A national referral hub for Health New Zealand

Health New Zealand triages referrals by hand. A clinician reads each one, checks it against criteria that are partly written down and partly held in experience, assigns a priority, and writes back when it doesn't fit. This proposal is for a national service that does most of that work automatically.

The referral hub would receive referrals from referrer systems through one published API, assess each against the receiving specialty's criteria using AI, and return one of five outcomes: accepted with a priority, declined with advice, more information needed, redirected to a better test or service, or passed to a clinician. Specialties define the criteria and decide how much of the decision the hub may make on its own. Health NZ owns the standard, the criteria registry, the decision rules and the data.

The benefit is triage done without a human for the majority of referrals. The API, the data standard and the specialty forms exist to make that safe. The metric that matters is the share of referrals decided with no human touch while staying inside clinically agreed error limits.

Community referred radiology (CRR) is the right first service. National primary care referral criteria for ultrasound, CT and X-ray have applied since September 2025, with national priority codes, and four regional triage hubs are being set up to apply them. Using the volumes and triage rates published in the CRR hub design, triaging every community ultrasound and CT referral by hand takes roughly 20,000 clinician hours a year, about 12 full-time equivalents (our estimate; see section 13). Most of those referrals are straightforward against written criteria. A working prototype already checks CRR referrals against all 38 exams of the national criteria (section 4). The return on investment in section 13 is for CRR alone, carrying the whole cost of the hub; every later service shares the platform, so that figure is the floor for the hub's return.

The AI is constrained by design. It reads the referral against the specialty's criteria as written, and must cite each criterion it relies on and quote the referral for each condition. Code checks every citation and quote, sets the priority from the criteria data, and applies a decision policy set by the specialty. Accepts can be automated; declines go to a person. Every automated decision records which criterion, which version and which model produced it. Automatic acceptance is switched on one exam and tier at a time, after the hub has run silently alongside human triagers and the agreement has been measured.

Health NZ should own the parts nobody sells and that carry public accountability: the referral standard, the criteria registry, the decision policy, the audit trail and the data. It should buy the parts the market already sells well: FHIR servers, integration engines, form rendering, and the language models themselves.

Section 16 lists the decisions this paper asks for. The first is permission to run a four-month discovery that confirms the procurement lever with primary care software vendors, loads the CRR criteria into a registry, and measures the current triage baseline.

Referral triage is skilled work done at volume, and the way it is done differs by service, district and region. Community radiology shows the pattern clearly because Health NZ has documented it.

Triage is manual, varied and partial

The CRR hub design describes a different arrangement in each region. Two Northern districts have dedicated GP liaison triagers working in the RIS. Te Manawa Taki and Central have no triage hub, some audit, and referrals on bpac forms. In Te Waipounamu, Canterbury and Southern triage in ERMS before the radiology service prioritises and protocols in the RIS. Across the country triage is done by radiologists, medical imaging technologists, sonographers and GP liaisons, depending on the district.

Coverage is incomplete. Direct-to-private referrals often never pass through a Health NZ RIS, so they can only be audited after the event. X-ray triage is left as it is because it is judged lower risk and the hubs are not resourced for it.

Quality varies, and so do decisions

Clinical audits by existing triagers between 2018 and 2024 found that between 10% and 50% of direct referrals did not meet the local criteria of the time. At Waitematā the decline rate was about 2% before a GP liaison triaged every community ultrasound and CT referral, rose to 17% once they did, then fell by about one point a year as written advice taught referrers what the criteria required.

Human triage also drifts. The Waitematā liaison team noticed their own prioritisation shifting over a twelve-month period, which is why the hub design builds in cross-auditing between triagers.

Systems are fragmented

Referrals arrive through HealthLink eReferral forms, ERMS, bpac forms, practice-specific templates and, still, fax. The national messaging standard for referrals, HISO 10011, is built on HL7 version 2.4 messages. Each district maintains its own forms and its own interfaces to its own patient administration system and RIS.

