A Living Systematic Review Engine: LLM-Automated Evidence Surveillance Validated Against a Published Meta-Analysis of Statins for Sepsis

A Living Systematic Review Engine: LLM-Automated Evidence Surveillance Validated Against a Published Meta-Analysis of Statins for Sepsis | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Method Article A Living Systematic Review Engine: LLM-Automated Evidence Surveillance Validated Against a Published Meta-Analysis of Statins for Sepsis Hayden Farquhar This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9308492/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Systematic reviews take a mean of 67 weeks to complete and are frequently outdated by publication. Living systematic reviews aim to address this through continuous updating, but the manual labour required makes them impractical at scale. Large language models (LLMs) offer the potential to automate the full review pipeline, yet most existing tools address only abstract screening. Objective To develop and validate an end-to-end LLM-automated living systematic review engine that performs literature searching, title/abstract screening, structured data extraction, risk of bias assessment, and random-effects meta-analysis, and to demonstrate its capacity for continuous evidence surveillance. Methods We built a Python pipeline using PubMed E-utilities for automated searching, Claude Haiku for triple-run title/abstract screening with majority voting, and Claude Sonnet for two-pass data extraction with self-verification and Cochrane Risk of Bias 2 (RoB 2) assessment. Meta-analysis used the DerSimonian–Laird random-effects estimator. The pipeline was validated against a published meta-analysis of statins for sepsis (Chen et al., 2018; 9 randomised controlled trials [RCTs], 2333 patients). Performance was evaluated on screening sensitivity/specificity, extraction accuracy against ground truth event counts, and concordance of pooled effect estimates. The living update capability was demonstrated by incorporating studies published after the reference search date. The study is reported following the PRISMA 2020 guidelines where applicable [ 1 ]. Results The PubMed search retrieved 215 candidate articles, capturing all 8 reference studies with known PubMed identifiers (100% recall). LLM screening achieved sensitivity of 1.00 (8/8 studies identified), specificity of 0.98 (203/207 correctly excluded), and Cohen’s kappa of 0.79 against the reference standard, with perfect inter-rater agreement (kappa = 1.00) across screening runs. Data extraction achieved exact agreement on mortality event counts for 4 of 7 studies with available ground truth (57%, 95% CI 18–90%); the remaining discrepancies were attributable to abstract-only access limitations rather than extraction failures. The LLM-derived pooled risk ratio (RR 0.94, 95% CI 0.74–1.21; k = 5; I² = 6%) was concordant with the reference estimate (RR 0.96, 95% CI 0.83–1.11; k = 7; I² = 20%), reaching the same clinical conclusion: no significant mortality benefit from statins in sepsis. The pipeline independently identified one new RCT published in 2025 (Rashid et al.) not available to the reference review. Incorporating this study updated the pooled estimate to RR 0.89 (95% CI 0.70–1.14; k = 6) without changing the overall conclusion. Total pipeline execution time was 32 minutes at a cost of US $ 1.95, compared with an estimated 58 hours and US $ 2,900–5,800 for manual review. Conclusions An LLM-automated pipeline can replicate the screening, extraction, and synthesis stages of a systematic review with high fidelity, at a fraction of the time and cost. The pipeline’s detection of a post-reference RCT demonstrates practical living review capability. Key limitations include reduced extraction accuracy when full text is unavailable and restriction to a single database. The open-source pipeline provides a foundation for scalable, continuously updated evidence synthesis. Registration The embedded systematic review was not prospectively registered on PROSPERO, as the primary study objective was pipeline validation. living systematic review large language models evidence synthesis automation meta-analysis statins sepsis Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Systematic reviews are a foundation of evidence-based medicine [ 2 ], yet the time required to produce them has become a barrier to their usefulness. An analysis of the PROSPERO registry found that systematic reviews take a mean of 67.3 weeks to complete, with screening and data extraction consuming the largest share of reviewer effort [ 3 ]. By the time a review is published, the evidence it summarises may already be outdated: Shojania et al. found that 7% of reviews already had a signal for updating at the time of publication, rising to 23% within two years [ 4 ]. Living systematic reviews — reviews that are continuously updated as new evidence emerges — have been proposed as a solution [ 5 ]. The concept is endorsed by Cochrane [ 2 ], and several living review protocols have been registered. However, the practical burden of monthly or quarterly searching, screening, extraction, and re-analysis has limited adoption. Most existing living reviews rely on partial automation of a single stage, typically abstract screening, while the remaining workflow remains manual [ 6 ]. Even high-profile implementations such as the COVID-NMA living review consortium have required sustained manual effort to maintain currency [ 7 ]. Automation tools for systematic reviews have existed for some time. ASReview uses active learning to prioritise screening [ 8 ], and RobotReviewer automates risk of bias assessment using machine learning [ 9 ]. The emergence of large language models (LLMs) has expanded what can be automated. LLMs have achieved sensitivities of 0.76–1.00 for abstract screening in several evaluations [ 10 , 11 , 12 ], and recent work has demonstrated their use for data extraction [ 13 ] and risk of bias assessment [ 14 , 15 ]. However, no published system integrates all stages — from literature search through meta-analytic synthesis — into a single automated pipeline with validated end-to-end performance. In this study, we present an open-source LLM-automated living systematic review engine and validate it against a published meta-analysis of RCTs of statins for sepsis [ 16 ]. We chose this demonstration case because it has a well-defined reference standard, an active trial landscape, and genuine clinical equipoise — making it a suitable test case for continuous updating. Methods Study design This is a methods validation study comparing the outputs of an LLM-automated systematic review pipeline against a published human-conducted meta-analysis. The study involved no patient data and required no ethics approval. The embedded review of statins for sepsis was not prospectively registered on PROSPERO, as the primary objective was pipeline validation rather than a de novo review. The study is reported following the PRISMA 2020 guidelines [ 1 ] where applicable to the automated review component (see Additional file 2: PRISMA Checklist). Reference standard We used the meta-analysis by Chen et al. (2018) as the primary reference standard [ 16 ]. This review included 9 RCTs (2333 patients) examining statin therapy versus placebo or no statin in adults with sepsis, septic shock, or related critical illness. The pooled risk ratio for in-hospital mortality was 0.96 (95% CI 0.83–1.11; I² = 20%; k = 7), and for 28-day mortality was 0.90 (95% CI 0.73–1.11; I² = 29%; k = 5). The review concluded that statin therapy had no significant effect on mortality. Eight of the nine included studies had PubMed identifiers (PMIDs) and were used for validation; one study (El Gendy 2014) lacked a PMID and was excluded from PMID-based validation. Pipeline architecture The pipeline consists of four automated stages executed sequentially in Python (Fig. 1 ): Stage 1 — Literature search. PubMed was searched via the NCBI E-utilities API using a Boolean query combining MeSH terms and free-text synonyms for statins (including individual drug names), sepsis and related conditions (septic shock, ARDS, ventilator-associated pneumonia, critical illness), and an RCT publication type filter. The search was designed for high sensitivity, deliberately broad to capture studies where sepsis-related conditions were secondary indications. Raw search results (XML) were archived for reproducibility. Stage 2 — Title/abstract screening. Each retrieved article was screened against predefined inclusion and exclusion criteria using Claude Haiku 4.5 (Anthropic; model identifier in Supplementary Materials). Inclusion criteria required: (a) primary RCT or quasi-experimental design, (b) adult patients with sepsis or infection-related critical illness, (c) statin intervention of any type, (d) placebo or no-statin comparator, and (e) reporting of mortality, ICU length of stay, or organ dysfunction outcomes. Exclusion criteria included: non-primary studies (reviews, secondary analyses, post-hoc analyses), paediatric populations, non-infection indications, and animal studies. The LLM was instructed to apply a “when in doubt, include” principle, excluding only when confidence exceeded 0.95. To assess decision stability, each article was screened three times at temperature 0.3, and the final decision was determined by majority vote. If no clear majority existed, the article defaulted to “include.” All screening decisions, confidence scores, criterion-level assessments, and token usage were logged to a JSONL file for full reproducibility. Stage 3 — Data extraction. Included studies underwent structured extraction using Claude Sonnet 4 (Anthropic; model identifier in Supplementary Materials) in a three-pass process: (a) initial PICO (Population, Intervention, Comparator, Outcome) and effect size extraction using a structured JSON prompt, (b) self-verification in which the LLM reviewed its own extraction against the source text with specific instructions to check for percentage-versus-count errors and arm size completeness, and (c) risk of bias assessment using the Cochrane RoB 2 framework [ 17 ]. Where available, full text was retrieved from PubMed Central (PMC) Open Access; otherwise, extraction was performed from the abstract alone. A title/author sanity check was implemented to detect and reject mismatched PMC articles. All extractions used temperature 0.0 for maximum determinism. Stage 4 — Meta-analytic synthesis. Pooled effect estimates were computed using the DerSimonian–Laird random-effects model [ 18 ]. Risk ratios were calculated from extracted event counts with continuity correction (0.5) applied to zero-cell studies. Heterogeneity was assessed using Cochran’s Q test and the I² statistic [ 19 ]. Prediction intervals were computed using the t-distribution with k − 2 degrees of freedom. Publication bias was evaluated using Egger’s regression test and visual inspection of funnel plots, noting that both methods have limited reliability with fewer than 10 studies [ 2 ]. Sensitivity analyses included leave-one-out analysis and subgroup analysis by statin type. Living update demonstration To demonstrate the pipeline’s capacity for