How AI Is Transforming Systematic Literature Reviews: PRISMA Compliance and Audit Trails

A systematic literature review takes a team of five researchers up to a year to complete. The output is a static document that may already be partially outdated by the time it is incorporated into an HTA dossier or regulatory submission. For HEOR teams under pressure to submit faster, cover broader evidence bases, and satisfy increasingly specific HTA body documentation requirements, the SLR bottleneck is a direct constraint on market access timelines.[1]

A pragmatic review published in Frontiers in Pharmacology in 2025 examined 25 studies on AI automation in evidence synthesis. In 17 of those studies, researchers observed a time reduction greater than 50%, with five to six-fold decreases in abstract review time in the most efficient deployments. [1] ISPOR CEO Rob Abbott, commenting on the 2026 to 2027 Top 10 HEOR Trends report where AI ranked number one, noted that AI can perform systematic literature reviews in a fraction of the time. [2] But the critical question for HEOR and regulatory teams is not whether AI is faster. It is whether AI-generated SLRs meet PRISMA compliance standards and can withstand HTA audit scrutiny. Pienomial's KnolAI, as an AI research platform built for life sciences, is designed to answer yes to both questions.[9]

1. What PRISMA 2020 Compliance Actually Requires

PRISMA 2020, the Preferred Reporting Items for Systematic reviews and Meta-Analyses updated statement published in 2021, is the international evidence synthesis reporting standard accepted by NICE, G-BA, ICER, Cochrane, and major peer-reviewed journals. It provides a 27-item checklist covering every stage of an SLR from protocol registration to results reporting.[4]

Five requirements in the PRISMA 2020 checklist are most frequently inadequately addressed in AI-assisted SLRs. First, protocol pre-registration: the SLR protocol including the PICOS framework, search strategy, inclusion and exclusion criteria, and data extraction plan must be registered before the search is conducted. AI tools that generate SLRs without documented protocol pre-registration fail this requirement. Second, documented search strings: the exact search strings used in each database, with date of search, must be reported in a form that allows independent replication. Search strings generated and applied internally by an AI tool without documentation fail this requirement. Third, dual independent screening: PRISMA requires that title, abstract, and full-text screening be conducted by at least two independent reviewers with inter-rater reliability reported. A single AI screening pass without dual-reviewer simulation fails this requirement. Fourth, PRISMA flow diagram: the flow of records from identification through inclusion must be documented with exact numbers at each stage including exclusion reasons. Fifth, risk of bias assessment: each included study must be assessed using a validated instrument, RoB 2 for randomised controlled trials and ROBINS-I for observational studies, with domain-level ratings and rationale.[8]

2. Where Generic AI SLR Tools Fail the Compliance Test

Five specific PRISMA failure modes appear consistently in generic AI literature review tools, each of which creates a vulnerability in HTA submissions.[3]

Failure Mode 1, Undocumented search strategy: Tools including general-purpose LLMs and some AI literature review platforms generate search results from proprietary internal processes without documenting the exact search strings, the databases searched, or the date of execution. A reviewer cannot independently replicate the search. PRISMA item 7 requires full search strategy documentation for all databases.

Failure Mode 2, Single-pass AI screening: AI tools typically screen abstracts in a single pass without simulating a second independent screener. PRISMA requires dual independent screening with inter-rater agreement reported, typically as Cohen's kappa. Without a documented dual-screen simulation and kappa calculation, the screening process fails PRISMA item 8.

Failure Mode 3, Missing PRISMA flow diagram: Most AI literature tools produce lists of included papers without automatically generating a PRISMA flow diagram showing the number of records at each stage with exclusion reasons at each step. PRISMA item 17 requires this diagram as a mandatory reporting element.

Failure Mode 4, Untraced extraction outputs: AI-extracted data, including efficacy endpoints, sample sizes, and statistical results, is often presented without linking each extracted data point to a specific table, figure, or section in the source paper. PRISMA item 20 requires that data can be verified against the source at a specific location within the document.