The opening

The CRR hub design itself anticipates change: it notes that who triages ultrasound and CT may shift as the Radiology Transformation and Data and Digital programmes progress, or where audit shows that some referral types are low risk. This proposal is one way to make that happen, and the same mechanism can then be applied to specialty outpatient referrals, where consultants and registrars do the triage.

The hub combines four functions that are usually separate systems: a single front door for referrals, assessment and triage against criteria, routing to the right service or provider, and visibility of demand. Bundled as one launch they would move at the speed of the slowest. The spine is the front door and the data standard behind it. Everything else plugs into that.

For each referral the hub returns one of five outcomes:

• Accepted, with a priority code and target date, and routed to the service or provider that will do the work.
• Declined, with the criterion it was assessed against, nationally consistent wording, and advice or a HealthPathways link.
• More information needed, naming exactly what is missing. Most of these are caught before the referral is sent (section 6).
• Redirected to a more appropriate test or service, where the specialty has defined that route.
• Passed to a clinician, with the AI's reason and quotes, when a red flag is present, the likely outcome is a decline, any automatic-acceptance check fails, or the exam and tier is not yet automated.

CRR is the strongest first candidate because most of the hard groundwork has been done. The National Community Referral Criteria for Imaging (Part I), version 2.0, published on 9 April 2026, cover 38 exams across adult CT, ultrasound and X-ray and paediatric imaging, with 758 individual criteria. They were developed from evidence-based guidelines and specialist advice, agreed through consultation with GPs, hospital specialists and radiologists, and endorsed by the Radiology National Clinical Network and the National Clinical Governance Group. They let GPs, urgent care doctors and primary care nurse practitioners refer directly for the listed examinations.

Every exam uses the same tiers: acute assessment (same-day hospital care rather than community imaging), acute (within 24 or 48 hours, set per criterion), P2 (within two weeks), P3 (within six weeks), P4 (deferrable, 6 to 12 weeks) and S codes for a specified date. Two more tiers carry no timeframe: an alternative pathway (a different exam, specialist advice or a HealthPathway), and not routinely funded. The CRR hub design already calls for a national change-control process for criteria, standardised wording for declined referrals, and cross-auditing of triage decisions. The referral hub mechanises those things.

A working prototype exists. The community triage tool in HNZ Radiology Tools, on a beta site and not in production, holds all 38 exams word for word, checks a referral against them for PCRLs, and lets GPs check a referral before they send it. This section follows how it works. Its only measurement so far is a run on 46 made-up referrals whose expected answers have not had clinical review, where Claude Opus 5.5 agreed with the expected outcome in 39. A check takes 10 to 20 seconds and costs about 4 cents once the criteria are cached.

Keeping the criteria as written

The criteria do not need rewriting as rules. They are held exactly as published: each exam as its rows of tiers, and each criterion, AND/OR connector, footnote and guidance box as a tree with a stable identifier such as us-abdomen-adult.p3.1. The priority belongs to the row a criterion sits in, so it is read from the data, not from the model. The extracted text is checked word for word against the PDF, and a clinician signs off each exam against it. A correction or a new national version loads alongside the old one, so every past decision stays tied to the wording it was checked against.

All of Part I comes to about 44,000 tokens, small enough to give the model in full and cache between calls. There is no step that picks a criterion first, so the model can say when a different exam fits better. The hub adds a thin layer of its own around each criterion: any look-back window for previous imaging, and whether its exam and tier are switched on for automatic acceptance. The record below is a real criterion; the hub fields are proposed.

How a referral is checked

The model reads the referral against the whole criteria set and returns a fixed structure: the outcome (meets criteria, needs more information, does not meet, acute assessment, or discuss with PCRL), the criteria it matched with a quote from the referral for each condition, any red flags, alternative pathways or not-funded items that fit, what is missing, follow-up questions, and a short reason. It does not set the priority and does not score its own confidence, because models rate themselves poorly.