continuous evidence surveillance, we compared: (a) a baseline analysis using only studies available to the 2018 reference review, and (b) an updated analysis incorporating studies published after the reference search date that the pipeline independently identified and included. Validation metrics Screening performance was evaluated by sensitivity (proportion of reference-included studies correctly identified), specificity (proportion of non-included studies correctly excluded), positive predictive value (PPV), negative predictive value (NPV), and Cohen’s kappa against the reference inclusion decisions. Inter-rater agreement between screening runs was assessed using pairwise Cohen’s kappa. Extraction accuracy was evaluated by comparing LLM-extracted mortality event counts against ground truth values from the reference review. Timepoint-aware matching was used: if the LLM extracted 90-day mortality, it was compared against 90-day ground truth where available. An exact match required identical event counts; a “within 5%” match allowed for rounding from percentage conversion. Given the small number of studies, exact binomial 95% confidence intervals were computed for accuracy proportions. Synthesis concordance was assessed by comparing the direction, magnitude, and statistical significance of pooled effect estimates, as well as heterogeneity statistics. Continuous deployment as a living review An important distinction exists between a pipeline that can be re-run and a review that is continuously updated. To bridge this gap, we deployed the pipeline as a GitHub Actions scheduled workflow that executes automatically on the 1st of each month. Each cycle performs the full pipeline (search, screen, extract, synthesise), commits updated results (forest plot, meta-analysis, extraction data) to the repository, and opens a GitHub issue if new studies are detected or the evidence base changes. This provides a transparent, version-controlled audit trail of every update cycle. The deployment requires two API keys (NCBI and Anthropic) stored as repository secrets, and no dedicated infrastructure — the workflow runs on GitHub’s free-tier runners. Monthly execution cost is approximately US $ 1.70 per cycle, dominated by screening API calls. Full deployment instructions are provided in the Additional file 1 (DEPLOYMENT.md) to enable readers to fork the repository and deploy a living review on any clinical topic by modifying only the search query, screening criteria, and extraction prompts. Software and reproducibility The pipeline was implemented in Python 3.14 using Pydantic for data modelling, httpx for asynchronous API calls, SciPy for statistical computations, and Matplotlib for figure generation. All API calls (prompts, responses, token counts, costs) were logged. Source code is available at https://github.com/hayden-farquhar/Living-Systematic-Review . In this study, we used Anthropic’s Claude Haiku 4.5 for screening and Claude Sonnet 4 for extraction. However, the pipeline architecture is model-agnostic: the LLM is accessed through a single abstraction layer, and any instruction-following model with structured output capability can be substituted. As of April 2026, suitable alternatives include OpenAI’s GPT-4o and GPT-4o-mini, Google’s Gemini 2.5 Pro and Flash, Meta’s Llama 4 (via API providers or local deployment), and Mistral Large. The choice of model involves a trade-off between cost, accuracy, and speed — smaller models suit high-volume screening, while larger models with stronger reasoning are preferred for extraction. We encourage future work comparing pipeline performance across model providers. Results Literature search The PubMed search retrieved 215 unique articles (Fig. 2 ). All 8 reference studies with known PMIDs were captured (100% recall). The search was intentionally broad, including terms for ARDS and ventilator-associated pneumonia to capture studies where the sepsis population overlapped with these conditions. Screening performance Of 215 articles screened, 12 were included (8 reference studies plus 4 additional) and 203 were excluded (Table 1 ). Screening sensitivity was 1.00 (8/8 reference studies correctly identified) and specificity was 0.98 (203/207). Cohen’s kappa against the reference was 0.79, indicating substantial agreement [ 20 ]. NPV was 1.00 — no reference studies were missed. Table 1 Screening Performance Against Reference Standard Metric Value Articles searched 215 Articles included by LLM 12 Reference studies identified 8/8 (100%) Sensitivity 1.00 Specificity 0.98 Positive predictive value 0.67 Negative predictive value 1.00 Cohen’s kappa (vs reference) 0.79 Inter-rater kappa (between runs) 1.00 False negatives 0 False positives 4* Screening model Claude Haiku 4.5 Runs per article 3 (majority vote) Temperature 0.3 Total API calls 645 Cost US $ 1.70 *Two of four false positives were post-2018 RCTs not available to the reference review (Rashid 2025, Sapey 2019); one was a long-term follow-up of an included trial (Agus 2017); one was a mechanistic study with a small embedded RCT (Ghosh 2015). The four false positives included two studies published after the reference search date (Rashid et al. 2025, a primary RCT; Sapey et al. 2019, a pilot RCT), one long-term follow-up of an included trial (Agus et al. 2017, HARP-2 12-month data), and one mechanistic study with a small embedded RCT component (Ghosh et al. 2015). Two of these represent genuine new evidence rather than true errors. Inter-rater agreement across the three screening runs was perfect (pairwise kappa = 1.00 for all pairs), with 100% unanimous agreement (215/215 articles). This reflects the near-deterministic behaviour of the LLM at temperature 0.3. Total screening cost was US $ 1.70 (645 API calls; 1,091,499 input tokens; 207,202 output tokens). Data extraction accuracy Of 8 studies extracted, 5 yielded mortality event counts suitable for meta-analysis. Three studies lacked extractable data: two because their abstracts reported only “no significant difference” without numeric event counts (Patel 2012, Truwit 2014), and one because the PMC full-text link pointed to an incorrect article and the abstract reported only percentages without arm-level denominators (Singh 2017). The full-text sanity check correctly rejected the two mismatched PMC articles. Among 7 studies with ground truth mortality data available for comparison, 4 (57%, 95% CI 18–90%) had exact matches on event counts (Table 2 ). The 3 discrepancies were: (a) Truwit 2014 (SAILS trial) — the abstract reported percentages (28.5% vs 24.9%) but not the per-arm sample sizes (379 vs 366), preventing accurate count derivation; (b) Papazian 2013 — the LLM correctly converted percentages to counts but assumed equal arm sizes (150/150) when the actual split was 146/138; and (c) Patel 2012 — the abstract contained no numeric mortality data. All three discrepancies were attributable to information absence in the abstract rather than extraction errors by the LLM. When the source text contained explicit event counts, extraction accuracy was 4/4 (100%). Table 2 Data Extraction Accuracy Study Year Source Mortality (LLM) Mortality (Reference) Timepoint Match Novack 2009 Abstract 0/42 vs 0/41 0/42 vs 0/41 Hospital Exact Kruger 2011 Abstract 6/75 vs 4/75 6/75 vs 4/75 Hospital Exact Patel 2012 Abstract —/49 vs —/51 2/49 vs 2/51 28-day Missing* Kruger 2013 Abstract 18/123 vs 24/127 18/123 vs 24/127 90-day Exact Papazian 2013 Abstract 32/150 vs 23/150 31/146 vs 21/138 28-day Close** McAuley 2014 Abstract 57/259 vs 75/281 57/259 vs 75/281 28-day Exact Truwit 2014 Abstract*** —/— vs —/— 108/379 vs 91/366 Hospital Missing* Singh 2017 Abstract*** —/— vs —/— N/A 28-day N/A Exact match: 4/7 (57%, 95% CI 18–90%). Abstract did not report numeric event counts. LLM assumed equal arm sizes (150/150); actual split was 146/138. PMC full-text link pointed to incorrect article; sanity check reverted to abstract. Extraction model: Claude Sonnet 4. Method: two-pass (initial extraction + self-verification) + RoB 2 assessment. Cost: US $ 0.25. Self-verification (the second LLM pass) successfully corrected percentage-to-count conversion errors in preliminary extraction, improving accuracy from 2/7 to 4/7 exact matches. Total extraction cost was US $ 0.25 (24 API calls across 8 studies). Meta-analytic synthesis The LLM-derived pooled risk ratio from 5 studies with extractable mortality data was 0.94 (95% CI 0.74–1.21; I² = 6.0%; τ² = 0.006; Cochran’s Q = 4.26, p = 0.37), compared with the reference estimate of 0.96 (95% CI 0.83–1.11; I² = 20%; k = 7) (Fig. 3 , Table 3 ). Both analyses reached the same conclusion: statin therapy does not significantly reduce mortality in sepsis. Table 3 Meta-Analytic Synthesis — LLM Pipeline vs Reference Metric LLM Pipeline Chen et al. 2018 (Reference) Studies contributing 5 7 Pooled RR (95% CI) 0.94 (0.74–1.21) 0.96 (0.83–1.11) I² 6% 20% τ² 0.006 Not reported Cochran’s Q (p-value) 4.26 (p = 0.37) Not reported Egger’s test p-value 0.46* Not reported Conclusion No significant effect No significant effect *Egger’s test has limited reliability with fewer than 10 studies. The wider confidence interval in the LLM-derived estimate reflects the smaller number of contributing studies (5 vs 7), a direct consequence of abstract-only access limiting data extraction. Egger’s test showed no evidence of publication bias (p = 0.46), though this test has limited power with fewer than 10 studies [ 2 ] (Fig. 4). Leave-one-out sensitivity analysis confirmed that the pooled estimate was robust to the exclusion of any single study. Subgroup analysis by statin type showed no significant effect for either atorvastatin (RR 0.87, 95% CI 0.52–1.44; k = 2) or simvastatin (RR 1.02, 95% CI 0.67–1.53; k = 3). Living update demonstration The pipeline independently identified one new primary RCT published in 2025 (Rashid et al.), an open-label trial of atorvastatin 20 mg versus standard care in 68 adults with sepsis. This study reported 28-day mortality of 36% (13/36) in the atorvastatin group versus 56% (18/32) in the control group (p = 0.10). Incorporating this study updated the pooled estimate to RR 0.89 (95% CI 0.70–1.14; k = 6; I² = 15%) without changing the overall conclusion (Fig. 5 , Table 4 ). Table 4 Living Update — Cumulative Meta-Analysis Analysis k Pooled RR 95% CI I² Conclusion Chen et al. 2018 (reference) 7 0.96 0.83–1.11 20% No significant effect LLM pipeline (baseline) 5 0.94 0.74–1.21 6% No significant effect + Rashid 2025 6 0.89 0.70–1.14 15% No significant effect + Rashid 2025 + Agus 2017 (12mo) 7 0.87 0.74–1.01 0% No significant effect* *Upper confidence bound approaching 1.0; continued monitoring warranted. Further incorporating the 12-month mortality data from the HARP-2 follow-up (Agus et al. 2017; 31.8% vs 37.3%) yielded a pooled RR of 0.87 (95% CI 0.74–1.01; k = 7; I² = 0%). While the upper confidence bound approached 1.0, the conclusion of no statistically significant benefit remained unchanged. This progressive narrowing of the confidence interval with cumulative evidence illustrates the value of continuous monitoring. The pipeline also detected a pilot RCT (Sapey et al. 2019) of high-dose simvastatin in community-acquired