“Automate the chore, keep the signature. The pipeline codes everything; a person still signs every run. That one line is what made clinical ops trust it — and what let sites get answers in days instead of weeks.
— The call I made Challenge:
Four manual, cross-tool jobs that depended on one person remembering
Coding adverse events and concomitant medications, reviewing the queries sites had answered, reconciling AEs against lab results, and knowing whether the trial master file was actually complete — each was a real job, each lived in a different tool, and each depended on a data manager doing it by hand on a schedule nobody enforced.
The failure mode was not dramatic. It was a coding run that slipped a week, a query a site answered that nobody closed, a TMF folder that looked full and wasn't.
- Coding by hand: Listings exported from the EDC, pasted into dictionary tools, results pasted back — hours of work every cycle, easy to get subtly wrong.
- Licensed, versioned dictionaries: MedDRA and WHODrug versions matter for safety reporting; a wrong or stale code is a reportable problem, not a typo.
- Queries nobody closed: Sites answered queries that then sat open, which looked to the site like nobody was listening.
- TMF completeness was unknowable: The only way to know whether a document was filed was to open the folder — and 'filed' said nothing about which version.
Solution:
Pipelines for the mechanics, a person for the judgement
Each job became a pipeline with a human sign-off at the end. The coding run pulls listings from the EDC, submits them to the MedDRA and WHODrug coding services, writes the results to the shared drive and posts a summary to Slack for the data manager to review and sign. Queries follow a written playbook. The TMF is tracked out of the document system into the analytics warehouse, one row per document version, and scored against the DIA reference model.
- Coding pipeline: EDC listings → MedDRA and WHODrug coding APIs → shared drive → Slack summary to the data manager, every two weeks
- A responded-query review workflow that resolves what the site actually answered and follows up on the rest
- AE / labs reconciliation across the trial's safety data
- eTMF tracking from the document system into the analytics warehouse — one row per document version — scored against the DIA TMF Reference Model
Before and after
The chores moved to software. The signature stayed with the data manager.
Workflow
Before
After
Medical coding
AE and ConMed listings exported, re-keyed into MedDRA and WHODrug browsers by hand, results pasted back.
Listings pulled from the warehouse, submitted to both coders automatically, results filed to Drive — reviewed and signed every two weeks.
Query review
Site-answered queries sat open for weeks; closing them was nobody's Tuesday job.
Responded queries surfaced weekly; the ones the site actually answered are closed, the rest get a short follow-up.
AE ↔ labs reconciliation
Cross-checks done in spreadsheets at database lock, when it was expensive to fix anything.
Reconciled continuously against the same warehouse tables the monitors use.
eTMF tracking
A document index that knew the current state and nothing about history.
One row per document version, scored against the DIA TMF Reference Model, queryable in SQL.
Impact:
Minutes of review instead of hours of hands-on work
The coding cycle went from [X hours] of hands-on work to [Y minutes] of review per run, on a cadence that no longer depends on anyone's calendar. Sites get their queries closed, and TMF completeness is a dashboard instead of a folder crawl.
Adoption followed the evidence rather than a pitch: the first automated coding pass ran beside the manual one, and the diff did the persuading.
Coding[X hours] → [Y minutes] per bi-weekly run; every run reviewed and signed by the data manager.
Queries[METRIC: queries resolved per week]; site-answered queries closed instead of left open.
TMFVersion-level completeness tracked against the DIA model, queryable in the warehouse.
TrustChanges shipped the same week they were asked for, with the data manager as a daily user rather than a stakeholder.
“Built with the data manager as a daily user, not a stakeholder. The pipelines plug into the tools the team already lives in — the EDC, the shared drive, Slack, the warehouse — so nothing new to log into.
— How it came together