Code then checks the result. Every cited criterion must exist in the requested exam, every quote must appear in the referral, the priority comes from the tier of the matched criteria, red flags override, and an answer that does not hold together goes to a PCRL. Confidence is worked out from that checked result. A referral is accepted automatically only when the outcome is meets criteria with no red flags and nothing missing, every condition has a quote found in the referral, all matched criteria sit in one tier, prior imaging dates rule out any look-back window, and that exam and tier has been switched on after a shadow pilot. Declines are not automated in CRR: the criteria say non-matching referrals are considered case by case, so every likely decline goes to a PCRL.

What changes for the CRR hubs

The regional hubs keep their purpose; the work inside them changes. Primary care radiology liaisons spend less time reading straightforward referrals and more on the review queue, the advice line, audit and referrer education, which the hub design identifies as where most of the long-term value lies. GPs can run the same check before they send, answer its questions in place and attach a check record the PCRL can look up, so fewer referrals come back for more information. Because automated assessment costs little per referral, direct-to-private referrals can be triaged before the scan rather than audited afterwards, and X-ray can be brought into scope without extra triage staff.

Four views: where the hub sits, what is inside it, how decisions are made, and how the integration estate changes over time.

Keeping the criteria as written, and checking the model's output, matters. Rewriting the 758 CRR criteria as rules would be a large clinical job, and would put a second version of the criteria between the specialty and the decision, one that could drift from what the specialty published. Instead the model reads the published wording, and its output is narrow and checkable: which criteria by identifier, a quote from the referral for each condition, and what is missing. Code rejects any identifier that does not exist in the requested exam and any quote that is not in the referral, and takes the priority from the tier the criterion sits in. A failure is then either a misreading, which the quotes show, or a gap in the criteria, which goes back to the owner. A criterion change is a new version in the registry. A model change is re-run against the evaluation set before it goes live.

One API for every referrer system: practice management systems, existing eReferral providers, urgent care systems and the hub's own web form. It is FHIR R4 on the HL7 New Zealand Base profiles, uses national identifiers, and asks for as little as possible up front. This is a starting point for the vendor working group, not a specification.

Base URL and namespaces are placeholders. Anything with an rh- prefix is proposed and does not exist yet. NHI, HPI and naming systems are the real NZ Base ones. Patient and practitioner identifiers in examples are test-style values.

Access

System-to-system calls use OAuth 2.0 client credentials with signed JWT assertions (the SMART Backend Services pattern) through Health NZ's API gateway. Every submission carries the referring practitioner's HPI person ID and the facility or organisation HPI ID, so the hub knows who is accountable without a separate user login. Conformant vendors receive production credentials only after passing the test suite.

The mandatory envelope

Versioning, conformance and governance

The API is a published national standard, so its governance is part of the design. Proposed: semantic versioning in the URL path; additive changes within a major version; breaking changes only in a new major version with at least 18 months of overlap; a public sandbox with synthetic patients; an automated conformance test suite; and a change board that includes vendors and clinical representatives, operating under HISO endorsement. Passing conformance becomes a condition in primary care software procurement and accreditation, which is the lever that turns a published standard into one that vendors actually implement.

When triage completes or changes, the hub emits one canonical event: a FHIR message Bundle. Every downstream consumer receives the same content. Systems that cannot yet read FHIR receive the same event rendered by an adapter as HL7 v2, following the HISO 10011 referral messages (REF and RRI, I12 to I15).

Mapped to the referral's path through the hub. The right-hand column is what each group holds, needs, or is likely to worry about; it is the agenda for engagement.

Autonomy in this proposal means the specialty's control over the hub, not the hub's independence from the specialty.

The division is simple to state. Specialties own clinical content and decide how much the hub may do on its own. The hub owns the mechanics: the standard, the bindings, the pipeline, the evidence. National governance owns a small number of floors that no specialty can lower. Within those floors, a specialty experiences the hub as its own: its forms, its criteria, its wording, its pace.

