Why Turnkey Health Integration Pitches Fail in an NHS Trust
The demo is clean. A sandbox Patient, a smiling Observation, a timeline that updates when someone clicks. Then the trust sends a file from the live EPR and the room goes quiet. The identifier is not an NHS number. Half the diagnoses live in a scanned letter. The “API” is a stored procedure someone wrote in 2016 and is afraid to touch. Turnkey was a slide. The deployment is a database.
What the pitch is selling
A certain kind of agency has learned the vocabulary and stopped there. Interoperability, FHIR-native, deployed in weeks. The sales artefact is a diagram with your logo, the hospital’s logo, and an arrow. Nobody in the room asks which table the arrow reads, who is allowed to read it, or what happens when the EPR supplier ships a patch on a Thursday night.
Search results reward that diagram. Trusts do not. A trust deployment fails in places a generic pitch never budgets for: a subject-access and sharing review, a supplier who owns the schema, a cancer or pathology feed that is monthly XML rather than a stream, and a test environment that is six months behind production. The arrow survives. The project does not.
Where the time actually went
When managing data ingestion for cancer registries across multiple NHS trusts, the pattern was repetitive enough to be boring. About nine in ten stalls were not in the transform. They were in getting a feed the organisation would switch on: an account on the integration engine or the file drop, a lawful basis the information governance lead would sign, and a patient identifier that survived the join to the national record. Legacy EPR data extraction hurdles look like engineering. On the ground they are access, identity, and a schema nobody has documented since the last upgrade.
The mapping, once the file existed, was the short part. Local codes, an extra Z-segment, a diagnosis sitting in a comment field: irritating, finite, testable. The long part was discovering that the “real-time” feed was a nightly batch, and that the nightly batch omitted the ward because a report writer had filtered it out in 2019.
Cancer registration is a bad place to bluff
England’s cancer data are not a vibe. Providers submit the Cancer Outcomes and Services Data set to the National Disease Registration Service. The items are specified. Pathology has its own schema, usually from the laboratory system, not from a hopeful scrape of the EPR. Linkage is on NHS number, at patient and tumour level. A pitch that says “we will FHIR the cancer registry” without naming COSD, the monthly return, and the reject path is selling a product that has not met the recipient.
Real-time cancer registry ingestion metrics, when they are honest, are dull: days from diagnosis to a complete record, NHS number match rate, schema rejects, and the proportion of site-specific items left blank. Messages per second is a vanity number if the registry takes a monthly file. Agencies that lead with throughput have usually never watched a liaison manager send a reject spreadsheet back to a trust.
Rule: If the demo data were authored by the vendor, it is not a deployment test. Ask for a de-identified extract from the EPR you will actually read.
Information governance is not a footer
Information Governance (IG) sign-off failure points show up late because the sales cycle treats them as paperwork. They are design constraints. A feed of identifiable clinical data needs a sharing arrangement, a retention story, a list of processors, and a security position the trust can map to the Data Security and Protection Toolkit. Turn up without those and the technical work stops, even if the parser is finished.
The agencies that dominate a Google results page rarely staff this. They staff a workshop. The workshop produces a backlog. The backlog includes “confirm lawful basis” in week one and it is still open in month four, because nobody senior at the trust was in the room when the arrow was drawn.
What to demand before you sign
Ask for three artefacts, not a platform tour. A field-level map from a real extract to the profile you claim to support. A reject report from that same extract. A one-page note on who signs IG, who owns the EPR change window, and what the feed does when the source system is upgraded. If those three are “in the next phase,” you are buying a slide.
Deployment work of this kind is healthcare data integration against live estates, not a connector catalogue. Where instruments and departmental systems are part of the path, the same scepticism belongs on medical instrument integration middleware: name the message you have seen, or do not claim the interface.
