Meta-analysis is a cornerstone of evidence synthesis, yet challenges arise when studies report heterogeneous summary statistics, such as means and standard deviations (SDs) versus medians, interquartile ranges (IQRs), or other percentiles. Excluding studies that report only medians and IQRs can introduce bias and reduce precision, particularly when outcomes are skewed, which is common in clinical research. Although several methods exist to estimate means and SDs from alternative summaries, many rely on strong normality assumptions, exhibit computational burden, or fail to adequately account for the precision of reported quantiles (e.g., extreme values versus medians). To address these limitations, we propose two flexible weighted estimators for estimating the mean and SD from reported quantiles. The methods leverage inverse-variance and inverse–variance–covariance weighting, respectively, to enhance both accuracy and precision. Additionally, our methods are flexible enough to accommodate any set of reported quantiles and various underlying distributions, and they can be readily implemented using standard statistical software. Simulation studies demonstrate that the weighted estimators provide nearly unbiased estimates of the mean and SD with high precision in most cases, especially for large sample sizes. In a real-world meta-analysis, the estimates obtained using the proposed estimators closely aligned with those derived from true sample statistics. These approaches are particularly valuable for skewed outcomes and offer a practical and user-friendly solution for researchers seeking to integrate heterogeneous data while improving accuracy and precision.

Systematic reviews play a critical role in evidence-based research but are labor-intensive, especially during title and abstract screening. Compact large language models (LLMs) offer potential to automate this process, balancing time/cost requirements and accuracy. The aim of this study is to assess the feasibility, accuracy, and workload reduction by three compact LLMs (GPT-4o mini, Llama 3.1 8B, and Gemma 2 9B) in screening titles and abstracts. Records were sourced from three previously published systematic reviews and LLMs were requested to rate each record from 0 to 100 for inclusion, using a structured prompt. Predefined 25-, 50-, 75-rating thresholds were used to compute performance metrics (balanced accuracy, sensitivity, specificity, positive and negative predictive value, and workload-saving). Processing time and costs were registered. Across the systematic reviews, LLMs achieved high sensitivity (up to 100%) and low precision (below 10%) for records included by full text. Specificity and workload savings improved at higher thresholds, with the 50- and 75-rating thresholds offering optimal trade-offs. GPT-4o-mini, accessed via application programming interface, was the fastest model (~40 minutes max.) and had usage costs ($0.14–$1.93 per review). Llama 3.1-8B and Gemma 2-9B were run locally in longer times (~4 hours max.) and were free to use. LLMs were highly sensitive tools for the title/abstract screening process. High specificity values were reached, allowing for significant workload savings, at reasonable costs and processing time. Conversely, we found them to be imprecise. However, high sensitivity and workload reduction are key factors for their usage in the title/abstract screening phase of systematic reviews.

Overlap of primary studies among multiple systematic reviews (SRs) is a major challenge when conducting overviews. The corrected covered area (CCA) is a metric computed from a matrix of evidence that quantifies overlap. Therefore, the assumptions used to generate the matrix may significantly affect the CCA. We aim to explore how these varying assumptions influence CCA calculations. We searched two databases for intervention-focused overviews published during 2023. Two reviewers conducted study selection and data extraction. We extracted overview characteristics and methods to handle overlap. For seven sampled overviews, we calculated overall and pairwise CCA across 16 scenarios, representing four matrix-construction assumptions. Of 193 included overviews, only 23 (11.9%) adhered to an overview-specific reporting guideline (e.g. PRIOR). Eighty-five (44.0%) did not address overlap; 14 (7.3%) only mentioned it in the discussion; and 94 (48.7%) incorporated it into methods or results (38 using CCA). Among the seven sampled overviews, CCA values varied depending on matrix-construction assumptions, ranging from 1.2% to 13.5% with the overall method and 0.0% to 15.7% with the pairwise method. CCA values may vary depending on the assumptions made during matrix construction, including scope, treatment of structural missingness, and handling of publication threads. This variability calls into question the uncritical use of current CCA thresholds and underscores the need for overview authors to report both overall and pairwise CCA calculations. Our preliminary guidance for transparently reporting matrix-construction assumptions may improve the accuracy and reproducibility of CCA assessments.