Decision rights

The autonomy ladder

Autonomy is set per exam and tier, not per specialty: US Abdomen P3, for example, rather than all of ultrasound. A specialty can have its clearest exams and tiers automated while contentious ones stay in assist mode indefinitely. Each step up needs evidence from the step below. The gate numbers are strawperson values for the clinical networks to set.

Floors no specialty can lower

• A detected red flag always goes to a clinician, marked urgent, and is never automatically declined.
• No referral is declined without the criterion it was assessed against and advice on what to do instead.
• The referrer can always ask for a human review of an automated decision.
• Automated decisions are sampled at or above the national floor rate.
• Every form field binds to the data standard.
• Outcomes are monitored by ethnicity, deprivation and rurality and reported to the specialty monthly.
• Records are kept for the statutory retention period, with the full evidence trail.

How this fits the CRR hub design

The CRR design already has most of the governance this needs: regional GP radiology liaison leads forming a national network, a change-control process for criteria, nationally consistent decline wording, and cross-auditing of triage decisions to catch drift. The hub gives that network a registry to own, a dashboard of agreement and outcomes, and a set of switches. It does not replace the network's judgement.

What the AI is asked to do

Read a referral against the whole criteria set for the service. Say which criteria it meets and quote the referral for each condition. Flag red flags, alternative pathways and not-funded items that fit, list what is missing and the questions that would settle it, and give a short reason. Answer "discuss with PCRL" when the referral doesn't support a decision. It does not set the priority or rate its own confidence; code does both from the checked result. It is not asked to diagnose or invent criteria, and it can suggest a better-fitting exam only from those the criteria list. That is a narrower task than most published evaluations of language models in medicine, which matters for what the evidence can tell us.

What the evidence says

The literature on AI referral triage is young and mostly retrospective, and a systematic review of AI applied to referral documentation is only now underway. The studies point in a consistent direction:

• On guideline-based imaging scenarios, current models are highly accurate; the best in one 2025 study scored 98.3% on American College of Radiology appropriateness scenarios and matched junior radiologists on more realistic cases.
• On GP referrals for lumbar spine MRI, the best model and prompt combination vetted with 0.86 accuracy in a prospective evaluation using synthetic referrals.
• On real brain CT referrals from Irish centres, a purpose-built prediction model outperformed general chatbots at judging justification, and the study looked specifically at how referral completeness changed the result.
• An institutional model that augmented musculoskeletal MRI requests from the clinical record cut the share of deficient requests from 12–23% to under 1% and improved protocol selection.
• Across 23 models triaging rheumatology referrals, accuracy was comparable to what has been reported for human triage systems, and a well-designed prompt narrowed the gap between expensive and cheap models.
• In ophthalmology referral classification, performance dropped for rare and ambiguous categories.
• In the CRR prototype, Claude Opus 5.5 agreed with the expected outcome on 39 of 46 synthetic referrals (Sonnet 5.5 38, Haiku 5.5 37), found the red flag criterion in all 8 red flag cases, and every quote it gave was found in the referral. An AI model wrote those referrals and their expected answers, and no clinician has reviewed them yet, so this shows the mechanics work, not how the check compares with PCRLs.

The read-across is straightforward. Models do well when the criteria are explicit and the facts are in the referral, and less well when referrals are incomplete or the category is unusual. The hub is designed around that: structured forms put the facts in the referral, the pre-submission check catches what is missing, code checks every citation and sets the priority, and uncertain cases go to a person. None of the published numbers substitutes for local validation, which is why every criterion starts in shadow mode.

There is precedent for autonomy at scale. Skin Analytics' DERM, which holds a Class III CE mark and a positive NICE early value assessment, discharges benign skin lesions from urgent referral pathways without a dermatologist. It has assessed more than 250,000 patients across 25 NHS trusts and is being extended to all 18 trusts in NHS England South East from autumn 2026.

