iPriorAI

NHS trusts buy scanners and AI software from big companies. Those systems capture enormous value on trust machines: the scans, the reports, the annotations consultants make — and, coming soon, AI models and everything they produce.

Who owns and controls all of that is decided by contracts. Most contracts say nothing. And silence is not neutral — it gets decided in the supplier’s favour later, at exactly the moment the trust needs the data, wants to switch supplier, or is asked to prove where a model came from.

iPriorAi is the audit that happens before that moment: what the law and the contracts actually say, what the queue numbers would actually do, and what has to be in place before anything touches a real patient — with every assumption labelled and every weak point on the table.

This is a catalogue of CT scan collections that universities and hospitals around the world have published for others to use. Think of it as checking which ingredients a company is allowed to bake with.

Why your trust should care: before you buy or build an AI tool, you are entitled to ask what was it trained on — and was that lawful? A model trained on collections marked “no commercial use” can be limited in who may sell it, on what terms, and what happens if the licence is ever challenged. A supplier’s warranty that it holds “all necessary rights” is a promise, not evidence.

• Green (Clear) — the licence allows this use, in writing.
• Amber (Contested) — possible, but only under conditions you would have to sign.
• Red (Silent) — no usable answer: prohibited, or no licence could be found at all.

Look at the downstream outcome column: almost nothing here carries both the timestamps and the patient outcomes that queue-prioritisation research actually needs (INSPECT is the one exception). The scarcity is the finding: the evidence base behind the sales pitch is thinner than it looks.

One thing this register does not, and cannot, verify: whether the patients in these collections consented to their anonymised data being used this way. That was each originating institution’s own process, under its own country’s law, before publication — not something iPriorAI checks or re-checks. Any future Northern Ireland dataset is a different, and closer, question: see the Governance tracker.

For NI trusts, the question is never simply “can we use the data?”. It is a family of questions, and each has its own rights position:

• Can NI-based NHS Trusts use medical imaging data for research?
• Can it be made available in open repositories for the global public good?
• Can it be made available to companies for commercial use?
• Who owns datasets generated by AI using the imaging data from the machines?
• Who at each Trust is responsible to approve such use of data generated in each Trust?

And, because this is an AI audit, the same family extended to the machines that learn:

• Can a supplier use trust data to train or improve its own models — and who owns what comes out?
• Is a model trained on trust data a derivative work of that data (engaging share-alike licences)? Unresolved — and unresolved is not the same as safe.
• May the data leave the trust at all — to vendor clouds, sub-processors, successor companies?
• On exit, can the trust take its data and every artefact derived from it?
• Can the trust publish or license the evidence it generates locally (annotations, validation results, real-world performance)?

How this hits a trust in practice:

• Switching supplier becomes a hostage negotiation: your data, and everything built on it, may live in their systems on their terms.
• Proving innocence becomes impossible: you may be unable to demonstrate that a vendor’s model was not trained on your patients.
• Value walks out: annotations your consultants created can end up inside a product sold back to you — and to other trusts.
• Deleting what you must becomes unenforceable: data in vendor-side logs is outside your retention and deletion regime.
• Nobody can sign: if no named person at the trust is responsible for approving data use, every request becomes an improvisation — and improvisation defaults to no, or to a bad yes.

None of these are exotic disputes. They are the normal consequences of a contract that said nothing.

This is a “what if” machine for the reporting queue. It takes published NI figures and asks: if AI is allowed to jump some scans to the front of the queue, who gets seen faster — and who gets pushed back?

The one idea to hold on to: there is no new radiologist in the machine. Moving one scan forward moves others back. Any tool that shows only the benefits is selling, not informing — so this one always shows both sides on the same screen.

Stroke — quantified, and the strongest case. Untreated stroke destroys brain tissue at roughly 1.9 million neurons per minute (Saver, Stroke 2006). The benefit of clot-dissolving treatment decays by the hour, working best the earlier it is given (Emberson et al, Lancet 2014), and for clot-removal therapy every minute saved measurably reduces disability (HERMES, JAMA 2016). The CT report sits upstream of all of it: every hour a report waits is an hour consumed from a benefit that decays.

Cancer — quantified, and it cuts both ways. A four-week delay in cancer treatment is associated with a mortality increase on the order of 6–13% depending on type (Hanna et al, Nat Rev Clin Oncol 2020 — modelling estimates quoted in that editorial). This matters on both sides of the simulator: prioritised incidental cancers benefit — and routine scans, which also contain undiagnosed cancer, pay the cost of waiting.

Pulmonary embolism and aortic dissection — delay is harmful, but the per-hour quantification is thin, and we say so. Thrombolysis delay is an independent predictor of mortality in acute PE (Beydilli et al, 2016 — small, single-centre), and slower out-of-hours pathways cost lives in a 25,608-case aortic dissection cohort (Kirchhoff et al, JTCVS Open 2026). Useful, honest signal — not a per-minute constant, and not claimed as one.

The hospital side — and where we refuse to invent numbers. Faster reports mean earlier decisions, and NHS England’s national radiology review treats reporting speed as a patient-flow measure (GIRFT Radiology National Report). But there is no robust published figure converting reporting minutes into freed beds or shorter stays — so this tool does not display one. Anyone quoting bed-days per hour of reporting is inventing it.

And the honest counterweight, always: the displaced routine cases are patients too. The same cancer-delay evidence applies to the scan that waited an extra week. That is why this simulator shows benefit and cost together, and why the reviewer question for every run is: net, for whom, and what did the other queue pay?

Every number carries a label. Published — a named study or official statistic says it. Derived — simple arithmetic on published numbers. Assumption — an educated guess, stated openly.

Why your trust should care: when this evidence reaches a committee, nobody can say the numbers were dressed up. The weak points are on the table on purpose — and that is exactly what makes the strong points credible.

Before anyone touches a real patient scan, this list has to be signed off step by step, by named people, with evidence. It protects patients — and it protects the staff involved.

Why every step says “open”: because that is the truth. Nothing has been approved yet, and this tool will not pretend otherwise.

And “not a medical device”? This tool talks about services, statistics and contracts. It never looks at a patient’s scan and never tells a clinician what to do. The day it started re-ordering real scans for real patients it would become regulated medical software — a different product, given a different name on purpose.

Two of these steps ask the question this tool gets asked most often: did the patient agree to this? Research use and commercial use are asked separately on purpose — agreeing to research is not the same as agreeing that a company can profit from what the research finds, and NI has not yet answered either one.