pneumonia with sepsis, though this study did not report extractable mortality event counts. Time and cost Total pipeline execution time was approximately 32 minutes, dominated by the screening stage (25 minutes for 645 sequential API calls). Total API cost was US $ 1.95 (screening: US $ 1.70; extraction: US $ 0.25). By comparison, manual conduct of the equivalent review stages would require an estimated 58 hours of reviewer time at a cost of US $ 2,900–5,800 for two trained reviewers (Table 5 ) [ 3 , 21 ]. The pipeline achieves a roughly 110-fold reduction in time and greater than 99% reduction in direct costs. Monthly living update cycles, requiring re-execution of the search and screening stages only for new publications, would cost approximately US $ 1.70 per cycle. Table 5 Time and Cost Comparison Stage Human time (hours) Pipeline time (min) Pipeline cost (US $ ) Speedup Literature search 8 0.5 0.00 960× Title/abstract screening 18 25 1.70 43× Full-text retrieval 4 1 0.00 240× Data extraction + RoB 16 5 0.25 192× Meta-analysis 8 < 0.1 0.00 4800× Report generation 4 < 0.1 0.00 2400× Total 58 32 1.95 110× Human estimates assume 2 trained reviewers. Hourly rate of US $ 50–100 is based on typical research assistant and systematic reviewer compensation in academic settings [ 3 ]. Pipeline estimates: single end-to-end execution. Monthly living update (search + screening only): US $ 1.70/cycle (~ US $ 20/year). Discussion Principal findings We demonstrate that an LLM-automated pipeline can replicate the core stages of a systematic review — screening, data extraction, risk of bias assessment, and meta-analytic synthesis — with sufficient fidelity to reach the same clinical conclusion as a human-conducted review, at a fraction of the time and cost. The pipeline’s independent detection of a post-reference RCT validates its utility for living evidence surveillance. Comparison with existing work Previous studies have demonstrated LLM performance on individual review stages. Abstract screening with GPT-4 has achieved sensitivities of 0.76–1.00 depending on the review topic and prompt design [ 10 , 11 , 12 ]. Our finding of 1.00 sensitivity with 0.98 specificity is consistent with the upper range of reported performance. The perfect inter-rater agreement (kappa = 1.00) at low temperature adds a dimension not previously reported, though it also reveals a limitation of the triple-run design: at temperature 0.3, the LLM is effectively deterministic, and the majority vote reduces to a single decision. A higher temperature setting would introduce genuine stochastic variation between runs, enabling meaningful inter-rater analysis — but at the potential cost of screening accuracy. Future work should explore this trade-off. Data extraction using LLMs has been less extensively studied. Khan et al. (2025) reported concordant extraction accuracy of 0.94 using a collaborative two-LLM approach with cross-critique [ 13 ]. Our finding of 57% exact match on mortality event counts is lower, but the comparison is not straightforward: our denominator includes studies where the source text contained no extractable data. When the abstract or full text explicitly reported event counts, our extraction accuracy was 100% (4/4). The gap between these figures highlights a structural constraint — abstract-only extraction has an inherent ceiling — rather than a model capability limitation. Existing SR automation tools address individual stages: ASReview handles screening prioritisation [ 8 ], RobotReviewer automates risk of bias assessment [ 9 ], and various tools assist with search strategy development [ 6 ]. To our knowledge, this is the first validated pipeline that integrates automated search, screening, extraction, synthesis, and living update capability into a single system with quantified performance at each stage. Strengths The pipeline logs every API call, prompt, response, and intermediate decision, enabling reproducibility at a granularity that manual reviews rarely achieve [ 21 ]. The cost structure makes continuous evidence surveillance economically feasible for any clinical question. The validation design reports failures (extraction from abstracts) alongside successes (screening, synthesis concordance), rather than selecting only favourable metrics. Limitations Extraction accuracy is substantially constrained by full-text availability. When only abstracts are accessible, studies reporting mortality as percentages without arm-level denominators cannot be accurately converted to event counts. This affected 3 of 8 studies in our validation and represents the most significant barrier to fully automated extraction. Institutional full-text access or publisher API integration would likely improve this, but was beyond the scope of this study. The pipeline searches PubMed only; comprehensive systematic reviews typically search multiple databases (Embase, CENTRAL, Web of Science), grey literature, and trial registries [ 2 ]. The reference standard was a published meta-analysis rather than a prospectively conducted review, introducing the possibility that the reference itself contained errors. Validation was performed on a single clinical topic with a small number of included studies (8 with PMIDs); the 57% extraction accuracy has wide confidence intervals (95% CI 18–90%), and generalisability to other review questions requires further evaluation. The pooled meta-analysis combined studies reporting mortality at different timepoints (hospital, 28-day, and 90-day), which introduces clinical heterogeneity that the I² statistic may not fully capture. A human reviewer would typically standardise timepoints or conduct separate analyses by timepoint; the automated pipeline extracted whichever timepoint was most prominently reported. LLM outputs are non-deterministic across model versions, and performance may vary as models are updated. The pipeline depends on commercial API services whose availability and pricing may change. Implications The economics of living reviews change when the marginal cost of an update cycle falls below US $ 2.00. Organisations maintaining portfolios of living reviews could automate monthly surveillance across dozens of topics simultaneously. The pipeline’s detection of Rashid et al. (2025), a new RCT published seven years after the reference meta-analysis, demonstrates that automated surveillance can identify evidence that static reviews miss — the failure mode that living reviews are designed to prevent. The automated pipeline logs every decision in a format that is difficult to achieve in manual review processes [ 21 ]. By providing a complete GitHub Actions workflow alongside an open-source, config-driven pipeline, readers can fork the repository, substitute their own PICO criteria and search terms, and have a living review running on any clinical question with no dedicated infrastructure. The annual cost of approximately US $ 20 for monthly updates makes this feasible for unfunded researchers and guideline groups in resource-limited settings. Sharing a deployable system — not just code — is essential for translating automation research into practice. Future directions Integration of additional databases (Embase, CENTRAL) and clinical trial registries (ClinicalTrials.gov) would improve search comprehensiveness. Institutional full-text access could substantially improve extraction accuracy. Parallel screening using multiple LLM providers would enable cross-model validation and reduce dependence on any single vendor. Evaluation across diverse clinical topics — including reviews with larger numbers of included studies, continuous outcomes, and more complex eligibility criteria — is needed to establish generalisability. Conclusions An LLM-automated living systematic review pipeline can replicate the findings of a human-conducted meta-analysis of statins for sepsis, achieving 100% screening sensitivity and concordant pooled effect estimates, in 32 minutes for US $ 1.95. The pipeline’s identification of a post-reference RCT demonstrates practical living review capability. While extraction accuracy remains limited by full-text access, the approach provides a validated, transparent, and economically viable foundation for continuous evidence synthesis, with potential for scaling across clinical topics. Declarations Ethics approval and consent to participate: No ethics approval was required. This study involved no patient data; all analyses used published literature. Consent for publication: Not applicable. Competing interests: The author declares no competing interests. Funding: No external funding was received for this study. Author Contribution HF conceived the study, designed and implemented the pipeline, conducted all analyses, and wrote the manuscript. The author read and approved the final manuscript. Acknowledgements: Large language models (Anthropic Claude Haiku 4.5 and Claude Sonnet 4) were used as integral components of the pipeline under evaluation, as described in the Methods. Claude Code was used as a coding assistant during pipeline development. The author takes full responsibility for the content and accuracy of this manuscript. Data Availability All code, data, prompts, and API logs are available at https://github.com/hayden-farquhar/Living-Systematic-Review. References Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. Higgins JPT, Thomas J, Chandler J, et al., editors. Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Cochrane; 2024. Available from: . Borah R, Brown AW, Capers PL, Kaiser KA. Analysis of the time and workers needed to conduct systematic reviews of medical interventions using data from the PROSPERO registry. BMJ Open. 2017;7(2):e012545. Shojania KG, Sampson M, Ansari MT, Ji J, Doucette S, Moher D. How quickly do systematic reviews go out of date? A survival analysis. Ann Intern Med. 2007;147(4):224–233. Elliott JH, Synnot A, Turner T, et al. Living systematic review: 1. Introduction — the why, what, when, and how. J Clin Epidemiol. 2017;91:23–30. Marshall IJ, Wallace BC. Toward systematic review automation: a practical guide to using machine learning tools in research synthesis. Syst Rev. 2019;8:163. Boutron I, Chaimani A, Meerpohl JJ, et al. The COVID-NMA project: building an evidence ecosystem for the COVID-19 pandemic. Ann Intern Med. 2020;173(12):1015–1017. van de Schoot R, de Bruin J, Schram R, et al. An open source machine learning framework for efficient and transparent systematic reviews. Nat Mach Intell. 2021;3(2):125–133. Marshall IJ, Kuiper J, Wallace BC. RobotReviewer: evaluation of a system for automatically assessing bias in clinical trials. J Am Med Inform Assoc. 2016;23(1):193–201. Guo E, Gupta M, Deng J, Park YJ, Paget M, Naugler C. Automated paper screening for clinical reviews using large language models: data analysis study. J Med Internet Res. 2024;26:e48996. Khraisha Q, Put S, Kappenberg J, Warraitch A, Hadfield K. Can large language models replace humans in systematic reviews? Evaluating GPT-4’s efficacy in screening and extracting data from peer-reviewed and grey literature in multiple languages. Res Synth Methods. 