The human baseline is not a fixed standard

Triage by people varies. Waitematā's decline rate moved from about 2% to 17% depending on who was triaging, audits found 10–50% of direct referrals didn't meet criteria, and experienced triagers saw their own prioritisation drift within a year. An automated assessment gives the same answer to the same referral every time, and when it is wrong, it is wrong in a way that can be measured and corrected.

Benefits

• Capacity. Clinician time moves from routine reading to the cases that need judgement, advice and education.
• Speed. Seconds instead of the current target of around two days, and no batching delays for booking teams.
• Coverage. Direct-to-private referrals can be triaged before the scan instead of audited after it. X-ray can come into scope.
• Consistency. The same referral gets the same answer in Kaitaia and Invercargill. Drift becomes visible.
• Feedback at the point of referral. The referrer learns what is missing while the patient is still in the room.
• Measurement. Every decision is structured data: demand by criterion, decline reasons, unmet need, and equity of access.

Designed for safety

• The model reads the criteria as written and cites them; code checks every citation and quote, sets the priority and applies the policy, so each failure has one place to look.
• Confidence worked out in code from the checked result rather than the model's own rating, with "discuss with PCRL" always available as an answer.
• Accepts and declines treated differently: accepts can be automated once every check passes, declines go to a person, and red flags override everything.
• Full provenance: criterion, model, prompt and policy versions on every decision.
• Random and targeted sampling of automated decisions, and seeded test cases for human reviewers.
• Monitoring for drift and for differences by ethnicity, deprivation and rurality, with automatic step-down.
• A model gateway: pinned model versions, revalidation before any change, more than one supplier, approved processing locations, and contracts that prohibit training on referral content.

Feasibility constraints

• Input quality. Scanned letters and attachments are harder. Early phases rely on structured forms; unstructured inputs are handled as evidence, never as the only basis for an automated decline.
• Data handling. NAIAEAG's stated caution on language models means hosting, processing location and supplier terms must be settled in Phase 0.
• Regulation. New Zealand has no dedicated regulatory regime for software as a medical device. The Medical Products Bill will regulate software with a therapeutic purpose; the Ministry has indicated introduction in early 2027 with a two-year transition, so late 2029 at the earliest. Whether referral triage falls inside it is not settled. The hub should be built to medical-device quality practice from the start so that classification, if it comes, is paperwork rather than a rebuild.
• Accountability. Clinical accountability stays with the specialty under its governance. The provenance record shows exactly what happened and why.
• Workforce. Triage roles change. Early, honest engagement with the people doing the work and their unions is part of the delivery plan, not an afterthought.

A scan of referral platforms and AI triage products that a buyer would look at, and what each teaches.

No product combines the three things this proposal needs: a New Zealand criteria registry under Health NZ clinical governance, an AI decision layer with autonomy set per exam and tier and full provenance, and a nationally owned FHIR standard that New Zealand vendors implement. The components exist and should be bought. The combination and its governance are the part only Health NZ can supply.

Buy the parts the market sells. Own the parts that decide who gets care.

Access criteria are public decisions

Who receives publicly funded imaging is decided by Health NZ under clinical governance. The logic that applies those decisions should be inspectable, versioned, and answerable to official information requests, the Ombudsman, the Health and Disability Commissioner and coroners. That is much harder when the logic sits inside a supplier's product.

The standard is the asset, and ownership is the lever

Ontario owns its eReferral specification and vendors compete to implement it. The Netherlands relies on a private platform that has changed owners twice. New Zealand's primary care software market is concentrated: Medtech supplies more than 70% of practices and has itself changed hands. A standard Health NZ owns, tied to procurement and accreditation, survives vendor changes. A standard a vendor owns is a dependency.

Accountability for automated decisions cannot be contracted out

NAIAEAG endorsement, acceptance checks, the ability to pause automation, monitoring and the safety case all need to sit with the organisation that is accountable for the outcome. If the Medical Products Bill brings triage software into scope, Health NZ will want to be the party that controls the evidence, not a customer waiting on a supplier's roadmap.

