a Plumber API sidecar that executes HADES analyses against a 1-million-patient CDM — and the specific, reproducible bugs we had to fix before a single analysis could complete.