Failure Mode 5, No risk of bias assessment: Generic AI tools do not have validated RoB 2 or ROBINS-I capability. The quality assessment step is missing entirely, leaving the SLR without the methodological quality characterisation that HTA bodies require for evidence weighting.[6]

3. What HTA Bodies Now Require for AI-Assisted SLRs

HTA bodies are developing explicit guidance on AI use in evidence generation, and the direction is consistent: declaration of AI use, methodology documentation, and human oversight evidence are becoming standard requirements.[3]

NICE published a formal position statement in 2024 requiring that when AI is used in evidence generation, reporting must be transparent and must use relevant checklists including PALISADE to justify AI tool selection and TRIPOD+AI for AI model methodology disclosure. [5] NICE also recommends priority screening techniques using machine learning for abstract screening, acknowledging AI efficiency benefits while requiring full methodology transparency. Submitting organisations remain fully accountable for all submitted content regardless of AI assistance.

G-BA's transparency requirements for evidence synthesis methodology are stringent. The dossier must contain sufficient detail for an independent reviewer to assess the credibility of the literature search and screening process. AI-generated searches that cannot be independently reproduced fail this requirement and expose the submission to formal clarification requests. IQWiG, G-BA's scientific advisory body, has not yet issued specific AI guidance but applies its standard transparency requirements to AI-assisted work without modification.[3]

ICER's evidence standards increasingly mirror PRISMA-level methodology documentation requirements for systematic reviews. For evidence reports that inform US payer negotiations, the quality of SLR methodology is a direct input to ICER's assessment of the certainty of clinical evidence, with methodology limitations downgrading the certainty rating in ways that affect the MFP range.[6]

4. The PRISMA-Compliant AI SLR Architecture

A PRISMA-compliant AI SLR architecture requires eight specific capabilities that most generic AI literature tools do not provide.[4]

1. Protocol module: PICOS framework definition with documentation, search string generation with full transparency, database selection justification, inclusion and exclusion criteria specification, and protocol pre-registration support for PROSPERO or OSF.

2. Documented search execution: Automated execution of documented, reproducible search strings across PubMed, Embase, Cochrane, and additional databases. Exact search strings recorded with date and time stamp. Search strings exportable in a format that allows independent replication.

3. Deduplication with documentation: Algorithmic deduplication with documentation of the method used and the number of duplicates removed at each stage. PRISMA flow diagram updated automatically as each stage is completed.

4. Dual-screen simulation: Independent AI screening of title and abstract with inter-rater reliability calculation, simulated kappa calculation, and human resolver workflow for disagreements. Screening decisions logged with the specific inclusion or exclusion criterion applied to each record.

5. Full-text screening with reason documentation: Included records from abstract screening retrieved for full-text review. AI applies inclusion and exclusion criteria to full texts with exclusion reason documentation at the record level.

6. Data extraction with sentence-level attribution: Structured extraction from included studies with each extracted data point linked to a specific table, figure, or section in the source paper. Exportable extraction tables with embedded source citations.

7. Risk of bias assessment: Structured RoB 2 assessment for RCTs with domain-level ratings and judgment rationale. ROBINS-I assessment for observational studies. Exportable risk of bias summary tables.[8]

8. Auto-generated PRISMA flow diagram: PRISMA 2020-compliant flow diagram generated automatically with record counts at each stage. Exportable as an editable figure for submission inclusion.

5. KnolAI SLR Workflow: Step-by-Step PRISMA Mapping

The KnolAI SLR workflow, within the Knolens AI research platform, maps to the PRISMA 2020 checklist at each stage.[9]

Stage 1, Protocol development: The HEOR team defines the PICOS framework. KnolAI generates the structured protocol document with search strategy documentation. The protocol is saved with a timestamp as the pre-registered evidence framework before search execution begins. Stage 2, Database search: KnolAI executes documented searches across PubMed, Embase, Cochrane, and additional databases specified in the protocol. Search strings and execution dates are recorded in the audit trail. Stage 3, Deduplication: Automated deduplication with documentation of the algorithm and the number of duplicates removed. The PRISMA flow diagram is initialised with the total records identified count.

