HEOR and market access guide
AI for HEOR evidence synthesis: from systematic review to HTA dossier
How AI fits the SLR, indirect comparison and HTA dossier workflow, what NICE, CDA-AMC, IQWiG, HAS and PBAC have said, and how to validate AI screening.
AI can now help at every step of HEOR evidence synthesis: search, screening, extraction, indirect comparison and dossier drafting. HTA bodies do not ban it. NICE and Canada's Drug Agency accept AI methods that are declared, justified, validated and overseen by a capable human, and IQWiG expects suitability shown against international standards such as RAISE. The submitting company stays accountable, so the tool must leave a reason for every decision.
The HEOR evidence workflow: SLR, ITC, dossier
Market access evidence follows a fixed chain. A systematic literature review (SLR) finds every relevant study. An indirect treatment comparison (ITC) or network meta-analysis estimates relative effects when no head-to-head trial exists. The results then go into a health technology assessment (HTA) dossier in each payer's format.
- Protocol. Define the question as population, intervention, comparator and outcome (PICO), with inclusion and exclusion criteria.
- Search. Build and document database queries.
- Screen. Title and abstract review, then full text, usually by two reviewers.
- Extract and appraise. Pull study data and assess risk of bias.
- Synthesize. Meta-analysis, ITC or network meta-analysis.
- Report. A PRISMA 2020 flow diagram, evidence tables and narrative.
- Adapt. Reformat and update the evidence for each HTA body.
The last step is growing. The EU Joint Clinical Assessment (JCA) adds a European layer. In July 2025 the G-BA changed its dossier templates so companies can reference documents already submitted at EU level.
Where AI helps at each step
NICE's position statement lists the uses it considers established and those it considers less so (NICE ECD11, paragraphs 1.17 to 1.20). Canada's Drug Agency repeats the same list (CDA-AMC position statement, paragraphs 1 to 5).
| Step | What AI does | How established, per NICE |
|---|---|---|
| Search | Generates search strategies; text analysis finds search terms | Supported use |
| Study classification | Tags records by study design, such as RCT classifiers | Supported use |
| Title, abstract and full-text screening | Ranks or decides on records against the criteria | Supported use |
| Data extraction | LLMs extract quantitative and qualitative data | "Less well established" |
| Synthesis | LLMs generate code for (network) meta-analysis | "Less well established" |
| Dossier | Drafts narrative and reformats evidence per market | Not covered in the systematic review section |
The practical gain is largest in screening, where thousands of abstracts meet a fixed set of criteria. Extraction and synthesis need closer human checking because they feed the numbers a committee examines.
What HTA bodies have said about AI in evidence synthesis
We read each body's own published documents. Here is what they say, as of October 2026.
NICE (England)
NICE published Use of AI in evidence generation: NICE position statement (ECD11) on 15 August 2024. Key points:
- AI methods "should only be used when there is demonstrable value from doing so" (1.1).
- Use should be "based on the principle of augmentation, not replacement, of human involvement," with "a capable and informed human in the loop" (1.7).
- Submitters "should conduct careful technical and external validation when AI methods are used, and present the results" (1.7).
- Submitters "should clearly declare its use, explain the choice of method and report how it was used, including human input." The submitter "remains accountable for the content" (1.8).
- Submitters should engage with NICE early to discuss their plans (1.3) and report the risks they found and how they addressed them (1.12).
CDA-AMC (Canada)
Canada's Drug Agency (formerly CADTH) published its Position Statement on the Use of Artificial Intelligence in the Generation and Reporting of Evidence in April 2025. It is based on the NICE statement and aligns with all of NICE's positions. It repeats the augmentation principle, the need for technical and external validation, and the duty to declare AI use (paragraphs 25 and 26). It adds Canadian legal and ethical references and asks submitters to engage early through CDA-AMC Scientific Advice.
IQWiG and G-BA (Germany)
The G-BA decides on benefit; IQWiG assesses dossiers under its General Methods, Version 8.0 (19 December 2025). Section 9.1.1 says that "the use of artificial intelligence (AI) for study selection cannot be ruled out in the future." It adds that suitability "should be based on international standards, such as those set out in the RAISE guidance." The same section already allows validated machine-learning classifiers, such as RCT classifiers, in search strategies. Study selection is still done by two people independently. We did not find a separate G-BA statement on AI in dossiers.
