Skip to content
HealthcarePatient Pathway AICASE_13

Inheriting and fixing a broken patient triage AI from a vendor who had gone dark

An NHS-style mid-size acute trust (engagement Q2 2024) where the prior vendor had shipped a triage classifier and ceased trading, model still in production, generating incorrect triage scores for ~8% of patients, with zero documentation, no codebase access, and no model cards. We reverse-engineered the existing model from its outputs (treating outputs as ground truth until we understood why it failed), identified a systematic under-triage bias against elderly patients with atypical presentations, and shipped a stabilisation patch within two weeks. Then rebuilt the full triage model on the trust's 5-year admission records with clinical-governance involvement at every feature-selection and threshold-setting decision. Triage accuracy 97.3% on held-out data, reviewed and approved by the clinical governance board. Unnecessary A&E admissions reduced ~40%. The trust now owns the full model, training pipeline, and monitoring stack.

97.3%
Triage accuracy
−40%
Unnecessary A&E admissions
6wk
Time to stable production

The Challenge

The vendor had delivered a triage classification model and then ceased trading. The model was in production but generating incorrect triage scores for ~8% of patients, and there was no documentation, no codebase access, and no model cards. The hospital needed the system stabilised urgently, then rebuilt to a standard that could survive clinical governance review.

Our Approach

AI Readiness Audit was forensic, we reverse-engineered the existing model from its outputs, identified the failure modes (systematic under-triage for elderly patients with atypical presentations), and built a stabilisation patch within 2 weeks. We then rebuilt the full triage model using the hospital's 5-year admission records, with clinical governance involvement at every stage of feature selection and threshold setting.

Outcome

Triage accuracy: 97.3% on held-out clinical data, reviewed and approved by the hospital's clinical governance board. Unnecessary A&E admissions reduced 40%. Clinical governance documentation complete. The hospital now owns the full model, training pipeline, and monitoring stack.

What We Learned

01

Vendor lock-in on ML models is a governance risk, always require full model documentation and portability at contract stage.

02

Forensic diagnosis of a broken model requires treating its outputs as ground truth until you understand why it fails.

03

Clinical stakeholder involvement in feature selection is not optional, it's how you get clinical governance approval.

Stages Engaged
AI Readiness Audit
Discovery & Blueprint
Production Build
Total Duration
4 months total
Artifacts Delivered
PRD
Clinical Governance Documentation
Model Cards
WBS
IT Runbook
Monitoring Dashboard Spec
Start with a Feasibility Call

2 hours. No cost. We'll tell you honestly whether AI makes sense for your case.

Book a call