The 2019 ICH E9(R1) addendum highlights the importance of estimands, including the specification of post-randomization events that may affect the interpretation of clinical trial outcomes (i.e., intercurrent events; ICEs) and strategies to handle these events. Compared to trial protocols, there is limited discussion of estimands in the context of evidence synthesis. We conducted a comprehensive review of the Cochrane Library for pairwise meta-analyses of immuno-, targeted, hormone, and other novel oncology therapies. Dates were restricted to 2021 and onwards to allow time for addendum adoption. Outcomes of interest were progression-free survival (PFS) and overall survival (OS). Information on treatment switching and analytic strategies to address treatment switching were extracted from each meta-analysis and the RCTs they included. Out of 162 oncology reviews published in the Cochrane Library since 2021, eight pairwise meta-analyses and 68 RCTs met selection criteria. Most RCTs were Phase 3 (68%; n = 46) and/or open-label (76%; n = 52). More than half of RCTs explicitly allowed switching (59%; n = 40), while more than one third (38%; n = 26) did not report on treatment switching. Among trials that allowed treatment switching, censoring mechanisms for treatment switching varied in analyses of PFS. No included RCTs censored OS at the time of treatment switching. Despite the high prevalence of treatment switching in included trials, none of the identified meta-analyses addressed treatment switching analytically. Poor reporting regarding treatment switching in the RCTs themselves hinders the utility of aggregate-level meta-analyses. To ensure accurate interpretation of meta-analytic results, improved reporting of ICEs and ICE handling strategies is needed.

In many areas of oncology, cancer drugs are now associated with long-term survivorship and mixture cure models (MCM) are increasingly being used for survival analysis. The objective of this article was to propose a methodology for conducting network meta-analysis (NMA) of MCM. This method was illustrated through a case study evaluating recurrence-free survival (RFS) with adjuvant therapy for stage III/IV resected melanoma. For the case study, the MCM NMA was conducted by: (1) fitting MCMs to each trial included within the network of evidence; and (2) incorporating the parameters of the MCMs into a multivariate NMA. Outputs included relative effect estimates for the MCM NMA as well as absolute estimates of survival (RFS), modeled within the Bayesian multivariate NMA, by incorporating absolute baseline effects of the reference treatment. The case study was intended for illustrative purposes of the MCM NMA methodology and is not meant for clinical interpretation. The case study demonstrated the feasibility of conducting an MCM NMA and highlighted key issues and considerations when conducting such analyses, including plausibility of cure, maturity of data, process for model selection, and the presentation and interpretation of results. MCM NMA provides a method of comparative survival that acknowledges the benefit newer treatments may confer on a subset of patients, resulting in long-term survival and reflection of this survival in extrapolation. In the future, this method may provide an additional metric to compare treatments that is of value to patients.

Effective data management is essential for tasks involving decisions based on data, including knowledge synthesis and literature reviews. Despite this, how to carry out data management in literature reviews effectively remains unclear. With the increasing volume of research papers and the expansion of computational techniques for processing data (e.g., machine learning or large language models), it becomes imperative to consider data management as a crucial element for the advancement of literature review practices and tools. Presently, there are shortcomings related to (1) handling the growth of research to be synthesized, (2) addressing data quality issues when applying computational techniques or facilitating the verification of content produced by generative artificial intelligence, (3) enabling efficient reuse of datasets and innovative recombination of tools, and (4) facilitating transparent collaboration across heterogeneous review teams. To address these shortcomings, we develop the C5-DM Framework with conceptual principles to address data management challenges across five areas relevant to literature reviews: data conceptualization, collection, curation, control, and consumption. Methodological guidance for researchers with respect to these five areas is necessary to reduce errors, save time on repetitive tasks, and allow review teams to develop insightful syntheses.

