From ambulance bay to OPCS-4—and why British coding AI needs a British test track. Starring one hi-vis pigeon, a surgical rubber duck and a deeply concerned finance department.
A patient episode begins when someone arrives for care. Doctors and nurses treat the patient and document the encounter in notes, scans and discharge summaries. Across NHS trusts, specialist clinical-coding teams—sometimes more than thirty people in a large trust—read those records and translate confirmed diagnoses and completed procedures into ICD-10 and OPCS-4 codes.
It is skilled, labour-intensive work that takes years to master. The resulting codes support reimbursement, epidemiology and service planning. Automation could help, but patient records are sensitive and difficult for AI companies to obtain. MIMIC-IV gives researchers reusable US hospital data, but it reflects an American institution and coding context rather than NHS workflows.
EPICODE is the MVP of a UK-specific synthetic episode dataset and benchmark for training, comparing and evaluating clinical-coding AI products before governed real-world validation.
US datasets made clinical AI evaluation possible—but they encode US care and billing. NHS teams need UK-native episodes mapped to ICD-10 and OPCS-4, without waiting for patient-level data access.
Clinical impression: acute appendicitis.
The UK does not lack health data. It lacks a ready-to-use, public, coherent episode benchmark for testing NHS clinical coding systems against both diagnosis and procedure standards.
MIMIC-IV provides de-identified hospital records, notes and ICD billing fields at research scale. It is an extraordinary resource—and it comes from one Boston health system, not the NHS.
Useful for model research ≠ evidence of NHS coding performance.HES contains nationally coded activity, but patient-level access is governed. Public Artificial HES helps test pipelines, yet does not preserve relationships between fields and cannot support analysis.
No public MIMIC-equivalent for coherent NHS ICD-10 + OPCS-4 episode evaluation.EPICODE supplies the missing test fixture: deterministic NHS-style episodes, expected ICD-10 and OPCS-4 classifications, and passage-level evidence contracts. Synthetic first; governed real-world validation later.
Inspect the test data ↓Choose a pathway and seed. The same inputs always produce the same validated episode.
Applying scenario constraints and evidence contracts…
America offers reusable clinical datasets at research scale. Britain offers rich national activity data behind governance and public artificial extracts that are useful for plumbing—not coherent coding evaluation. EPICODE is designed for that specific space between them.
364k+ de-identified patients; hospital, ICU and ICD billing data.
Public synthetic Medicare claims in realistic CMS file formats.
Commercial effect: teams can prototype and compare sooner—although US coding and care pathways do not transfer neatly to the NHS. The Atlantic is not a validation strategy.
Three public, large-scale files for schemas, teaching and pipelines—not coherent patient journeys.
England-wide admitted care, outpatient and emergency activity.
Longitudinal UK primary-care records with linked health data.
Commercial effect: superb real-world evidence exists, and public artificial HES helps test plumbing. But coherent product evaluation can still inherit access lead-time, governance cost and restricted portability.
A curated UK synthetic benchmark lets product, clinical and procurement teams agree acceptance criteria early. Use it for rapid regression tests, evidence demos and vendor comparisons; reserve governed real data for later-stage external validation.
Landscape snapshot, 18 July 2026. “Fewer ready-made benchmarks” is our inference from the access models shown—not a claim that UK health data is absent. Synthetic results establish engineering confidence, not clinical validity.
Purpose-built synthetic records let teams develop evaluation workflows before governed data access.
Every classification points back to explicit evidence, making expected outputs inspectable.
Seeded generation turns a polished demo into a reproducible engineering asset.