How AI and Living SLRs Can Power Dynamic JCA Evidence Mapping

Living Literature Review and JCA: An Overview

A living systematic literature review (SLR) is a critical asset for health technology developers facing the new operational reality under the European Union Health Technology Assessment Regulation. Joint Clinical Assessments (JCAs) demand comprehensive, transparent, and up-to-date clinical evidence packages that address multiple Population, Intervention, Comparator, and Outcome (PICO) questions. Traditional static SLRs struggle to keep pace with rapidly evolving evidence and the tight timelines of JCA processes.

Living systematic literature reviews that continuously incorporate new studies, powered by AI platforms for Life Sciences, create a pathway toward dynamic evidence mapping that keeps insights current and actionable. This approach treats evidence as living infrastructure rather than a one-time project deliverable. It supports both JCA dossier preparation and subsequent national health technology assessment (HTA) needs.

Why Static SLRs Break Under JCA Conditions

Search Currency Becomes a Compliance Item

The dossier guidance sets a hard recency rule. The cut-off date for searches can be no more than three months before dossier submission. It also asks developers to report whether and when new data relevant to the assessment scope might become available, and to list new studies that could emerge during the assessment period.

Three months is tight by HTA standards. Methodologists have noted that NICE works to a six-month currency window, so the JCA halves the tolerance while multiplying the number of questions. Read the requirements together and the implication is clear: a one-time review cannot satisfy a forward-looking reporting duty. You are being asked not only what the evidence says today, but what is likely to land next quarter.

The Dossier Is Versioned by Design

The same guidance carries a revision history that anticipates repeat submissions: updated dossiers where a JCA report specifies the need for an update, updates filed on the developer’s own initiative when additional evidence becomes available, and updates following re-initiation of an assessment, each mapped to specific articles of Implementing Regulation (EU) 2024/1381.

In other words, the regulation treats the evidence base as something that reopens. A frozen SLR forces you to rebuild from scratch each time, usually with a different analyst and a lost audit trail.

Scope Volatility Outpaces Fixed Protocols

Scoping runs before the final therapeutic indication is settled, so the scope your dossier must answer can shift underneath you. That is why many developers simulate PICO scoping in advance. A static review keyed to one predicted scope is brittle; an evidence base indexed by PICO can be re-cut when the real scope arrives.

What Dynamic JCA Evidence Mapping Means

Two ideas combine here. A living systematic literature review is, in the original formulation by Elliott and colleagues, a systematic review that is continually updated as relevant new evidence appears. It is an approach to updating rather than a new methodology: standard systematic review methods, plus an explicit commitment to a predetermined search and update frequency.

Dynamic evidence mapping is what you build on top of it. Instead of one narrative synthesis, you maintain a structured matrix in which rows are PICOs (confirmed or predicted), columns are comparators and requested outcomes, and each cell records what evidence exists, whether it is direct or indirect, whether a connected network is available, and where the gap sits. Studies are indexed at the claim level, so a single trial can serve several PICOs without being re-extracted. The shift is from producing a document to maintaining an asset.

Architecture for AI-Assisted Living Evidence Mapping

AI-Assisted Living Evidence Mapping

Layer 1: Continuous, Reproducible Retrieval

Validated strategies run on a schedule across the required sources, with deduplication against the existing library and version stamping for every strategy change. Because reproducibility is the point, report searches using the PRISMA-S checklist, and keep the RIS exports

Layer 2: Recall-Calibrated Machine Triage

The Cochrane RCT Classifier was calibrated for 99% recall and then correctly retrieved 43,783 of 44,007 randomized trials included in Cochrane reviews, missing 0.5%, with older records more likely to be missed. In early pilots of the associated Screen4Me workflow, manual screening workload fell by between 40% and 70% depending on trial prevalence in the search results. A study-specific classifier built for the Cochrane COVID-19 Study Register reached 0.99 recall with a 24.1% net reduction in screening workload.

The lesson is not that automation removes screening. Classifiers can be calibrated to a recall target you are willing to defend, and the workload saving follows from that target rather than the reverse.

Layer 3: Structured Extraction With Targeted Verification

Extraction is where performance becomes uneven, and where oversight has to be specific rather than general. A systematic review of large language model performance in data extraction for evidence synthesis found categorical and string variables extracted more reliably (74% to 96%) than numerical data (47% to 88%), with omissions the dominant error type rather than fabrication.

That finding is directly actionable. Route numeric fields, effect estimates, confidence intervals, event counts, and follow-up times to mandatory human verification. Use model output as a first pass on descriptive fields. Log every field’s provenance either way.

Layer 4: The PICO Evidence Map

Each refresh cycle regenerates the map: new studies land in the cells they inform, gap flags update, and feasibility signals for indirect treatment comparisons change as connecting studies appear. Because the guidance requires you to justify any PICO for which no evidence is submitted, a live gap register is the source document for that justification.