HAS (France)
HAS has not published a position on AI in industry submissions that we could find. It has reported on its own internal experiments (April 2025). Sixteen bibliographic search tools it tested brought no significant added value. It tested ASReview, an active-learning screening tool, retrospectively on five finished HAS reviews with promising results, and is now testing it prospectively.
PBAC (Australia)
The current PBAC Guidelines (Version 5.0) do not mention artificial intelligence or machine learning. We found no separate PBAC statement on AI-assisted reviews. Submitters should follow the general literature review requirements and declare any AI use.
The methods community
In November 2025, Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence published a joint position statement based on the RAISE recommendations. It calls for human oversight, transparent reporting and justification. Authors must ensure AI "will not compromise the methodological rigour or integrity" of the review. IQWiG points to RAISE directly. PRISMA 2020 items 8 and 9 already ask authors to report "details of automation tools used" in selection and data collection.
Will HTA bodies accept a review screened with AI?
Yes, on conditions. No HTA body we checked forbids AI-assisted screening. NICE and CDA-AMC set out the conditions explicitly. IQWiG leaves the door open and points to RAISE.
Across these documents, the conditions are the same:
- Declare it. State which tool was used, at which step and how.
- Keep a human in the loop. A qualified reviewer checks and owns the decisions.
- Validate it. Show technical and external validation results, not vendor claims.
- Make it traceable. Each inclusion and exclusion should be explainable and reproducible.
- Talk to the agency early. NICE and CDA-AMC both invite early engagement.
The risk is not that AI is used. The risk is a review whose decisions cannot be explained when an evidence review group asks.
How to validate AI screening before you rely on it
NICE asks for "careful technical and external validation" but does not prescribe a protocol. A practical approach that a reviewer can follow:
- Pre-specify the role of AI in the protocol. Say whether AI ranks records, acts as a second reviewer, or makes first-pass decisions that humans check.
- Use a reference set. Run the tool on a completed review in the same disease area, with known included studies, before the live review.
- Measure recall first. Missing an eligible study is the costly error in an HTA review. Report recall (sensitivity) and the workload saved, with confidence intervals.
- Check a human sample. Have a human screen a random sample of AI exclusions in the live review, and report agreement.
- Audit the reasons. Read the recorded reason for a sample of decisions. A correct decision with a wrong reason is still a problem.
- Freeze and log versions. Record the model version, criteria text and settings, so the screening can be reproduced.
- Report it. Describe the tool, the human role and the validation results under PRISMA 2020 items 8 and 9, and list risks and mitigations as NICE paragraph 1.12 asks.
Ask every vendor for peer-reviewed or independently checkable accuracy data, and run your own check on a review you know well.
How TrialMind handles HEOR evidence synthesis
TrialMind for health economics and market access runs the review as one traceable workflow:
- Protocol: the question is defined by PICO before the search begins.
- Search: visible PubMed queries and filters, with results reviewed before screening.
- Screening: abstracts are scored against predefined criteria, with reviewer oversight. Every inclusion and exclusion has a recorded reason.
- Extraction: tables linked to source papers, with custom fields and human review.
- Synthesis: report sections and evidence tables with linked citations, plus risk-of-bias assessment.
- Indirect comparisons: where no head-to-head trial exists, for example a Bucher comparison of dapagliflozin and empagliflozin in heart failure using the published DAPA-HF and EMPEROR-Reduced trials.
- Dossier formats: NICE, G-BA, CADTH, HAS and PBAC, from one evidence base updated per market.
The methods are published. In A foundation model for human-AI collaboration in medical literature mining (LEADS, Nature Communications 2025), 16 clinicians and researchers from 14 institutions used the model in their workflow. In study selection, experts with LEADS reached 0.81 recall against 0.78 without it and saved 20.8% of their time. In data extraction, accuracy was 0.85 against 0.80, with 26.9% time savings. An earlier paper, Accelerating clinical evidence synthesis with large language models (npj Digital Medicine 2025), reports a pilot in which human-AI collaboration improved screening recall by 71.4% and cut screening time by 44.2%.