2024;15(4):616–626. Syriani E, David I, Kumar G. Screening articles for systematic reviews with ChatGPT. J Comput Lang. 2024;80:101287. Khan MA, Ayub U, Naqvi SAA, et al. Collaborative large language models for automated data extraction in living systematic reviews. J Am Med Inform Assoc. 2025;32(4):638–647. Lai H, Ge L, Sun M, et al. Assessing the risk of bias in randomized clinical trials with large language models. JAMA Netw Open. 2024;7(5):e2412687. Hasan B, Saadi S, Rajjoub NS, et al. Integrating large language models in systematic reviews: a framework and case study using ROBINS-I for risk of bias assessment. BMJ Evid Based Med. 2024;29(6):394–398. Chen M, Ji M, Si X. The effects of statin therapy on mortality in patients with sepsis: A meta-analysis of randomized trials. Medicine (Baltimore). 2018;97(31):e11578. Sterne JAC, Savović J, Page MJ, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ. 2019;366:l4898. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials. 1986;7(3):177–188. Higgins JPT, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ. 2003;327(7414):557–560. Landis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977;33(1):159–174. Buscemi N, Hartling L, Vandermeer B, Tjosvold L, Klassen TP. Single data extraction generated more errors than double data extraction in systematic reviews. J Clin Epidemiol. 2006;59(7):697–703. Additional Declarations No competing interests reported. Supplementary Files Additionalfile1supplementarymaterials.docx Additionalfile2PRISMAchecklist.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9308492","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Method Article","associatedPublications":[],"authors":[{"id":623657131,"identity":"4693c705-77f7-46be-b770-77bc9248ea0f","order_by":0,"name":"Hayden Farquhar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCklEQVRIiWNgGAWjYBACAyCWAGJ+EIcZiOXYGJgbGBjYCGuRbIBqMWZjYCRRS2IDIS3m7GcMb3zcwSDBP/vwNumCmnvpfewHGx/zlDHI84sdwKrFsifH2HLmGQYJiXNpZdIzjhXntvEkNhvznGMwnDk7AbvDDuSYSfO2MdQxnOExk+ZhS8htk2BsA4kkGNzGoeX8GzPpv20MEvJgLf8S0tkIarkBtIURqMUApIW3LSGBCC3Pii172yQkDM+wFVvz9iUYgvxiOOecBG6/nE/eeONnm42E3Bnmjbd5viXIy7cfPvjgTZmNPL80di0MDBywqAHHERxI4FAOAuwP4DbiUTUKRsEoGAUjGQAAlPtS+Lqtq/cAAAAASUVORK5CYII=","orcid":"","institution":"Independent Researcher","correspondingAuthor":true,"prefix":"","firstName":"Hayden","middleName":"","lastName":"Farquhar","suffix":""}],"badges":[],"createdAt":"2026-04-03 03:54:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9308492/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9308492/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107707620,"identity":"d9c399ed-43d8-47a9-8ab2-f73560605f65","added_by":"auto","created_at":"2026-04-24 09:20:45","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":178846,"visible":true,"origin":"","legend":"\u003cp\u003ePipeline architecture diagram.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/481ac6410754e57b1dc7a422.png"},{"id":107675352,"identity":"6ce57a22-847c-4f51-bfe4-7221691c08c3","added_by":"auto","created_at":"2026-04-24 00:42:25","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":334075,"visible":true,"origin":"","legend":"\u003cp\u003ePRISMA flow diagram.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/76bc4674cac8247fee54df35.png"},{"id":107675353,"identity":"7613bc53-07be-4884-bf20-4f743e85cada","added_by":"auto","created_at":"2026-04-24 00:42:25","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":161073,"visible":true,"origin":"","legend":"\u003cp\u003eForest plot of baseline meta-analysis.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/44bc654092638746d7103154.png"},{"id":107707747,"identity":"697b2281-74f6-456b-9681-4fdc3d1019ed","added_by":"auto","created_at":"2026-04-24 09:21:03","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":114525,"visible":true,"origin":"","legend":"\u003cp\u003eLeave-one-out sensitivity analysis confirmed that the pooled estimate was robust to the exclusion of any single study. Subgroup analysis by statin type showed no significant effect for either atorvastatin (RR 0.87, 95% CI 0.52–1.44; k = 2) or simvastatin (RR 1.02, 95% CI 0.67–1.53; k = 3).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/b3e1335d39f69ea43e73c588.png"},{"id":107675355,"identity":"360c839f-00c6-4856-bb1c-c2b18b9d24fe","added_by":"auto","created_at":"2026-04-24 00:42:25","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":175821,"visible":true,"origin":"","legend":"\u003cp\u003eUpdated forest plot with Rashid 2025.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/18b350f525bdc9fe7e813795.png"},{"id":108685578,"identity":"07f17ec4-6813-4519-a5bd-69b2e8cbb63e","added_by":"auto","created_at":"2026-05-07 09:58:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1168062,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/3a9c0315-f69e-4aa1-ada8-214e6d77aa62.pdf"},{"id":107675349,"identity":"68269897-8118-4348-9d9f-85a39a50705e","added_by":"auto","created_at":"2026-04-24 00:42:25","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":14562,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile1supplementarymaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/158231e4cc951b73df735717.docx"},{"id":107675350,"identity":"4eeefa13-def7-4e76-906b-4b060341caa6","added_by":"auto","created_at":"2026-04-24 00:42:25","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":12338,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile2PRISMAchecklist.docx","url":"https://assets-eu.researchsquare.com/files/rs-9308492/v1/fe8613db5a864b262b7e9a5b.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Living Systematic Review Engine: LLM-Automated Evidence Surveillance Validated Against a Published Meta-Analysis of Statins for Sepsis","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSystematic reviews are a foundation of evidence-based medicine [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], yet the time required to produce them has become a barrier to their usefulness. An analysis of the PROSPERO registry found that systematic reviews take a mean of 67.3 weeks to complete, with screening and data extraction consuming the largest share of reviewer effort [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. By the time a review is published, the evidence it summarises may already be outdated: Shojania et al. found that 7% of reviews already had a signal for updating at the time of publication, rising to 23% within two years [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eLiving systematic reviews \u0026mdash; reviews that are continuously updated as new evidence emerges \u0026mdash; have been proposed as a solution [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. The concept is endorsed by Cochrane [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e], and several living review protocols have been registered. However, the practical burden of monthly or quarterly searching, screening, extraction, and re-analysis has limited adoption. Most existing living reviews rely on partial automation of a single stage, typically abstract screening, while the remaining workflow remains manual [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Even high-profile implementations such as the COVID-NMA living review consortium have required sustained manual effort to maintain currency [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAutomation tools for systematic reviews have existed for some time. ASReview uses active learning to prioritise screening [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], and RobotReviewer automates risk of bias assessment using machine learning [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The emergence of large language models (LLMs) has expanded what can be automated. LLMs have achieved sensitivities of 0.76\u0026ndash;1.00 for abstract screening in several evaluations [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], and recent work has demonstrated their use for data extraction [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] and risk of bias assessment [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. However, no published system integrates all stages \u0026mdash; from literature search through meta-analytic synthesis \u0026mdash; into a single automated pipeline with validated end-to-end performance.\u003c/p\u003e \u003cp\u003eIn this study, we present an open-source LLM-automated living systematic review engine and validate it against a published meta-analysis of RCTs of statins for sepsis [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. We chose this demonstration case because it has a well-defined reference standard, an active trial landscape, and genuine clinical equipoise \u0026mdash; making it a suitable test case for continuous updating.\u003c/p\u003e "},{"header":"Methods","content":"\u003cp\u003eStudy design\u003c/p\u003e \u003cp\u003eThis is a methods validation study comparing the outputs of an LLM-automated systematic review pipeline against a published human-conducted meta-analysis. The study involved no patient data and required no ethics approval. The embedded review of statins for sepsis was not prospectively registered on PROSPERO, as the primary objective was pipeline validation rather than a de novo review. The study is reported following the PRISMA 2020 guidelines [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e] where applicable to the automated review component (see Additional file 2: PRISMA Checklist).\u003c/p\u003e \u003cp\u003eReference standard\u003c/p\u003e \u003cp\u003eWe used the meta-analysis by Chen et al. (2018) as the primary reference standard [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. This review included 9 RCTs (2333 patients) examining statin therapy versus placebo or no statin in adults with sepsis, septic shock, or related critical illness. The pooled risk ratio for in-hospital mortality was 0.96 (95% CI 0.83\u0026ndash;1.11; I\u0026sup2; = 20%; k\u0026thinsp;=\u0026thinsp;7), and for 28-day mortality was 0.90 (95% CI 0.73\u0026ndash;1.11; I\u0026sup2; = 29%; k\u0026thinsp;=\u0026thinsp;5). The review concluded that statin therapy had no significant effect on mortality. Eight of the nine included studies had PubMed identifiers (PMIDs) and were used for validation; one study (El Gendy 2014) lacked a PMID and was excluded from PMID-based validation.\u003c/p\u003e \u003cp\u003ePipeline architecture\u003c/p\u003e \u003cp\u003eThe pipeline consists of four automated stages executed sequentially in Python (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e):\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eStage 1 \u0026mdash; Literature search.\u003c/b\u003e PubMed was searched via the NCBI E-utilities API using a Boolean query combining MeSH terms and free-text synonyms for statins (including individual drug names), sepsis and related conditions (septic shock, ARDS, ventilator-associated pneumonia, critical illness), and an RCT publication type filter. The search was designed for high sensitivity, deliberately broad to capture studies where sepsis-related conditions were secondary indications. Raw search results (XML) were archived for reproducibility.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStage 2 \u0026mdash; Title/abstract screening.