To examine the extent to which information sources other than journal articles are sought for systematic reviews. Cross-sectional study of published systematic reviews. We examined all published systematic reviews included in MEDLINE in a 4-week period in 2019. Both systematic reviews and protocols of reviews were eligible for inclusion. (1) Number and types of information sources sought in systematic reviews; (2) proportion of reviews that explicitly searched for study reports other than journal articles; (3) proportion of reviews that searched resources containing study reports other than journal articles. A total of 1,262 systematic reviews fulfilled the eligibility criteria. The median number of information resources searched for all systematic reviews was 4. Of the 1,262 reviews, study reports other than journal articles were sought in 40% (n = 502) of systematic reviews (97% (n = 64) of Cochrane reviews and 37% (n = 438) of non-Cochrane reviews). Trial registers were searched in 88% of Cochrane reviews and 21% of non-Cochrane reviews. In 99.3% (n = 1,253) of all the systematic reviews, the searches performed had the potential to identify study reports other than journal articles. Between a third and a half of systematic reviews search for study reports other than journal articles. Systematic review searches often search resources that include study reports other than journal articles, whether or not the reviewers explicitly sought them.

Network meta-analysis (NMA) facilitates the comparison of multiple treatments by integrating both direct and indirect evidence. Applications of NMA in medical decision making have grown exponentially. However, the validity of NMA findings depends on key assumptions: homogeneity, transitivity, and consistency. A lack of consistent assessment of these assumptions potentially compromises the reliability of NMA outcomes. The objective of this study is to evaluate the extent to which researchers address NMA assumptions and report the assessment of evidence certainty in NMA publications. A total of 22,079 studies were identified from PubMed, Embase, and Cochrane CENTRAL (January 2010–August 2024). A sample of 393 NMAs was calculated to represent this population and randomly selected. Data on study characteristics, NMA assumptions, and the certainty of evidence were extracted and analyzed descriptively. Of the 393 NMAs, 71.8% were published between 2020 and 2024. Homogeneity was assessed in 300 (76.3%) NMAs, transitivity in 45 (11.5%) NMAs, and consistency in 265 (67.4%) NMAs. The certainty of evidence was assessed in 110 (28.0%) studies, predominantly using GRADE (71 NMAs; 18.1%) or CINeMA (29 NMAs; 7.4%). NMAs published in journals with high-impact factors more frequently evaluate these aspects than those published in low-impact journals. The assessment of NMA assumptions is inconsistently reported across studies, particularly for transitivity and consistency assumptions. Our findings highlight the need for standardized protocols or reporting guidelines to ensure these assessments are conducted and transparently reported.

Systematic reviews (SRs) are critical for evidence-based research but are time-consuming and labor-intensive. The rapid expansion of academic publications further challenges the performance and applicability of existing screening and classification methods. While large language models (LLMs) present new opportunities for automation, limited research has examined whether they can achieve classification performance comparable to human reviewers in large-scale, multi-class settings. With the goal of improving classification performance, we proposed an LLM-based framework that leverages full-text key-insight extraction to enhance literature classification. We constructed a manually curated dataset of 900 articles from 17 published SRs to quantitatively evaluate the classification capabilities of LLMs. The results provided empirical evidence of LLMs’ potential in supporting large-scale SRs and introduced a practical pathway for improving efficiency and reliability in evidence synthesis. Empirical results showed that key-insight-based classification (KBC) significantly outperforms abstract-based classification (ABC). We implemented a confidence-weighted voting (CWV) mechanism using multiple LLMs to improve robustness. The CWV method achieved the highest macro F1-score of 0.796, substantially exceeding KBC (0.732), ABC (0.676), and unsupervised K-means clustering (0.446). By employing zero-shot LLMs, our approach demonstrated the potential for enhanced adaptability across diverse domains and classification tasks without requiring fine-tuning, demonstrating that a carefully designed pipeline can enable LLMs to achieve classification performance comparable to human reviewers.