Stage 4, Abstract screening: KnolAI screens each record against the PICOS inclusion and exclusion criteria. Each screening decision is logged with the specific criterion applied. Inter-rater reliability simulation generates a kappa value for the screening batch. Stage 5, Full-text screening: Included abstracts are retrieved for full-text review. KnolAI conducts full-text screening with exclusion reason documentation at the record level. The PRISMA flow diagram is updated automatically with counts at each step. Stage 6, Data extraction: Structured extraction from included studies with sentence-level attribution linking each extracted data point to its source location. Stage 7, Risk of bias assessment: RoB 2 or ROBINS-I assessment for each included study with domain-level judgment and rationale. Stage 8, Output generation: Auto-generated PRISMA flow diagram, structured evidence table in submission format, and complete audit trail for all AI actions at each stage.

6. Time and Cost Comparison: AI-Augmented vs Traditional SLR

The quantified efficiency evidence from Frontiers in Pharmacology is compelling: across 25 studies examining AI automation in evidence synthesis, the majority observed time reductions greater than 50%, with the most efficient deployments showing five to six-fold reductions in abstract review time and 55 to 64% decreases in the number of abstracts requiring human review. [1] Studies examining work saved at 95% recall reported six to ten-fold reductions in workload with automation.

Translating this to a realistic HEOR SLR for an HTA submission, a traditional SLR requiring 5 researchers across 6 months, covering 10,000 abstracts and 200 full texts, becomes a KnolAI-augmented SLR requiring 2 researchers across 6 weeks, with AI handling abstract screening, full-text retrieval, extraction, risk of bias assessment, and PRISMA flow diagram generation. Human researcher time is concentrated on protocol design, borderline screening decisions, extraction validation, and methodology documentation review.

The cost implication is direct. HEOR teams that can reduce SLR timelines from months to weeks, without compromising PRISMA compliance, can produce more evidence, respond to HTA body requests faster, and maintain living SLRs across multiple indications simultaneously rather than treating each SLR as a single-project resource commitment.[6]

7. Living SLRs: The Continuous Evidence Update Model

A traditional SLR is a point-in-time snapshot. The ISPOR 2026 to 2027 report and HTA body guidance from NICE increasingly reference living systematic reviews as the standard for products with ongoing post-market evidence generation, managed access agreements, or frequent indication expansions. A living SLR continuously monitors literature databases for new publications meeting the SLR PICOS criteria, screens new records against the protocol, and integrates new evidence into the existing evidence table.[2]

KnolAI enables living SLRs through the Knolens knowledge layer: the SLR protocol and evidence layer are maintained in the knowledge graph. New publications are ingested and screened continuously against the PICOS protocol. New inclusions trigger notification and automated extraction. The HEOR team reviews and approves new evidence additions before they are incorporated into the active evidence layer.

The value for HTA teams managing conditional reimbursement: NICE managed access agreements frequently require evidence updates at 12 to 36-month intervals. A living SLR infrastructure means these updates are generated from a continuously maintained evidence base rather than requiring a full new SLR project at each review cycle. The audit trail from the original SLR extends through each update, maintaining a complete methodology record from first submission through each conditional review.[9]

8. Integrating SLR Outputs into the HEOR Evidence Workflow

An SLR is rarely an end product. It is an input to downstream HEOR work: network meta-analysis inputs, cost-effectiveness model evidence inputs, and value dossier evidence tables. The integration challenge is that SLR outputs are typically delivered as static documents that must be manually re-entered into NMA software and HE models.[7]

KnolAI addresses this challenge by storing structured extraction outputs in the Knolens knowledge graph as entity-relationship triples: compound, trial, endpoint, result, and source. These triples are directly queryable for NMA input data without manual re-entry. The same structured data feeds the cost-effectiveness model input pack through KnolComposer. Value dossier evidence tables are generated from the structured data layer with full source attribution.