\u003c/b\u003e Each retrieved article was screened against predefined inclusion and exclusion criteria using Claude Haiku 4.5 (Anthropic; model identifier in Supplementary Materials). Inclusion criteria required: (a) primary RCT or quasi-experimental design, (b) adult patients with sepsis or infection-related critical illness, (c) statin intervention of any type, (d) placebo or no-statin comparator, and (e) reporting of mortality, ICU length of stay, or organ dysfunction outcomes. Exclusion criteria included: non-primary studies (reviews, secondary analyses, post-hoc analyses), paediatric populations, non-infection indications, and animal studies. The LLM was instructed to apply a \u0026ldquo;when in doubt, include\u0026rdquo; principle, excluding only when confidence exceeded 0.95.\u003c/p\u003e \u003cp\u003eTo assess decision stability, each article was screened three times at temperature 0.3, and the final decision was determined by majority vote. If no clear majority existed, the article defaulted to \u0026ldquo;include.\u0026rdquo; All screening decisions, confidence scores, criterion-level assessments, and token usage were logged to a JSONL file for full reproducibility.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStage 3 \u0026mdash; Data extraction.\u003c/b\u003e Included studies underwent structured extraction using Claude Sonnet 4 (Anthropic; model identifier in Supplementary Materials) in a three-pass process: (a) initial PICO (Population, Intervention, Comparator, Outcome) and effect size extraction using a structured JSON prompt, (b) self-verification in which the LLM reviewed its own extraction against the source text with specific instructions to check for percentage-versus-count errors and arm size completeness, and (c) risk of bias assessment using the Cochrane RoB 2 framework [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Where available, full text was retrieved from PubMed Central (PMC) Open Access; otherwise, extraction was performed from the abstract alone. A title/author sanity check was implemented to detect and reject mismatched PMC articles. All extractions used temperature 0.0 for maximum determinism.\u003c/p\u003e \u003cp\u003e \u003cb\u003eStage 4 \u0026mdash; Meta-analytic synthesis.\u003c/b\u003e Pooled effect estimates were computed using the DerSimonian\u0026ndash;Laird random-effects model [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Risk ratios were calculated from extracted event counts with continuity correction (0.5) applied to zero-cell studies. Heterogeneity was assessed using Cochran\u0026rsquo;s Q test and the I\u0026sup2; statistic [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Prediction intervals were computed using the t-distribution with k\u0026thinsp;\u0026minus;\u0026thinsp;2 degrees of freedom. Publication bias was evaluated using Egger\u0026rsquo;s regression test and visual inspection of funnel plots, noting that both methods have limited reliability with fewer than 10 studies [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Sensitivity analyses included leave-one-out analysis and subgroup analysis by statin type.\u003c/p\u003e \u003cp\u003eLiving update demonstration\u003c/p\u003e \u003cp\u003eTo demonstrate the pipeline\u0026rsquo;s capacity for continuous evidence surveillance, we compared: (a) a baseline analysis using only studies available to the 2018 reference review, and (b) an updated analysis incorporating studies published after the reference search date that the pipeline independently identified and included.\u003c/p\u003e \u003cp\u003eValidation metrics\u003c/p\u003e \u003cp\u003e \u003cb\u003eScreening performance\u003c/b\u003e was evaluated by sensitivity (proportion of reference-included studies correctly identified), specificity (proportion of non-included studies correctly excluded), positive predictive value (PPV), negative predictive value (NPV), and Cohen\u0026rsquo;s kappa against the reference inclusion decisions. Inter-rater agreement between screening runs was assessed using pairwise Cohen\u0026rsquo;s kappa.\u003c/p\u003e \u003cp\u003e \u003cb\u003eExtraction accuracy\u003c/b\u003e was evaluated by comparing LLM-extracted mortality event counts against ground truth values from the reference review. Timepoint-aware matching was used: if the LLM extracted 90-day mortality, it was compared against 90-day ground truth where available. An exact match required identical event counts; a \u0026ldquo;within 5%\u0026rdquo; match allowed for rounding from percentage conversion. Given the small number of studies, exact binomial 95% confidence intervals were computed for accuracy proportions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eSynthesis concordance\u003c/b\u003e was assessed by comparing the direction, magnitude, and statistical significance of pooled effect estimates, as well as heterogeneity statistics.\u003c/p\u003e \u003cp\u003eContinuous deployment as a living review\u003c/p\u003e \u003cp\u003eAn important distinction exists between a pipeline that \u003cem\u003ecan\u003c/em\u003e be re-run and a review that \u003cem\u003eis\u003c/em\u003e continuously updated. To bridge this gap, we deployed the pipeline as a GitHub Actions scheduled workflow that executes automatically on the 1st of each month. Each cycle performs the full pipeline (search, screen, extract, synthesise), commits updated results (forest plot, meta-analysis, extraction data) to the repository, and opens a GitHub issue if new studies are detected or the evidence base changes. This provides a transparent, version-controlled audit trail of every update cycle.\u003c/p\u003e \u003cp\u003eThe deployment requires two API keys (NCBI and Anthropic) stored as repository secrets, and no dedicated infrastructure \u0026mdash; the workflow runs on GitHub\u0026rsquo;s free-tier runners. Monthly execution cost is approximately US\u003cspan\u003e$\u003c/span\u003e1.70 per cycle, dominated by screening API calls. Full deployment instructions are provided in the Additional file 1 (DEPLOYMENT.md) to enable readers to fork the repository and deploy a living review on any clinical topic by modifying only the search query, screening criteria, and extraction prompts.\u003c/p\u003e \u003cp\u003eSoftware and reproducibility\u003c/p\u003e \u003cp\u003eThe pipeline was implemented in Python 3.14 using Pydantic for data modelling, httpx for asynchronous API calls, SciPy for statistical computations, and Matplotlib for figure generation. All API calls (prompts, responses, token counts, costs) were logged. Source code is available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/hayden-farquhar/Living-Systematic-Review\u003c/span\u003e\u003cspan address=\"https://github.com/hayden-farquhar/Living-Systematic-Review\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIn this study, we used Anthropic\u0026rsquo;s Claude Haiku 4.5 for screening and Claude Sonnet 4 for extraction. However, the pipeline architecture is model-agnostic: the LLM is accessed through a single abstraction layer, and any instruction-following model with structured output capability can be substituted. As of April 2026, suitable alternatives include OpenAI\u0026rsquo;s GPT-4o and GPT-4o-mini, Google\u0026rsquo;s Gemini 2.5 Pro and Flash, Meta\u0026rsquo;s Llama 4 (via API providers or local deployment), and Mistral Large. The choice of model involves a trade-off between cost, accuracy, and speed \u0026mdash; smaller models suit high-volume screening, while larger models with stronger reasoning are preferred for extraction. We encourage future work comparing pipeline performance across model providers.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eLiterature search\u003c/p\u003e \u003cp\u003eThe PubMed search retrieved 215 unique articles (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAll 8 reference studies with known PMIDs were captured (100% recall). The search was intentionally broad, including terms for ARDS and ventilator-associated pneumonia to capture studies where the sepsis population overlapped with these conditions.\u003c/p\u003e \u003cp\u003eScreening performance\u003c/p\u003e \u003cp\u003eOf 215 articles screened, 12 were included (8 reference studies plus 4 additional) and 203 were excluded (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Screening sensitivity was 1.00 (8/8 reference studies correctly identified) and specificity was 0.98 (203/207). Cohen\u0026rsquo;s kappa against the reference was 0.79, indicating substantial agreement [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. NPV was 1.00 \u0026mdash; no reference studies were missed.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eScreening Performance Against Reference Standard\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArticles searched\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e215\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eArticles included by LLM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReference studies identified\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8/8 (100%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive predictive value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative predictive value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCohen\u0026rsquo;s kappa (vs reference)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInter-rater kappa (between runs)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse negatives\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse positives\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eScreening model\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eClaude Haiku 4.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRuns per article\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (majority vote)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTemperature\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal API calls\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e645\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUS\u003cspan\u003e$\u003c/span\u003e1.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e*Two of four false positives were post-2018 RCTs not available to the reference review (Rashid 2025, Sapey 2019); one was a long-term follow-up of an included trial (Agus 2017); one was a mechanistic study with a small embedded RCT (Ghosh 2015).\u003c/p\u003e \u003cp\u003eThe four false positives included two studies published after the reference search date (Rashid et al. 2025, a primary RCT; Sapey et al. 2019, a pilot RCT), one long-term follow-up of an included trial (Agus et al. 2017, HARP-2 12-month data), and one mechanistic study with a small embedded RCT component (Ghosh et al. 2015). Two of these represent genuine new evidence rather than true errors.\u003c/p\u003e \u003cp\u003eInter-rater agreement across the three screening runs was perfect (pairwise kappa\u0026thinsp;=\u0026thinsp;1.00 for all pairs), with 100% unanimous agreement (215/215 articles). This reflects the near-deterministic behaviour of the LLM at temperature 0.3.\u003c/p\u003e \u003cp\u003eTotal screening cost was US\u003cspan\u003e$\u003c/span\u003e1.70 (645 API calls; 1,091,499 input tokens; 207,202 output tokens).