To explore the perceptions of, and barriers to, grey literature searching among medical researchers and journal editors. A cross-sectional survey of authors of systematic reviews and the editors of the journals in which the reviews were published. Systematic reviews indexed in MEDLINE, spanning a 4-week period in 2019. We excluded protocols. We asked whether the reviewers were performing a grey literature search. If they were, we asked about their approach to the grey literature search and relevant guidance. If they were not, we asked about their rationale for this. We elucidated understandings of grey literature from all reviewers. The survey to journal editors asked about their perceptions towards grey literature. A consecutive sample of 1,229 systematic reviews was included. A total of 155 authors responded, and a total of 46 journal editors responded. The majority (57%) of reviewers reported performing a grey literature search. However, there was no consensus on types of grey literature items or sources among the reviewers. The most frequent barrier to grey literature searches was concern about the quality/detail of the grey literature items. Editors expressed negative perceptions towards grey literature, rooted in suspicion of its quality. While many reviewers reported performing a grey literature search, there was a diverse understanding of the term grey literature. Concerns about the quality of grey literature exist among reviewers and editors alike. There is a marked discrepancy between the best-practice guidelines and the gatekeepers of journals.

Meta-analytic-predictive (MAP) priors have been proposed as a generic approach to deriving informative prior distributions, where external empirical data are processed to learn about certain parameter distributions. The use of MAP priors is also closely related to shrinkage estimation (also sometimes referred to as dynamic borrowing). A potentially odd situation arises when the external data consist only of a single study. Conceptually, this is not a problem, it only implies that certain prior assumptions gain in importance and need to be specified with particular care. We outline this important, not uncommon special case and demonstrate its implementation and interpretation based on the normal–normal hierarchical model. The approach is illustrated using example applications in clinical medicine.

Evaluation of clinical prediction models across multiple clusters, whether centers or datasets, is becoming increasingly common. A comprehensive evaluation includes an assessment of the agreement between the estimated risks and the observed outcomes, also known as calibration. Calibration is of utmost importance for clinical decision making with prediction models, and it often varies between clusters. We present three approaches to take clustering into account when evaluating calibration: (1) clustered group calibration (CG-C), (2) two-stage meta-analysis calibration (2MA-C), and (3) mixed model calibration (MIX-C), which can obtain flexible calibration plots with random effects modeling and provide confidence interval (CI) and prediction interval (PI). As a case example, we externally validate a model to estimate the risk that an ovarian tumor is malignant in multiple centers (N = 2489). We also conduct a simulation study and a synthetic data study generated from a true clustered dataset to evaluate the methods. In the simulation study, MIX-C and 2MA-C (splines) gave estimated curves closest to the true overall curve. In the synthetic data study, MIX-C produced cluster-specific curves closest to the truth. Coverage of the PI across the plot was best for 2MA-C with splines. We recommend using 2MA-C with splines to estimate the overall curve and 95% PI and MIX-C for cluster-specific curves, especially when the sample size per cluster is limited. We provide ready-to-use code to construct summary flexible calibration curves, with CI and PI to assess heterogeneity in calibration across datasets or centers.

The ICH E9(R1) addendum provides guidelines on accounting for intercurrent events in clinical trials using the estimands framework. However, there has been limited attention to the estimands framework for meta-analysis. Using treatment switching, a well-known intercurrent event that occurs frequently in oncology, we conducted a simulation study to explore the bias introduced by pooling together estimates targeting different estimands in a meta-analysis of randomized clinical trials (RCTs) that allowed treatment switching. We simulated overall survival data of a collection of RCTs that allowed patients in the control group to switch to the intervention treatment after disease progression under fixed effects and random effects models. For each RCT, we calculated effect estimates for a treatment policy estimand that ignored treatment switching, and a hypothetical estimand that accounted for treatment switching either by fitting rank-preserving structural failure time models or by censoring switchers. Then, we performed random effects and fixed effects meta-analyses to pool together RCT effect estimates while varying the proportions of trials providing treatment policy and hypothetical effect estimates. We compared the results of meta-analyses that pooled different types of effect estimates with those that pooled only treatment policy or hypothetical estimates. We found that pooling estimates targeting different estimands results in pooled estimators that do not target any estimand of interest, and that pooling estimates of varying estimands can generate misleading results, even under a random effects model. Adopting the estimands framework for meta-analysis may improve alignment between meta-analytic results and the clinical research question of interest.

