The bottleneck was never the learning. Automated biobank-to-OMOP harmonisation for federated analysis.
Project 5 at the Nordic Biobank × NVIDIA Federated Learning Hackathon, Copenhagen 2026
ETL pipeline and federated learning framework for OMOP Common Data Model data.
Federated learning has worked for years, and it still almost never happens: every biobank stores its data differently, and a hand-built ETL per source takes months. We automated that step. Raw biobank data in, harmonised OMOP out, quality controlled, trained across sites. Nothing leaves the institution. Collaboration stops being a technical problem. It becomes a choice.
At every site, an AI agent writes the mapping from a profile of the raw data, a domain expert approves it, a QC gate checks the four OMOP tables, and the dataloader streams them to the GPU.
NVFlare trains across the sites. Only model weights cross a boundary.
Instructions for any coding agent: omop_skill/SKILL.md.
| Mapping | Written blind by an agent. Against OHDSI ETL-Synthea on 1,162 patients: 254,882 of 254,891 measurement values identical, no difference in diagnoses. 8 s for 2.1 GB, no server. |
| Quality control | Caught what both conversions let through, negative LDL and an HbA1c median of 3.9 %, and shows before training whether the sites differ enough to be worth federating. |
| UK Biobank | The same skill maps the raw extract. All eight assessment centres pass the contract. |
One real cohort of 964 patients, split across 1 to 16 sites. Alone, a site falls from 0.87 to 0.77 AUROC. Federated, it holds at 0.87 to 0.92, level with pooled. At 1 million patients and 3.39 billion rows, the scan takes 7.9 s and 6.2 GB of memory.
Before training, the cohort overview shows whether the sites differ enough to be worth federating. The Synthea sites do, the UK Biobank centres are copies of one cohort.
Five Synthea sites: federated matches pooled (AUROC 0.582 against 0.586), while the smallest site alone scores 0.433.
Eight UK Biobank centres, the negative control: the data has no signal and every interval covers 0.5. Single sites still show apparent signal, federation averages it out.
Details: omop_skill/RESULTS.md, examples/synthea_diabetes, examples/ukb_diabetes, and the final slides.
Subproject 1, raw Synthea data of one site to OMOP, then checked against the contract:
pip install -r omop_skill/requirements.txt
python omop_skill/scripts/run_mapping.py --mapping omop_skill/mappings/synthea_3.3.0.yaml --source synthea_cohorts/cohort_2/data/source_compact/site_a --out omop_skill/output/site_a
python omop_skill/scripts/validate_contract.py omop_skill/output/site_a --contract omop_skill/contract_rev1.yamlSubproject 2, federated training over the five sites of cohort_2:
pip install -e .
python examples/omop_t2dm/run.py --sites synthea_cohorts/cohort_2/data/omopA full NVFlare job is in src/omopflare/README.md.
| Path | What |
|---|---|
omop_skill/ |
Subproject 1: raw data to OMOP, and the checks |
src/omopflare/ |
Subproject 2: OMOP dataloader and federated learning with NVFlare |
examples/ |
End-to-end runs (ukb_diabetes, synthea_diabetes) and the benchmarks |
contract/ |
The data contract both subprojects hand over on |
synthea_cohorts/ |
Synthetic cohorts |
presentation/, docs/ |
Slides and project page |
Proof of concept: synthetic data, simulated sites, one real cohort for the benchmark. An AI drafts every mapping, a person approves it. Next: HUNT Cloud over AWS.
charles, solvi, lukas, maria, max, nik, mia, kev