TrialMind deploys in the cloud, a private cloud or your own environment. Keiji AI is SOC 2 Type I and Type II audited, ISO 27001 certified and HIPAA compliant.
The tool landscape in brief
Most HEOR teams compare TrialMind with established review platforms and newer AI tools. See the full table on DistillerSR, Covidence and AI-native alternatives. In short:
- DistillerSR offers generative AI screening, re-ranking and custom classifiers, with Part 11 and Annex 11 controls.
- Covidence keeps dual human screening and uses machine learning for relevance sorting and RCT classification.
- Rayyan adds an AI Reviewer that screens against your criteria with reasoning notes.
- Nested Knowledge pairs AI screening with built-in network meta-analysis.
- Laser AI focuses on ranked screening and source-linked extraction.
- Elicit automates screening and extraction with per-criterion scores and quotes.
Buyer checklist
Ask each vendor these questions before a pilot:
- Does the tool record a reason for every inclusion and exclusion, and can I export it?
- Where is screening accuracy published, and can I check it on my own completed review?
- What is the human role at each step, and can I configure it to match my protocol?
- Which databases can I search or import from, and is the search strategy exported in full?
- Does it produce a PRISMA 2020 flow diagram and the details PRISMA items 8 and 9 require?
- Can it run an indirect treatment comparison or network meta-analysis, and document the method?
- Which HTA dossier formats does it support, and how is the evidence base reused for the next market?
- Where does it run, which certifications does it hold, and is my data used for model training?
- Are model versions and settings logged, so the review can be reproduced a year later?
Frequently asked questions
- Do NICE and other HTA agencies accept AI-assisted systematic literature reviews?
- Yes, with conditions. NICE and Canada's Drug Agency accept AI methods that are declared, justified, validated and overseen by a capable human, and the submitter stays accountable. IQWiG says AI for study selection cannot be ruled out and points to the RAISE guidance.
- Is AI screening acceptable for HTA submissions?
- It is acceptable when it augments human reviewers rather than replacing them, and when you report how it was used and validated. Keep a recorded reason for each decision so the evidence review group can check it.
- Can AI help prepare an HTA dossier for NICE or G-BA?
- Yes. AI can run the review, build indirect comparisons and draft evidence sections, but the company remains accountable for the content. TrialMind supports NICE, G-BA, CADTH, HAS and PBAC dossier formats from one evidence base.
- Which step of an SLR benefits most from AI?
- Title and abstract screening, where many records meet fixed criteria. NICE describes extraction and meta-analysis code generation by LLMs as less well established, so they need closer checking.
- How do I report AI use in a systematic review?
- Name the tool and version, the step and the human role, and give validation results. PRISMA 2020 items 8 and 9 ask for details of automation tools in study selection and data collection.
See TrialMind on your own question
A 30-minute walkthrough on a study like yours, with sources and code you can check.
Sources · checked 2026-10-21
Product details come from each company’s public pages. Spotted something out of date? Tell us and we will correct it.
- NICE: Use of AI in evidence generation, position statement (ECD11), 2024
- NICE ECD11 (PDF), systematic review and evidence synthesis paragraphs 1.17 to 1.21
- CDA-AMC: Position Statement on the Use of AI in the Generation and Reporting of Evidence, April 2025
- IQWiG: General Methods, Version 8.0, 19 December 2025
- G-BA press release, 17 July 2025: procedures after EU assessment
- HAS: La HAS évalue le potentiel de l'IA pour assister le processus de revue de littérature, April 2025
- PBAC Guidelines, Version 5.0
- Cochrane: Setting the standards for responsible AI use in evidence synthesis (RAISE), November 2025
- PRISMA 2020 checklist
- LEADS: A foundation model for human-AI collaboration in medical literature mining, Nature Communications 2025
- Accelerating clinical evidence synthesis with large language models, npj Digital Medicine 2025
- DistillerSR AI
- Covidence: how its machine learning works
- Rayyan ResearchPilot
- Nested Knowledge: network meta-analysis
- Laser AI product
- Elicit for systematic reviews