Meta-research and evidence synthesis require considerable resources. Large language models (LLMs) have emerged as promising tools to assist in these processes, yet their performance varies across models, limiting their reliability. Taking advantage of the large availability of small size (<10 billion parameters) open-source LLMs, we implemented an agreement-based framework in which a decision is taken only if at least a given number of LLMs produce the same response. The decision is otherwise withheld. This approach was tested on 1020 abstracts of randomized controlled trials in rheumatology, using 2 classic literature review tasks: (1) classifying each intervention as drug or nondrug based on text interpretation and (2) extracting the total number of randomized patients, a task that sometimes required calculations. Re-examining abstracts where at least 4 LLMs disagreed with the human gold standard (dual review with adjudication) allowed constructing an improved gold standard. Compared to a human gold standard and single large LLMs (>70 billion parameters), our framework demonstrated robust performance: several model combinations achieved accuracies above 95% exceeding the human gold standard on at least 85% of abstracts (e.g., 3 of 5 models, 4 of 6 models, or 5 of 7 models). Performance variability across individual models was not an issue, as low-performing models contributed fewer accepted decisions. This agreement-based framework offers a scalable solution that can replace human reviewers for most abstracts, reserving human expertise for more complex cases. Such frameworks could significantly reduce the manual burden in systematic reviews while maintaining high accuracy and reproducibility.

The exponential growth of scientific literature poses increasing challenges for evidence synthesis. Systematic reviews (SRs) usually rely on keyword-based database searches, which are limited by inconsistent terminology and indexing delays. Citation searching—identifying studies that cite or are cited by known relevant articles—offers a complementary route to uncover additional evidence but remains poorly automated and integrated into screening workflows. We developed BibliZap, an open-source, fully automated citation-searching tool built on Lens.org data, performing multi-level forward and backward citation searches with relevance-based ranking. Its performance was evaluated across 66 published SRs, comparing five approaches: (1) PubMed-only searches; (2) PubMed followed by BibliZap restricted to the top 500 ranked results; (3) PubMed followed by full BibliZap screening; and (4–5) two exploratory early-stop strategies where BibliZap was initiated after identifying the first or the first three PubMed relevant records. The primary outcome was sensitivity, with secondary assessments of screening workload and precision. When used after PubMed screening, BibliZap increased mean sensitivity from 75% to 97%, achieving complete recall in over half of the reviews. Screening only the top 500 outputs still allowed over 90% of reviews to reach or exceed 80% recall. BibliZap recovered a median of three additional included articles per review, not retrieved by PubMed, while adding a median of 6,450 additional records. Citation searching via BibliZap enhances the completeness of evidence retrieval in SRs based on restricted database searches and supports transparent, scalable workflows adaptable to rapid and exploratory review contexts.

Network meta-analysis (NMA) is an increasingly popular evidence synthesis tool that can provide a ranking of competing treatments, also known as a treatment hierarchy. Treatment–covariate interactions (TCIs) can be included in NMA models to allow relative treatment effects to vary with covariate values. We show that in an NMA model that includes TCIs, treatment hierarchies should be created with a particular covariate profile in mind. We outline the typical approach for creating a treatment hierarchy in standard Bayesian NMA and show how a treatment hierarchy for a particular covariate profile can be created from an NMA model that estimates TCIs. We demonstrate our methods using a real network of studies for the treatment of major depressive disorder.