Demand data is planning infrastructure

The CRR programme is moving to a single waitlist for hospital and community imaging, with a dashboard on the Health Data Platform. A hub that structures every referral produces the demand signal that planning needs: volumes by criterion, decline reasons, unmet need and where it sits.

Equity obligations need control of the data

Monitoring automated decisions by ethnicity, deprivation and rurality, and acting on what it shows, depends on Health NZ holding the data and the monitoring. So does honouring commitments on Māori data sovereignty.

The economics favour owning the thin layer

Per-referral pricing scales badly at national volume. The hub is mostly integration, checks and orchestration, work Health NZ already does. The expensive-sounding part, the language model, is bought by the call and costs cents per referral.

The obvious objection

The Minister of Health has questioned Health NZ building tools that can be bought; the example was an in-house AI scribe. That objection is right for scribes, and this proposal applies it: the model, the FHIR platform, integration engines and form tooling are all bought. What is owned is the policy layer around them, which nobody sells for New Zealand.

An adaptable model. Every number is an input, the presets load starting points, and the outputs update as you type. Inputs are remembered in this browser only.

Benefits fall into three groups: clinician triage time released, administrative and rework time released, and low-value exams avoided on pathways that are not triaged today. Released time is valued two ways. The economic view counts all of it at what it costs. The cash view counts only the share you expect to turn into money, by avoiding planned growth in triage staffing or redeploying hours to waitlist work. Most released clinician time is capacity, not cash, and the business case should say so.

How the model works

For each stream, manual triage hours are referrals × minutes ÷ 60. After the hub, the hours still needed are the manual hours × (share not automated × assisted review time + share automated × QA sample rate). The difference is released time, valued at the triager cost. Administrative time is valued the same way. Low-value exams avoided are untriaged referrals × share not meeting criteria × share the hub catches × cost per exam, and are counted in both views because avoided per-exam payments are cash. Running costs include the model cost for every referral, including in shadow mode. Payback is the first year cumulative net turns positive. The break-even line solves for the no-touch share at which steady-state net benefit, economic view, is zero.

Where the defaults come from

• Volumes: community ultrasound and CT exams completed January to August 2024, by district, from Appendix 2 of the CRR hub design, annualised (about 164,000 ultrasound and 51,000 CT; breast and maternity ultrasound excluded) and rounded up to allow for declined referrals.
• Manual triage time: the hub design's planning rates of 10 ultrasound or 15 CT referrals an hour, observed at Waitematā.
• Untriaged referrals and inappropriate share: placeholders. The hub design reports 10–50% of direct referrals not meeting criteria in past audits; 15% is used.
• Outpatient stream: an illustrative placeholder only, to show how the case changes with scale. Replace it with Health NZ referral data.
• Costs, automation rates and hourly rates: strawperson values for challenge. They are not quotes or Health NZ figures.

Staged so that each phase produces evidence the next one needs, and so that nothing is automated before it has been measured. Timings are indicative.

1. Endorse the direction: a Health NZ-owned referral hub with AI triage, starting with community referred radiology.
2. Fund Phase 0, a four-month discovery, with named clinical, digital and commissioning leads.
3. Invite the CRR programme and the Radiology National Clinical Network to sponsor CRR as the first service.
4. Register the initiative with NAIAEAG now, so hosting and evidence requirements shape the design.
5. Ask procurement to identify the mechanism that makes conformance a condition for primary care software.
6. Agree that the referral API is taken through HISO as a national standard.

Open questions

• Which region hosts the shadow pilot?
• Where does the hub sit organisationally: Data and Digital, a clinical service, or the CRR programme in the first instance?
• What processing locations and model suppliers are acceptable?
• What error limits will clinical leadership accept for automated accepts, and should any specialty ever automate declines?
• How does the timeline align with the national radiology digital modernisation programme and its RIS?
• Which specialties follow CRR?