\u003c/p\u003e \u003cp\u003eData extraction accuracy\u003c/p\u003e \u003cp\u003eOf 8 studies extracted, 5 yielded mortality event counts suitable for meta-analysis. Three studies lacked extractable data: two because their abstracts reported only \u0026ldquo;no significant difference\u0026rdquo; without numeric event counts (Patel 2012, Truwit 2014), and one because the PMC full-text link pointed to an incorrect article and the abstract reported only percentages without arm-level denominators (Singh 2017). The full-text sanity check correctly rejected the two mismatched PMC articles.\u003c/p\u003e \u003cp\u003eAmong 7 studies with ground truth mortality data available for comparison, 4 (57%, 95% CI 18\u0026ndash;90%) had exact matches on event counts (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The 3 discrepancies were: (a) Truwit 2014 (SAILS trial) \u0026mdash; the abstract reported percentages (28.5% vs 24.9%) but not the per-arm sample sizes (379 vs 366), preventing accurate count derivation; (b) Papazian 2013 \u0026mdash; the LLM correctly converted percentages to counts but assumed equal arm sizes (150/150) when the actual split was 146/138; and (c) Patel 2012 \u0026mdash; the abstract contained no numeric mortality data. All three discrepancies were attributable to information absence in the abstract rather than extraction errors by the LLM. When the source text contained explicit event counts, extraction accuracy was 4/4 (100%).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eData Extraction Accuracy\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eYear\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSource\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMortality (LLM)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMortality (Reference)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTimepoint\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMatch\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNovack\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0/42 vs 0/41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0/42 vs 0/41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eHospital\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eExact\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKruger\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6/75 vs 4/75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6/75 vs 4/75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eHospital\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eExact\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2012\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;/49 vs \u0026mdash;/51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2/49 vs 2/51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e28-day\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMissing*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKruger\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18/123 vs 24/127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e18/123 vs 24/127\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e90-day\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eExact\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePapazian\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2013\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e32/150 vs 23/150\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e31/146 vs 21/138\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e28-day\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eClose**\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMcAuley\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e57/259 vs 75/281\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e57/259 vs 75/281\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e28-day\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eExact\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTruwit\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2014\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract***\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;/\u0026mdash; vs \u0026mdash;/\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e108/379 vs 91/366\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eHospital\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMissing*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSingh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAbstract***\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026mdash;/\u0026mdash; vs \u0026mdash;/\u0026mdash;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e28-day\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eN/A\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eExact match: 4/7 (57%, 95% CI 18\u0026ndash;90%). \u003cem\u003eAbstract did not report numeric event counts.\u003c/em\u003e \u003cb\u003eLLM assumed equal arm sizes (150/150); actual split was 146/138.\u003c/b\u003e PMC full-text link pointed to incorrect article; sanity check reverted to abstract. Extraction model: Claude Sonnet 4. Method: two-pass (initial extraction\u0026thinsp;+\u0026thinsp;self-verification)\u0026thinsp;+\u0026thinsp;RoB 2 assessment. Cost: US\u003cspan\u003e$\u003c/span\u003e0.25.\u003c/p\u003e \u003cp\u003eSelf-verification (the second LLM pass) successfully corrected percentage-to-count conversion errors in preliminary extraction, improving accuracy from 2/7 to 4/7 exact matches.\u003c/p\u003e \u003cp\u003eTotal extraction cost was US\u003cspan\u003e$\u003c/span\u003e0.25 (24 API calls across 8 studies).\u003c/p\u003e \u003cp\u003eMeta-analytic synthesis\u003c/p\u003e \u003cp\u003eThe LLM-derived pooled risk ratio from 5 studies with extractable mortality data was 0.94 (95% CI 0.74\u0026ndash;1.21; I\u0026sup2; = 6.0%; τ\u0026sup2; = 0.006; Cochran\u0026rsquo;s Q\u0026thinsp;=\u0026thinsp;4.26, p\u0026thinsp;=\u0026thinsp;0.37), compared with the reference estimate of 0.96 (95% CI 0.83\u0026ndash;1.11; I\u0026sup2; = 20%; k\u0026thinsp;=\u0026thinsp;7) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBoth analyses reached the same conclusion: statin therapy does not significantly reduce mortality in sepsis.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMeta-Analytic Synthesis \u0026mdash; LLM Pipeline vs Reference\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLLM Pipeline\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eChen et al.\u0026nbsp;2018 (Reference)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudies contributing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePooled RR (95% CI)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.94 (0.74\u0026ndash;1.21)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.96 (0.83\u0026ndash;1.11)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eI\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eτ\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.006\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNot reported\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCochran\u0026rsquo;s Q (p-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.26 (p\u0026thinsp;=\u0026thinsp;0.37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNot reported\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEgger\u0026rsquo;s test p-value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.46*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNot reported\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConclusion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNo significant effect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNo significant effect\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e*Egger\u0026rsquo;s test has limited reliability with fewer than 10 studies.\u003c/p\u003e \u003cp\u003eThe wider confidence interval in the LLM-derived estimate reflects the smaller number of contributing studies (5 vs 7), a direct consequence of abstract-only access limiting data extraction. Egger\u0026rsquo;s test showed no evidence of publication bias (p\u0026thinsp;=\u0026thinsp;0.46), though this test has limited power with fewer than 10 studies [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] (Fig.\u0026nbsp;4).\u003c/p\u003e \u003cp\u003e Leave-one-out sensitivity analysis confirmed that the pooled estimate was robust to the exclusion of any single study. Subgroup analysis by statin type showed no significant effect for either atorvastatin (RR 0.87, 95% CI 0.52\u0026ndash;1.44; k\u0026thinsp;=\u0026thinsp;2) or simvastatin (RR 1.02, 95% CI 0.67\u0026ndash;1.53; k\u0026thinsp;=\u0026thinsp;3).\u003c/p\u003e \u003cp\u003eLiving update demonstration\u003c/p\u003e \u003cp\u003eThe pipeline independently identified one new primary RCT published in 2025 (Rashid et al.), an open-label trial of atorvastatin 20 mg versus standard care in 68 adults with sepsis. This study reported 28-day mortality of 36% (13/36) in the atorvastatin group versus 56% (18/32) in the control group (p\u0026thinsp;=\u0026thinsp;0.10). Incorporating this study updated the pooled estimate to RR 0.89 (95% CI 0.70\u0026ndash;1.14; k\u0026thinsp;=\u0026thinsp;6; I\u0026sup2; = 15%) without changing the overall conclusion (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e, Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eLiving Update \u0026mdash; Cumulative Meta-Analysis\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnalysis\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ek\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePooled RR\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eI\u0026sup2;\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eConclusion\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChen et al.\u0026nbsp;2018 (reference)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.83\u0026ndash;1.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo significant effect\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLLM pipeline (baseline)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u0026ndash;1.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo significant effect\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e+ Rashid 2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u0026ndash;1.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e15%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo significant effect\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e+ Rashid 2025\u0026thinsp;+\u0026thinsp;Agus 2017 (12mo)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u0026ndash;1.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo significant effect*\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e*Upper confidence bound approaching 1.0; continued monitoring warranted.\u003c/p\u003e \u003cp\u003eFurther incorporating the 12-month mortality data from the HARP-2 follow-up (Agus et al. 2017; 31.8% vs 37.3%) yielded a pooled RR of 0.87 (95% CI 0.74\u0026ndash;1.01; k\u0026thinsp;=\u0026thinsp;7; I\u0026sup2; = 0%). While the upper confidence bound approached 1.0, the conclusion of no statistically significant benefit remained unchanged. This progressive narrowing of the confidence interval with cumulative evidence illustrates the value of continuous monitoring.