Categorical moderators are often found in meta-analysis and examined using meta-regression models. When multiple effect sizes are present within studies, several methods can be used for meta-regression: multivariate models, three-level models, correlated-effects models with robust variance estimation (RVE), three-level models with RVE, and correlated-effects models with RVE and cluster wild bootstrapping (CWB). This study aimed to compare the performance of these methods through a simulation study. Cohen’s d values were generated under a multivariate model, incorporating a binary variable that could represent either study-level or effect size-level characteristics. When the moderator referred to an effect size-level characteristic, its effect was allowed to vary across studies. Factors manipulated in the simulation included number of studies, number of outcomes per study, and the distribution of effect sizes across the categories of the moderator variable, ranging from balanced to highly unbalanced. The methods were applied and compared in terms of bias, Type I error, and power. The results showed that all methods exhibited lower power to detect effects when the moderator variable referred to study-level characteristics and the effect size distribution was very unbalanced. Methods based on RVE (correlated-effects with RVE or with RVE and CWB, and three-level models with RVE) effectively controlled Type I error rates but tended to be overconservative. In contrast, three-level models achieved higher power but at the cost of inflated Type I error. The best balance between Type I error control and power was observed when using a combination of three-level models and RVE.

Accurate assessment of adverse event (AE) incidence is critical in clinical research for drug safety. While meta-analysis serves as an essential tool to comprehensively synthesize the evidence across multiple studies, incomplete AE reporting in clinical trials remains a persistent challenge. In particular, AEs occurring below study-specific reporting thresholds are often omitted from publications, leading to left-censored data. Failure to account for these censored AE counts can result in biased AE incidence estimates. We present an R Shiny application that implements a Bayesian meta-analysis model specifically designed to incorporate censored AE data into the estimation process. This interactive tool provides a user-friendly interface for researchers to conduct AE meta-analyses and estimate the AE incidence probability using an unbiased approach. It also enables direct comparisons between models that either incorporate or ignore censoring, highlighting the biases introduced by conventional approaches. This tutorial demonstrates the Shiny application’s functionality through an illustrative example on meta-analysis of PD-1/PD-L1 inhibitor safety and highlights the importance of this tool in improving AE risk assessment. Ultimately, the new Shiny app facilitates more accurate and transparent drug safety evaluations. The Shiny-MAGEC app is available at: https://zihanzhou98.shinyapps.io/Shiny-MAGEC/.

Bayesian hypothesis testing via Bayes factors offers a principled alternative to classical p-value methods in meta-analysis, particularly suited to its cumulative and sequential nature. Unlike p-values, Bayes factors allow for quantifying support both for and against the existence of an effect, facilitate ongoing evidence monitoring, and maintain coherent long-run behavior as additional studies are incorporated. Recent theoretical developments further show how Bayes factors can flexibly control Type I error rates through connections to e-value theory. Despite these advantages, their use remains limited in the meta-analytic literature. This article provides a critical overview of their theoretical properties, methodological considerations—such as prior sensitivity—and practical advantages for evidence synthesis. Two illustrative applications are provided: one on statistical learning in individuals with language impairments, and another on seroma incidence following post-operative exercise in breast cancer patients. New tools supporting these methods are available in the open-source R package BFpack.

To understand a treatment’s potential impact at the individual level, it is crucial to explore whether the effect differs across patient subgroups and covariate values. Meta-analysis provides an important tool for detecting treatment–covariate interactions, as it can improve power compared to a single study. However, aggregation bias can occur when estimating individual-level treatment–covariate interactions in meta-analysis, due to trial-level confounding. This refers to when the association between the covariate and treatment effect across trials (at the aggregate level) differs from that observed within trials (at the individual level). It is, thus, recommended that heterogeneity in the treatment effect at the individual level should be disentangled from that at the trial level, ideally using an individual participant data (IPD) meta-analysis. Here, we explain this issue and provide new intuition about how trial-level confounding is impacted by differences in within-trial distributions of covariates and how this corresponds to asymmetry in subgroup-specific funnel plots in the case of categorical covariates. We then propose a sensitivity analysis to assess the robustness of interaction estimates to potential trial-level confounding. We illustrate these concepts using simulated and real data from an IPD meta-analysis of trials conducted on the TICO/ACTIV-3 platform, which assessed passive immunotherapy treatments for inpatients with COVID-19.