Skip to content

mczuzu/verum

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Verum — HTA Evidence Intelligence Platform

Concept architecture · March 2026 · Portfolio artifact

Prototype: https://mczuzu.github.io/verum/ Interactive demo · Pembrolizumab NSCLC · Real HTA data · Full four-step workflow


What It Is

Verum mines published HTA (Health Technology Assessment) decision documents across EU agencies — Spanish IPTs (Informes de Posicionamiento Terapéutico), NICE (National Institute for Health and Care Excellence) STAs (Single Technology Appraisals), G-BA (Gemeinsamer Bundesausschuss) dossiers, HAS decisions — extracts the implicit PICO (Population, Intervention, Comparator, Outcome) specifications each agency actually accepted, and maps them cross-country to show where requirements converge and where they diverge.

The result: before you lock a Phase III protocol, you know which comparator satisfies all target markets, which biomarker thresholds need dual assays, and which endpoints must be co-powered.

Status: Concept validated. Data model complete. UI prototype live. No paying customers yet.


The Core Insight

Every published HTA decision contains an implicit study specification: the population, comparator, endpoints, and study design that the agency actually accepted. That specification is machine-extractable from narrative PDFs using structured LLM (Large Language Model) extraction.

An IPT is not just a decision document — it is a reverse-engineered study protocol. Mine enough of them and you have a rules engine. The corpus is entirely public. Nobody has structured it into a queryable cross-country database.


The Problem It Solves

550+ AEMPS IPTs published. 2,500+ EU HTA decisions publicly available. 0 structured as a queryable database.

Pharma companies run the trials. They have the data. The problem is that AEMPS, NICE, G-BA, and HAS each ask a slightly different question about that same evidence — and each question changes depending on what they've accepted for similar drugs before. Nobody has structured that institutional memory.

HEOR (Health Economics and Outcomes Research) teams face three recurring, avoidable failures:

01 — Wrong comparator, wrong country G-BA uses platinum-pemetrexed as SoC (Standard of Care) for non-squamous NSCLC (Non-Small Cell Lung Cancer). AEMPS and NICE accept platinum-based chemotherapy broadly. A trial designed against one SoC fails AMNOG (Arzneimittelmarktneuordnungsgesetz) assessment regardless of OS (Overall Survival) results.

02 — Biomarker fragmentation AEMPS accepted PD-L1 CPS ≥ 10. NICE and G-BA require TPS ≥ 50%. Different assays, different thresholds, different eligible populations. Without both pre-specified in Phase III, you cannot satisfy all three markets.

03 — 100 days from JCA scope to dossier deadline Since January 2025, the EU HTAR (Health Technology Assessment Regulation) mandates JCAs (Joint Clinical Assessments) for all new oncology products. The scope document defining required evidence arrives only 100 days before the dossier deadline. Preparation must start years earlier, against precedents that haven't been published yet.

Today, answering these questions is a manual consulting engagement costing €50,000–200,000 over several weeks. Verum answers it with every inference traceable to a specific published HTA decision.


How It Works — Four Steps

01 — Mine decisions Every published HTA decision — Spanish IPTs, NICE STAs, G-BA dossiers, HAS decisions — is ingested and parsed. LLM extraction pulls the implicit PICO from each document, always linked to the source paragraph.

02 — Structure evidence Extractions are validated by a human reviewer before being committed to the database. Every EvidenceUnit is versioned, auditable, and traceable — not a summarisation, a structured data object with a complete audit trail.

03 — Map divergences Cross-country PICO matrix showing where agencies align and where they diverge — population, biomarker, comparator, endpoint, study design — including JCA consolidated scope as European reference layer.

04 — Recommend action Strategic recommendations calibrated to your stage: Phase II→III protocol design, pre-launch market sequencing, dossier preparation, new indication planning, or competitive intelligence. The analysis tells you what to do next.


Who It's For — Five Decision Moments

Moment Users Value
Phase II → III HEOR · Clinical Development Lock the right protocol before the trial runs
Pre-launch Market Access · HEOR Sequence markets by evidence gaps, not assumptions
Dossier preparation Regulatory Affairs · HEOR Write what each agency expects, not a generic package
New indication Clinical Development Reuse existing evidence before commissioning new studies
Market intelligence Market Access · Strategy Track agency PICO signals as the corpus updates

The EvidenceUnit — Core Data Object

The minimum reusable evidence object. One EvidenceUnit can populate multiple country dossiers. Stored as JSONB in PostgreSQL/Supabase, indexed by indication composite key: {drug_id, condition_code, subgroup_filter, line_of_therapy}.