\u003c/p\u003e \u003cp\u003eThe pipeline also detected a pilot RCT (Sapey et al. 2019) of high-dose simvastatin in community-acquired pneumonia with sepsis, though this study did not report extractable mortality event counts.\u003c/p\u003e \u003cp\u003eTime and cost\u003c/p\u003e \u003cp\u003eTotal pipeline execution time was approximately 32 minutes, dominated by the screening stage (25 minutes for 645 sequential API calls). Total API cost was US\u003cspan\u003e$\u003c/span\u003e1.95 (screening: US\u003cspan\u003e$\u003c/span\u003e1.70; extraction: US\u003cspan\u003e$\u003c/span\u003e0.25). By comparison, manual conduct of the equivalent review stages would require an estimated 58 hours of reviewer time at a cost of US\u003cspan\u003e$\u003c/span\u003e2,900\u0026ndash;5,800 for two trained reviewers (Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e) [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The pipeline achieves a roughly 110-fold reduction in time and greater than 99% reduction in direct costs. Monthly living update cycles, requiring re-execution of the search and screening stages only for new publications, would cost approximately US\u003cspan\u003e$\u003c/span\u003e1.70 per cycle.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTime and Cost Comparison\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026times;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStage\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHuman time (hours)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePipeline time (min)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePipeline cost (US\u003cspan\u003e$\u003c/span\u003e)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpeedup\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLiterature search\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e960\u0026times;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTitle/abstract screening\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e43\u0026times;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFull-text retrieval\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e240\u0026times;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData extraction\u0026thinsp;+\u0026thinsp;RoB\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e192\u0026times;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMeta-analysis\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e4800\u0026times;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReport generation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e2400\u0026times;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003e58\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e32\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003e1.95\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026times;\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e110\u0026times;\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eHuman estimates assume 2 trained reviewers. Hourly rate of US\u003cspan\u003e$\u003c/span\u003e50\u0026ndash;100 is based on typical research assistant and systematic reviewer compensation in academic settings [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Pipeline estimates: single end-to-end execution. Monthly living update (search\u0026thinsp;+\u0026thinsp;screening only): US\u003cspan\u003e$\u003c/span\u003e1.70/cycle (~\u0026thinsp;US\u003cspan\u003e$\u003c/span\u003e20/year).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003ePrincipal findings\u003c/p\u003e \u003cp\u003eWe demonstrate that an LLM-automated pipeline can replicate the core stages of a systematic review \u0026mdash; screening, data extraction, risk of bias assessment, and meta-analytic synthesis \u0026mdash; with sufficient fidelity to reach the same clinical conclusion as a human-conducted review, at a fraction of the time and cost. The pipeline\u0026rsquo;s independent detection of a post-reference RCT validates its utility for living evidence surveillance.\u003c/p\u003e \u003cp\u003eComparison with existing work\u003c/p\u003e \u003cp\u003ePrevious studies have demonstrated LLM performance on individual review stages. Abstract screening with GPT-4 has achieved sensitivities of 0.76\u0026ndash;1.00 depending on the review topic and prompt design [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Our finding of 1.00 sensitivity with 0.98 specificity is consistent with the upper range of reported performance. The perfect inter-rater agreement (kappa\u0026thinsp;=\u0026thinsp;1.00) at low temperature adds a dimension not previously reported, though it also reveals a limitation of the triple-run design: at temperature 0.3, the LLM is effectively deterministic, and the majority vote reduces to a single decision. A higher temperature setting would introduce genuine stochastic variation between runs, enabling meaningful inter-rater analysis \u0026mdash; but at the potential cost of screening accuracy. Future work should explore this trade-off.\u003c/p\u003e \u003cp\u003eData extraction using LLMs has been less extensively studied. Khan et al. (2025) reported concordant extraction accuracy of 0.94 using a collaborative two-LLM approach with cross-critique [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Our finding of 57% exact match on mortality event counts is lower, but the comparison is not straightforward: our denominator includes studies where the source text contained no extractable data. When the abstract or full text explicitly reported event counts, our extraction accuracy was 100% (4/4). The gap between these figures highlights a structural constraint \u0026mdash; abstract-only extraction has an inherent ceiling \u0026mdash; rather than a model capability limitation.\u003c/p\u003e \u003cp\u003eExisting SR automation tools address individual stages: ASReview handles screening prioritisation [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], RobotReviewer automates risk of bias assessment [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], and various tools assist with search strategy development [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. To our knowledge, this is the first validated pipeline that integrates automated search, screening, extraction, synthesis, and living update capability into a single system with quantified performance at each stage.\u003c/p\u003e \u003cp\u003eStrengths\u003c/p\u003e \u003cp\u003eThe pipeline logs every API call, prompt, response, and intermediate decision, enabling reproducibility at a granularity that manual reviews rarely achieve [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. The cost structure makes continuous evidence surveillance economically feasible for any clinical question. The validation design reports failures (extraction from abstracts) alongside successes (screening, synthesis concordance), rather than selecting only favourable metrics.\u003c/p\u003e \u003cp\u003eLimitations\u003c/p\u003e \u003cp\u003eExtraction accuracy is substantially constrained by full-text availability. When only abstracts are accessible, studies reporting mortality as percentages without arm-level denominators cannot be accurately converted to event counts. This affected 3 of 8 studies in our validation and represents the most significant barrier to fully automated extraction. Institutional full-text access or publisher API integration would likely improve this, but was beyond the scope of this study.\u003c/p\u003e \u003cp\u003eThe pipeline searches PubMed only; comprehensive systematic reviews typically search multiple databases (Embase, CENTRAL, Web of Science), grey literature, and trial registries [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The reference standard was a published meta-analysis rather than a prospectively conducted review, introducing the possibility that the reference itself contained errors. Validation was performed on a single clinical topic with a small number of included studies (8 with PMIDs); the 57% extraction accuracy has wide confidence intervals (95% CI 18\u0026ndash;90%), and generalisability to other review questions requires further evaluation.\u003c/p\u003e \u003cp\u003eThe pooled meta-analysis combined studies reporting mortality at different timepoints (hospital, 28-day, and 90-day), which introduces clinical heterogeneity that the I\u0026sup2; statistic may not fully capture. A human reviewer would typically standardise timepoints or conduct separate analyses by timepoint; the automated pipeline extracted whichever timepoint was most prominently reported.\u003c/p\u003e \u003cp\u003eLLM outputs are non-deterministic across model versions, and performance may vary as models are updated. The pipeline depends on commercial API services whose availability and pricing may change.\u003c/p\u003e \u003cp\u003eImplications\u003c/p\u003e \u003cp\u003eThe economics of living reviews change when the marginal cost of an update cycle falls below US\u003cspan\u003e$\u003c/span\u003e2.00. Organisations maintaining portfolios of living reviews could automate monthly surveillance across dozens of topics simultaneously. The pipeline\u0026rsquo;s detection of Rashid et al. (2025), a new RCT published seven years after the reference meta-analysis, demonstrates that automated surveillance can identify evidence that static reviews miss \u0026mdash; the failure mode that living reviews are designed to prevent. The automated pipeline logs every decision in a format that is difficult to achieve in manual review processes [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eBy providing a complete GitHub Actions workflow alongside an open-source, config-driven pipeline, readers can fork the repository, substitute their own PICO criteria and search terms, and have a living review running on any clinical question with no dedicated infrastructure. The annual cost of approximately US\u003cspan\u003e$\u003c/span\u003e20 for monthly updates makes this feasible for unfunded researchers and guideline groups in resource-limited settings. Sharing a deployable system \u0026mdash; not just code \u0026mdash; is essential for translating automation research into practice.\u003c/p\u003e \u003cp\u003eFuture directions\u003c/p\u003e \u003cp\u003eIntegration of additional databases (Embase, CENTRAL) and clinical trial registries (ClinicalTrials.gov) would improve search comprehensiveness. Institutional full-text access could substantially improve extraction accuracy. Parallel screening using multiple LLM providers would enable cross-model validation and reduce dependence on any single vendor. Evaluation across diverse clinical topics \u0026mdash; including reviews with larger numbers of included studies, continuous outcomes, and more complex eligibility criteria \u0026mdash; is needed to establish generalisability.