The compounding value: a SLR conducted with KnolAI does not produce a static document. It produces a structured evidence layer that powers all subsequent HEOR work on the same product and indication. Validated evidence from one SLR becomes a reusable asset for related submissions, adaptation studies, and indication extensions, without requiring re-extraction from source documents.[9]

9. How Fast Can Your Team Run a PRISMA-Compliant SLR with KnolAI?

Running a PRISMA-compliant SLR does not need to take months. KnolAI ships with every component of the PRISMA 2020 workflow pre-built: documented search execution across PubMed, Embase, and Cochrane, dual-screen simulation with kappa calculation, auto-generated PRISMA flow diagrams, sentence-level extraction attribution, and RoB 2 and ROBINS-I assessment frameworks. Your team is not configuring a research methodology from scratch. You are activating a product that is already built for PRISMA compliance. [9]

Most HEOR teams complete their first KnolAI-augmented SLR within four to six weeks of onboarding, compared to four to six months for a traditional manual SLR. Here is what the deployment looks like.

Sprint 1, Week 1, PICOS framework and protocol registered: The HEOR team defines the PICOS framework in the KnolAI protocol module. KnolAI generates the structured protocol document with search strings across all target databases, inclusion and exclusion criteria, and data extraction plan. The protocol is time-stamped and saved as the pre-registered evidence framework before any search is executed. PROSPERO or OSF registration is supported directly from the protocol module. No manual protocol document drafting required.

Sprint 2, Weeks 2 to 3, Search, screening, and extraction running: KnolAI executes documented database searches with exact strings recorded and time-stamped. Deduplication runs automatically with counts logged. Abstract screening runs with dual-screen simulation and kappa calculation. Full-text screening follows with exclusion reason documentation at the record level. Extraction runs with sentence-level source attribution linking every data point to its specific location in the source paper. The PRISMA flow diagram updates automatically at each stage. Human researcher time is concentrated on borderline screening decisions and extraction spot-checks.

Sprint 3, Weeks 4 to 5, RoB assessment, output generation, and validation: RoB 2 assessments run for RCTs and ROBINS-I for observational studies with domain-level ratings and rationale. The auto-generated PRISMA flow diagram and structured evidence table are produced in submission format. The validation process runs in parallel: senior reviewer checks a sample of search results, human reviewers independently screen 10 to 15% of AI screening decisions to generate the documented kappa, and domain experts spot-check extraction accuracy. The validation report is generated automatically as part of the methodology documentation package required by NICE's 2024 position statement. [5]

From this point, the SLR output is not a static document. It is a living evidence layer in the Knolens knowledge graph. New publications are screened continuously against the PICOS protocol. NICE managed access agreement update cycles are served from the same continuously maintained evidence base, not a new full SLR project. [2]

Conclusion

AI-powered systematic literature reviews are not a compromise on quality. They are a quality improvement opportunity if the AI architecture is built for PRISMA compliance and HTA audit readiness. The risk is not that AI will produce lower-quality SLRs than manual processes. The risk is that teams will deploy AI tools not designed for PRISMA compliance and submit SLRs that fail HTA scrutiny, wasting the time savings and damaging submission credibility.[3]

Pienomial's KnolAI delivers AI-augmented SLRs as a purpose-built AI research platform and AI governance platform for life sciences, with PRISMA 2020-compliant workflows, source attribution at the sentence level, auto-generated PRISMA flow diagrams, and full audit trail documentation designed to satisfy NICE, G-BA, ICER, and Cochrane requirements. [9] CTA: Request a KnolAI SLR Demo or Download the PRISMA AI Compliance Checklist.