AI-assisted colonoscopy triage tool

Problem

In Aotearoa New Zealand, demand for colonoscopy outstrips supply. Referrals arrive from family doctors written in very different styles (some thorough, some terse), and the urgency framing inside each letter is the only signal a specialist has to triage with. The cost of getting it wrong is asymmetric. A high-acuity patient can sit unrecognised in a pile of waiting referrals while the inbox is still being worked through. There was no structured layer between the family doctor’s letter and the specialist’s inbox, and no scalable way to embed the national guidelines into that handoff.

What I built

The triage tool reads the referral letter, follows the national colonoscopy triage guidelines, and returns a triage band (urgent, routine, or not accepted) alongside the specific guideline rules that justified the recommendation.

It runs as a private web application for the specialist service. Every recommendation is auditable: the clinician sees which guideline criteria fired, which were ruled out, and what supporting test (where applicable) was recommended. The clinical logic is treated as a regulated medical artefact. Every update is change-controlled, version-tracked, and gated by a regression suite before it reaches a user.

Evidence

• A prospective clinical-validation pass against an experienced gastroenterologist’s independent triage on a fifty-referral set returned 75% full agreement on both triage category and supporting-test recommendation, and 82% agreement on the triage category alone. Both figures sit above the working target for this stage of development.
• Ten previously-disputed cases were retrospectively agreed by the clinician once the underlying guideline citations were laid out side-by-side.
• National endorsements: the Endoscopy Guidance Group of New Zealand, Health New Zealand Digital Services as the technology integration partner, and the National AI Algorithm Expert Advisory Group for algorithmic governance.
• Full deployment approvals are in place across health and disability ethics, the local district health authority, an independent privacy review, and an intellectual property review.

A meaningful share of cases (around one in four) still needs a judgement call from the specialist. The triage tool is structured decision support; it is not a substitute for clinical review.

Impact

The next phase of validation is being readied for clinical recruitment, with an equity-as-active-intervention design: half of the recruited cases are Māori or Pacific, in direct response to known disparities in endoscopy access. The same underlying approach is being explored for cardiology referral triage at a second district, which suggests it generalises beyond one specialty.

Five concurrent funding applications support the next phase, including a national career-advancement fellowship that frames the triage tool as the platform on which a broader programme can be built.