Every quarter I pick up to five new applications of AI in health economics, outcomes research, real-world evidence and HTA. Each one is chosen from primary sources (papers, conference abstracts, HTA-body documents), summarised in plain words with what it shows and where it falls short, and linked straight to the original.
Autumn 2026 edition Published 29 Oct 2026
The first edition has three picks, one each from HTA, modelling and real-world evidence, all from ISPOR Europe 2026 abstracts released this autumn ahead of the November meeting in Vienna. Later editions list up to five.
HTA & JCA01
AI predicted the first JCA scope without seeing it
Blinded test against a real JCA
The team ran an AI-enabled PICO scoping workflow for tovorafenib in paediatric low-grade glioma while the published Joint Clinical Assessment report was kept out of its reach. They then scored the prediction against that report, which defined 3 populations and 8 PICOs, and report high concordance with the official scope.
Why it matters
Most published PICO simulators have been graded against simulation exercises or the predictor's own reasoning. This is among the first tests against a real, published JCA scope.
Watch out
It is a single case. Read the full abstract for how concordance was scored before generalising.
An AI agent that runs TECH-VER checks on health economic models
Benchmarked against expert reviewers
The agent pairs a retrieval-augmented language model with deterministic parsers for Excel workbooks and R or Python code. It maps each model component to TECH-VER and AdViSHE checks and returns a graded error log. On 25 models seeded with 120 known errors, it caught 88% of them with 96% specificity against three experienced modellers (Cohen's κ 0.82). Verification time fell from about 6.2 hours to 9 minutes.
Why it matters
Technical verification is the quality step most often squeezed under deadline. A checklist-anchored agent makes it routine and documented rather than optional.
Watch out
The errors were planted, not found in real projects, and the agent missed about one in eight. It did worse on script-based models than on Excel (84% vs 91%). Treat it as a first pass, not a sign-off.
LLM chart abstraction in a real RWE pilot, with honest numbers
Pilot against adjudicated study data
The pilot used de-identified notes from 24 patients with a haematological malignancy, with several hundred to more than 1,000 notes each. A structured prompt extracted treatment response, ECOG, regimens and dates, transplant status and discontinuation reasons, and attached source snippets to each answer. First-pass outputs matched the adjudicated study values in 70.8% of 96 non-date fields, dates fell within 7 days in 86.3%, and treatment response agreed in 66.7%.
Why it matters
Most published LLM-abstraction results come from vendors' tuned pipelines. This shows what an enterprise LLM with a well-built prompt delivers, and why source-referenced output matters for traceability in HTA submissions.
Watch out
The sample was small, and agreement was lowest for co-mutations and transplant status. Human adjudication remained essential.
This is the first edition. Past picks will collect here each quarter.
How picks are chosen
New and primary
Published or presented in the past quarter, and taken from the original source: a journal article, conference abstract, preprint, or a document from a regulator or HTA body. Marketing pages count only when they report data.
Validated beats impressive
Work tested against human experts or a real-world benchmark ranks above demonstrations. Every pick states its evidence level and its main limitation next to its headline result.
Spread across the field
No more than two picks per workstream in an edition, and nothing repeats from an earlier one. Summaries are written fresh; read the source before relying on any figure.
DisclaimerPersonal project, not sponsored by or affiliated with any employer (current or former) or any university I attended; views are my own. Picks are selected and summarised with AI assistance and checked against each source, but errors may remain, so always rely on the linked original. Inclusion is not an endorsement of any product, company or method. Full disclaimer