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eAn LLM-automated living systematic review pipeline can replicate the findings of a human-conducted meta-analysis of statins for sepsis, achieving 100% screening sensitivity and concordant pooled effect estimates, in 32 minutes for US\u003cspan\u003e$\u003c/span\u003e1.95. The pipeline\u0026rsquo;s identification of a post-reference RCT demonstrates practical living review capability. While extraction accuracy remains limited by full-text access, the approach provides a validated, transparent, and economically viable foundation for continuous evidence synthesis, with potential for scaling across clinical topics.\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate:\u003c/strong\u003e \u003cp\u003eNo ethics approval was required. This study involved no patient data; all analyses used published literature.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication:\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests:\u003c/h2\u003e \u003cp\u003eThe author declares no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eNo external funding was received for this study.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eHF conceived the study, designed and implemented the pipeline, conducted all analyses, and wrote the manuscript. The author read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements:\u003c/h2\u003e \u003cp\u003eLarge language models (Anthropic Claude Haiku 4.5 and Claude Sonnet 4) were used as integral components of the pipeline under evaluation, as described in the Methods. Claude Code was used as a coding assistant during pipeline development. The author takes full responsibility for the content and accuracy of this manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eAll code, data, prompts, and API logs are available at https://github.com/hayden-farquhar/Living-Systematic-Review.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003ePage MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHiggins JPT, Thomas J, Chandler J, et al., editors. Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Cochrane; 2024. Available from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e\u003c/span\u003e\u003cspan address=\"http://www.training.cochrane.org/handbook\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBorah R, Brown AW, Capers PL, Kaiser KA. Analysis of the time and workers needed to conduct systematic reviews of medical interventions using data from the PROSPERO registry. BMJ Open. 2017;7(2):e012545.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShojania KG, Sampson M, Ansari MT, Ji J, Doucette S, Moher D. How quickly do systematic reviews go out of date? A survival analysis. Ann Intern Med. 2007;147(4):224\u0026ndash;233.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElliott JH, Synnot A, Turner T, et al. Living systematic review: 1. Introduction \u0026mdash; the why, what, when, and how. J Clin Epidemiol. 2017;91:23\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarshall IJ, Wallace BC. Toward systematic review automation: a practical guide to using machine learning tools in research synthesis. Syst Rev. 2019;8:163.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoutron I, Chaimani A, Meerpohl JJ, et al. The COVID-NMA project: building an evidence ecosystem for the COVID-19 pandemic. Ann Intern Med. 2020;173(12):1015\u0026ndash;1017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evan de Schoot R, de Bruin J, Schram R, et al. An open source machine learning framework for efficient and transparent systematic reviews. Nat Mach Intell. 2021;3(2):125\u0026ndash;133.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarshall IJ, Kuiper J, Wallace BC. RobotReviewer: evaluation of a system for automatically assessing bias in clinical trials. J Am Med Inform Assoc. 2016;23(1):193\u0026ndash;201.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuo E, Gupta M, Deng J, Park YJ, Paget M, Naugler C. Automated paper screening for clinical reviews using large language models: data analysis study. J Med Internet Res. 2024;26:e48996.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhraisha Q, Put S, Kappenberg J, Warraitch A, Hadfield K. Can large language models replace humans in systematic reviews? Evaluating GPT-4\u0026rsquo;s efficacy in screening and extracting data from peer-reviewed and grey literature in multiple languages. Res Synth Methods. 2024;15(4):616\u0026ndash;626.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSyriani E, David I, Kumar G. Screening articles for systematic reviews with ChatGPT. J Comput Lang. 2024;80:101287.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhan MA, Ayub U, Naqvi SAA, et al. Collaborative large language models for automated data extraction in living systematic reviews. J Am Med Inform Assoc. 2025;32(4):638\u0026ndash;647.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLai H, Ge L, Sun M, et al. Assessing the risk of bias in randomized clinical trials with large language models. JAMA Netw Open. 2024;7(5):e2412687.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHasan B, Saadi S, Rajjoub NS, et al. Integrating large language models in systematic reviews: a framework and case study using ROBINS-I for risk of bias assessment. BMJ Evid Based Med. 2024;29(6):394\u0026ndash;398.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen M, Ji M, Si X. The effects of statin therapy on mortality in patients with sepsis: A meta-analysis of randomized trials. Medicine (Baltimore). 2018;97(31):e11578.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSterne JAC, Savović J, Page MJ, et al. RoB 2: a revised tool for assessing risk of bias in randomised trials. BMJ. 2019;366:l4898.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials. 1986;7(3):177\u0026ndash;188.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHiggins JPT, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ. 2003;327(7414):557\u0026ndash;560.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLandis JR, Koch GG. The measurement of observer agreement for categorical data. Biometrics. 1977;33(1):159\u0026ndash;174.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBuscemi N, Hartling L, Vandermeer B, Tjosvold L, Klassen TP. Single data extraction generated more errors than double data extraction in systematic reviews. J Clin Epidemiol. 2006;59(7):697\u0026ndash;703.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"living systematic review, large language models, evidence synthesis automation, meta-analysis, statins, sepsis","lastPublishedDoi":"10.21203/rs.3.rs-9308492/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9308492/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eSystematic reviews take a mean of 67 weeks to complete and are frequently outdated by publication. Living systematic reviews aim to address this through continuous updating, but the manual labour required makes them impractical at scale. Large language models (LLMs) offer the potential to automate the full review pipeline, yet most existing tools address only abstract screening.\u003c/p\u003e\u003ch2\u003eObjective\u003c/h2\u003e \u003cp\u003eTo develop and validate an end-to-end LLM-automated living systematic review engine that performs literature searching, title/abstract screening, structured data extraction, risk of bias assessment, and random-effects meta-analysis, and to demonstrate its capacity for continuous evidence surveillance.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe built a Python pipeline using PubMed E-utilities for automated searching, Claude Haiku for triple-run title/abstract screening with majority voting, and Claude Sonnet for two-pass data extraction with self-verification and Cochrane Risk of Bias 2 (RoB 2) assessment. Meta-analysis used the DerSimonian\u0026ndash;Laird random-effects estimator. The pipeline was validated against a published meta-analysis of statins for sepsis (Chen et al., 2018; 9 randomised controlled trials [RCTs], 2333 patients). Performance was evaluated on screening sensitivity/specificity, extraction accuracy against ground truth event counts, and concordance of pooled effect estimates. The living update capability was demonstrated by incorporating studies published after the reference search date. The study is reported following the PRISMA 2020 guidelines where applicable [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe PubMed search retrieved 215 candidate articles, capturing all 8 reference studies with known PubMed identifiers (100% recall). LLM screening achieved sensitivity of 1.00 (8/8 studies identified), specificity of 0.98 (203/207 correctly excluded), and Cohen\u0026rsquo;s kappa of 0.79 against the reference standard, with perfect inter-rater agreement (kappa\u0026thinsp;=\u0026thinsp;1.00) across screening runs. Data extraction achieved exact agreement on mortality event counts for 4 of 7 studies with available ground truth (57%, 95% CI 18\u0026ndash;90%); the remaining discrepancies were attributable to abstract-only access limitations rather than extraction failures. The LLM-derived pooled risk ratio (RR 0.94, 95% CI 0.74\u0026ndash;1.21; k\u0026thinsp;=\u0026thinsp;5; I\u0026sup2; = 6%) was concordant with the reference estimate (RR 0.96, 95% CI 0.83\u0026ndash;1.11; k\u0026thinsp;=\u0026thinsp;7; I\u0026sup2; = 20%), reaching the same clinical conclusion: no significant mortality benefit from statins in sepsis. The pipeline independently identified one new RCT published in 2025 (Rashid et al.) not available to the reference review. Incorporating this study updated the pooled estimate to RR 0.89 (95% CI 0.70\u0026ndash;1.14; k\u0026thinsp;=\u0026thinsp;6) without changing the overall conclusion. Total pipeline execution time was 32 minutes at a cost of US\u003cspan\u003e$\u003c/span\u003e1.95, compared with an estimated 58 hours and US\u003cspan\u003e$\u003c/span\u003e2,900\u0026ndash;5,800 for manual review.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eAn LLM-automated pipeline can replicate the screening, extraction, and synthesis stages of a systematic review with high fidelity, at a fraction of the time and cost. The pipeline\u0026rsquo;s detection of a post-reference RCT demonstrates practical living review capability. Key limitations include reduced extraction accuracy when full text is unavailable and restriction to a single database. The open-source pipeline provides a foundation for scalable, continuously updated evidence synthesis.\u003c/p\u003e\u003ch2\u003eRegistration\u003c/h2\u003e \u003cp\u003eThe embedded systematic review was not prospectively registered on PROSPERO, as the primary study objective was pipeline validation.\u003c/p\u003e","manuscriptTitle":"A Living Systematic Review Engine: LLM-Automated Evidence Surveillance Validated Against a Published Meta-Analysis of Statins for Sepsis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-24 00:42:21","doi":"10.21203/rs.3.rs-9308492/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f36f1663-78eb-42ee-a26f-95f5784f4976","owner":[],"postedDate":"April 24th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"Rejected","date":"2026-05-07T09:40:18+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-05-07T09:56:06+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-24 00:42:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9308492","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9308492","identity":"rs-9308492","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}