Schema aligned with GRADE evidence tables and PRISMA reporting standards.

Field Description Example
study_reference Source ID, database, study type PubMed 38291045, RCT
population_json PICO P: inclusion criteria, sample size NSCLC stage III/IV, n=847
intervention_json PICO I: drug, dose, duration Pembrolizumab 200mg Q3W
comparator_json PICO C: SoC, active comparator Platinum-based chemo
outcome_json PICO O: endpoint, value, CI OS HR 0.73 [0.58–0.92]
evidence_quality GRADE level, risk of bias — human-assigned High, low risk of bias
hta_context_json Country, agency, applicable guidelines ES, AEMPS, IPT criteria
version Append-only audit trail — never overwrite v1.2, 2026-01-15
extraction_meta LLM model, prompt version, confidence per field claude-3, prompt-v3, OS: 0.91

Human-in-the-Loop Design

Every pipeline phase produces a draft requiring human approval before the pipeline advances. The audit trail captures who approved, what version, when, and with what role.

This is not a constraint — it is the trust model. No AI tool in HTA has been accepted at submission level without full human accountability.

Phase Human task Estimated effort
P1 — Ingestion HEOR Scientist reviews retrieved study set; approves or refines query ~1–2 hrs / indication
P2 — Structuring Field-by-field EvidenceUnit review; GRADE assignment validated ~3–5 min / unit
P3 — Localisation Section mapping review; evidence gaps resolved; comparator confirmed ~2–4 hrs / country
P4 — Output Full dossier draft reviewed; citations verified; Medical Director sign-off ~4–8 hrs / dossier

Regulatory Context

The EU HTAR entered into force January 2025. All new oncology products are subject to JCAs. The JCA implementing act requires SLR (Systematic Literature Review) searches no older than 3 months, with the final scope issued only 100 days before the submission deadline.

This compresses timelines brutally. Reusable, versioned EvidenceUnits are the architectural answer to a blank-page rebuild under time pressure.


The Corpus

All sources are publicly available. The gap is not access — it is structure.

Source Volume Notes
AEMPS IPTs (Spain) 550+ total · 125/year · 60 oncology in 2024 PDF parsing + OCR for older docs
NICE STAs (UK) Part of 2,500+ EU decisions Structured URLs, downloadable
G-BA dossiers (Germany) Part of 2,500+ EU decisions Public PDFs
HAS decisions (France) Part of 2,500+ EU decisions Public PDFs
PubMed / ClinicalTrials.gov Public REST APIs, rate-limited
Embase Licensed Client-provided or CRO partnership

Tech Stack (Designed)

Layer Technology
Database PostgreSQL / Supabase — JSONB, append-only EvidenceUnit records
Extraction Claude / GPT-4 — structured PICO prompts, confidence scoring per field
Document output python-docx, pandoc
Frontend React + TypeScript + Tailwind CSS
Hosting GitHub Pages (prototype)

What This Is Not

  • Not an AI that writes HTA dossiers — output is structured evidence intelligence, not regulatory narrative
  • Not a literature search engine — does not replace PubMed or Embase queries
  • Not a substitute for HEOR expert judgement — informs decisions, does not make them
  • Not a built product — validated concept with complete architecture and documented gap

Related

Evidence Mapper — live AI-assisted pipeline over 63,000+ completed ClinicalTrials.gov trials. Built with the same HITL (Human-in-the-Loop) and structured data extraction principles that underpin Verum's architecture. evidence-mapper.com


María Castro · March 2026


Related

Evidence Mapper — live AI-assisted pipeline over 63,000+ ClinicalTrials.gov completed trials. Built with the same HITL and structured data extraction principles that underpin Verum's architecture.


María Castro · March 2026

About

Designed a four-layer LLM extraction pipeline to mine published EU HTA decisions (Spanish IPTs, NICE STAs, G-BA dossiers) and convert them into structured, versioned EvidenceUnit objects.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages