Specialist foot and ankle judgement, at the point of first contact.

I am building a clinical decision-support tool that helps non-specialist clinicians make safer decisions about foot and ankle injuries - the presentations most commonly missed, and most expensively missed.

It is the first vertical of a platform designed to extend across musculoskeletal care and beyond.

Get in touch See how it works

In development. Not yet UKCA or CE marked. Not currently available for clinical use

Diagram of the foot and ankle

50%

General-purpose AI gets foot and ankle triage wrong about half the time.

Foot and ankle injuries are a small proportion of musculoskeletal presentations and a disproportionate share of missed diagnoses.


A Lisfranc injury seen on a Friday and discharged as a sprain becomes a midfoot fusion two years later.


The clinicians seeing these injuries first — physiotherapists, urgent care doctors, GPs, emergency nurse practitioners — are not the people who see them often. They are working without sub-specialist input at exactly the moment that input matters most.


Published evidence in 2026 has repeatedly found that general-purpose large language models perform poorly and unpredictably on exactly this task. That is not an argument against using AI here. It is an argument against using a general one.

Foot and ankle is the
first vertical, not the ceiling.

The hard problem in clinical AI is not language. It is constraint: building something narrow enough to be safe and validated enough to be trusted.


Foot and ankle is a deliberate first domain. It is narrow enough that genuine subspecialist depth is achievable, and clinically significant enough that getting it right matters. It is where the evidence of failure is clearest and the cost of a missed injury is easiest to demonstrate.


But nothing underneath it is foot-specific. A curated subspecialist case base, a deterministic safety layer, structured intake that keeps identifiable data out of the system, and an auditable output with provenance - the architecture scales; the subspecialist content and validation are what make it trustworthy.

The same pattern applies anywhere a first-contact clinician is making decisions above their level of exposure: the rest of musculoskeletal care first, and in principle any subspecialty with the same structure — high-stakes decisions, made early, by people who do not see these injuries often.

Each vertical added strengthens the ones before it: the same regulatory groundwork, the same safety architecture, the same governance, the same deployment relationships.

Watch it triage a real case.

Ninety seconds, start to finish — from structured intake through the deterministic safety layer to a draft referral letter.

A narrow tool, built deliberately narrow.

A subspecialist knowledge base

Over 110 anonymised foot and ankle cases, curated and written by a practising consultant foot and ankle surgeon. Not scraped, not synthetic, not general orthopaedic content.

A deterministic safety layer

Hard-coded rules for the presentations that must never be missed — Lisfranc, Maisonneuve, syndesmotic injury, Achilles rupture, DVT. Where these rules fire, they override the model. The recommendation is fixed, not generated.

Structured intake only

No free-text patient narrative is ever transmitted. Clinicians select findings on an anatomical interface. The system is designed so that patient-identifiable information does not reach the model at any point.

An artefact at the end

Every assessment produces a structured recommendation and a draft referral letter — something the clinician can act on, not a conversation they have to summarise themselves.

The difference between a chatbot and a clinical tool is what it refuses to do.

A general assistant will always produce an answer. That is precisely the failure mode in clinical triage, where the most dangerous output is a confident one built on incomplete information.


My tool will not triage a presentation until the findings that determine urgency have been positively answered — weight-bearing status, neurovascular status, whether there is an open wound. Where a red-flag pattern is present, a deterministic rule fires and fixes the output regardless of what the model infers.

Every recommendation carries its provenance: whether it came from a hard-coded clinical rule or from model reasoning over the case base. That distinction is visible to the clinician using it, and auditable afterwards.

Built to be regulated, from the start.

I have taken the view throughout that a clinical AI tool which cannot survive regulatory scrutiny is not worth building.

Independent validation is underway. An external study of TwentySix Triage with the University of Pittsburgh's HexAI Lab is currently under institutional review board review - separate from and in addition to the case base I curated myself.

This product is in development. It is not yet UKCA or CE marked and is not currently available for clinical use.

Built for the first ten minutes,
not the clinic letter.

Urgent and injury care clinics

Where a patient with a twisted ankle is seen once, by a clinician who may not see another midfoot injury for months.

First-contact practitioners

Physiotherapists and advanced practitioners operating as the first point of assessment, increasingly without medical backup.

Insurer-funded MSK pathways

Where the cost of a missed injury and the cost of an unnecessary specialist referral are borne by the same organisation.

Built by the specialist,
not for one.

I am Maurice O'Flaherty, a consultant orthopaedic foot and ankle surgeon practising in Belfast and Dublin.

I built this because I spent years receiving referrals that arrived too late, and referrals that should never have been made at all. The knowledge base is my own casework. The clinical rules are the ones I apply myself. The architecture I taught myself to build, evenings and weekends, alongside a full surgical practice.

The result is a tool with an unusual property for clinical AI: the person who defined what correct means is the same person who built it, and is accountable for it — which is exactly why an independent validation study, run by a research group with no stake in the outcome, matters as much to me as the product itself.

Maurice

If you are building musculoskeletal pathways, underwriting them, or investing in them, I would like to talk.

Get in touch

Maurice O'Flaherty
TwentySix Triage Limited
+44 (0)7736 324073