Layer 5: Provenance and Audit

Every automated decision needs a reason code, a model and version reference, a threshold, and a named human sign-off. That expectation now has institutional weight. The NICE position statement on AI in evidence generation asks submitting organizations to justify AI use, prefer more explainable methods, and keep humans in decision-making. The joint position statement from Cochrane, the Campbell Collaboration, JBI, and the Collaboration for Environmental Evidence goes further, backing the RAISE framework and stating that any use of AI or automation that makes or suggests judgments should be fully and transparently reported.

Comparing the Three Operating Models

Key Benefits for Health Technology Developers

Living evidence mapping delivers measurable operational advantages. Time to access HTA-relevant evidence drops from months to on-demand. Consistency improves because the same underlying evidence base serves multiple PICOs and later national requirements. Traceability strengthens through audit logs of AI decisions and human overrides. Resource allocation shifts from repetitive screening toward higher-value synthesis and strategic interpretation.

The long-term cost advantage becomes clear after initial setup. Although a living systematic literature review requires upfront investment, continuous updates demand far fewer resources than restarting evidence reviews from scratch. Teams avoid the common pattern of commissioning multiple overlapping SLRs for different markets or questions. Comparative analysis of AI-assisted versus traditional approaches further clarifies these efficiency differences. See AI-Assisted vs. Traditional Systematic Literature Reviews for side-by-side considerations.

Practical Use Cases in the JCA Context

Consider an oncology asset approaching marketing authorization. A living SLR covering the broader disease area runs continuously. When the JCA scope arrives with three distinct PICO questions, the team filters the existing repository, extracts the relevant subset, and generates tailored evidence tables within days. Congress data published after the initial search integrate automatically.

In another scenario, a developer maintains living SLRs across a portfolio of assets in the same therapeutic area. Shared infrastructure reduces marginal cost for each new indication. When national agencies request additional comparators or real-world evidence, the same map supports rapid expansion.

These approaches also support earlier strategic planning. Teams can explore potential comparator landscapes and evidence gaps months before formal scoping, reducing last-minute surprises.

Best Practices for Implementation

Successful deployment follows several principles. First, ground the living SLR in a publicly registered protocol that follows PRISMA and Cochrane standards. Transparency builds trust with assessors.

Second, keep humans in the loop for all final inclusion, extraction, and synthesis decisions. Document AI model performance characteristics, including recall, precision, and any known limitations, so that reviewers can evaluate reliability. The future of systematic reviews rests on AI assistance rather than full automation, as detailed in The Future of Systematic Literature Reviews is AI-Assisted, Not AI-Automated.

Third, design the data model around PICO elements from the start. This structure enables rapid filtering when assessment questions change.

Fourth, establish clear governance for updates. Define triggers for full versus incremental review and maintain version control of the evidence map.

Fifth, prepare for audit. Retain complete search logs, screening decisions, and AI prompt histories. JCA and national processes require full reproducibility.

Risks, Limitations, and Mitigation Strategies

Key Risks and Mitigation Strategies for Trustworthy Living SLRs

AI tools introduce new risks. Model performance varies by therapeutic area and study type. Hallucinations or extraction errors remain possible, especially with complex tables or poorly structured full texts. Over-reliance on automation can erode methodological rigor if teams skip validation steps.

Guidance from HTA bodies on AI use in SLRs remains limited and inconsistent. Some agencies accept validated classifiers for screening under specific conditions; others provide little formal direction. Developers must therefore apply conservative standards and be prepared to justify methods.

Living approaches also require sustained organizational commitment. Continuous surveillance needs dedicated resources and clear ownership. Without proper governance, a living SLR can drift or become outdated in practice.

Mitigation centers on hybrid design, transparent reporting, independent validation of AI components, and alignment with emerging standards such as extensions to PRISMA for living reviews.

What Comes Next

Three shifts are worth planning for. Portfolio load rises as orphan medicines enter scope in 2028 and all new centrally authorized medicines follow in 2030, which turns per-product effort into an operating model problem. Reporting standards for AI-assisted synthesis are consolidating around frameworks like RAISE, which makes disclosure easier to standardize and harder to skip. And the same evidence base that feeds a JCA dossier increasingly feeds living guidelines and national appraisals, so the value of maintaining it keeps rising.

Conclusion

The JCA does not reward faster document production. It rewards teams whose evidence base is already current, already indexed by PICO, and already auditable when the scope lands. Living SLRs supply the currency, AI supplies the throughput at the screening and first-pass extraction stages, and human methodologists supply the judgment that assessors are actually evaluating. Build the asset before the clock starts, and the 100 days become an assembly window rather than a research project.

If you are mapping your own readiness, the MadeAi literature review platform documentation covers how continuous SLR workflows and traceability work in practice, and the market access and HEOR pages set out the wider workflow. Our analysis of continuous evidence generation is a useful companion to this piece.

Author’s Note: This article was supported by AI-based research and writing, with Claude 5 assisting in the creation of text and images.