Status of this document

This is the draft protocol, written before any screening, data extraction or analysis. Items marked TO BE COMPLETED need a decision or information from the review team before registration. After registration the protocol is frozen; every later change is recorded in protocol/amendments.csv with its date, rationale and whether it was made before or after any results for the affected outcome were seen.

The machine-readable parts of the protocol live next to this file and are authoritative where the text summarizes them: populations.csv, outcomes.csv, nodes.csv, node_classes.csv, effect_modifiers.csv, amendments.csv and ../search/strategies/concepts.yaml.

Administrative information

Registration. The protocol will be registered in PROSPERO and deposited on OSF before title and abstract screening begins. Registration numbers: TO BE COMPLETED.

Review type. Living systematic review with network meta-analysis (LSR NMA) of randomized trials, with a separate living synthesis of nonrandomized studies for harms, and an inventory of all trials, observational studies, registrations and published systematic reviews of GLP-1 receptor agonist containing drugs.

Reporting guidance. The protocol follows PRISMA-P 2015 (moher2015prismap; shamseer2015prismap). Reports will follow PRISMA 2020 (page2021prisma), the PRISMA extension for network meta-analyses (hutton2015prismanma), the PRISMA extension for living systematic reviews (PRISMA-LSR) (akl2024prismalsr) and PRISMA-S for searches (rethlefsen2021prismas).

Contributors. Guarantor: Ahmad Sofi-Mahmudi. Second and third reviewers, clinical advisors (endocrinology, cardiology, nephrology, hepatology, obesity medicine, pediatrics), an information specialist for peer review of the search and patient partners: TO BE COMPLETED. Dual independent screening, risk of bias assessment and certainty ratings require at least two human reviewers; no stage begins until the team is named here.

Amendments. Logged in protocol/amendments.csv and summarized in each release note.

Support and competing interests. Funding: TO BE COMPLETED. Competing interests of every contributor, including any relationship with manufacturers of incretin based drugs or their comparators in the previous 36 months: TO BE COMPLETED. No funder or manufacturer has a role in the design, conduct, analysis, interpretation or reporting.

Introduction

Rationale

Glucagon-like peptide-1 (GLP-1) receptor agonists moved within two decades from a second line glucose lowering option for type 2 diabetes to a class used for obesity, cardiovascular and kidney risk reduction, heart failure, steatotic liver disease, obstructive sleep apnea and osteoarthritis pain, and investigated in neurodegenerative, psychiatric and substance use disorders. The class now includes daily and weekly injectable peptides, an oral peptide, oral small molecules, and molecules that also act on the GIP, glucagon or amylin pathways. Trials are being published and registered faster than conventional systematic reviews can be updated, and published network meta-analyses differ in scope, node definitions, time points and certainty assessment, so their conclusions are hard to compare and quickly outdated.

A living network meta-analysis addresses both problems: it compares every regimen against every other within each population, including regimens never compared head to head, and it is updated as evidence accrues (elliott2017lsr; nikolakopoulou2018livingnma). Making the full inventory of studies and documents public, with all code and data, lets others verify, reuse and extend each step.

Objectives

Primary objectives, for each population in populations.csv that meets the connectivity rule (section connectivity):

1. To estimate the comparative benefits and harms of all GLP-1 receptor agonist containing regimens, at the level of molecule, route, formulation and maintenance dose, against each other and against placebo, standard care, lifestyle intervention and other active drugs.
2. To rate the confidence in every network estimate for critical outcomes and to draw conclusions with the GRADE minimally contextualized framework.

Secondary objectives:

3. To maintain a public, continuously updated inventory of every randomized and nonrandomized study of these drugs, linking each study to all of its publications, registrations, protocols, statistical analysis plans, regulatory documents, results postings, supplements and shared code and data.
4. To synthesize nonrandomized evidence on rare and long latency harms.
5. To inventory published systematic reviews and meta-analyses of these drugs, appraise them, and compare their scope, methods and conclusions with ours.
6. To keep all of the above current, with dated, citable versions.

Methods

Eligibility criteria

Study designs

There is no minimum sample size or duration for inclusion. Trials shorter than the outcome windows in outcomes.csv contribute only to outcomes whose windows they meet. Single arm trials, pharmacokinetic studies in healthy volunteers and mechanistic studies are inventoried and excluded from both syntheses.

Participants

All indications are eligible. Each population in populations.csv forms its own set of networks; populations are never pooled in a primary analysis because doing so would violate transitivity. Rules:

• A trial is assigned to every population whose definition its whole sample meets, and contributes to each population's networks for that population's outcomes. Within any single network a trial contributes once.
• Trials with mixed samples contribute when at least 80 percent of participants belong to the population or a subgroup result for the population is reported.
• Type 2 diabetes (P01) glycemic and weight networks are built within background therapy strata S0, S1 and S2 (definitions in populations.csv). Patient important outcomes (mortality, cardiovascular, kidney and heart failure events) are analyzed in one type 2 diabetes network.
• Randomized withdrawal and weight maintenance trials (P02W) are analyzed separately from treatment initiation trials.
• Conditions not listed receive a population code by amendment when their first eligible trial is included (P99 rule).

Interventions

Any drug whose mechanism includes GLP-1 receptor agonism, alone or in fixed combination, at any dose, route and formulation, used for any duration. This includes the classes in node_classes.csv: short and long acting injectable GLP-1 receptor agonists, oral semaglutide, oral small molecule GLP-1 receptor agonists, GIP and GLP-1 co-agonists, glucagon and GLP-1 co-agonists, triple agonists, amylin and GLP-1 combinations or co-agonists, GIP receptor antagonist and GLP-1 agonist conjugates, and fixed ratio combinations with basal insulin. Withdrawn or discontinued agents (albiglutide, taspoglutide) are included because they inform the network, and their exclusion is a sensitivity analysis.

Comparators

Placebo; standard care or no added treatment; a randomized lifestyle intervention arm; any other GLP-1 receptor agonist containing regimen (including another dose of the same drug); and any other glucose lowering or weight lowering drug (metformin, sulfonylureas, thiazolidinediones, DPP-4 inhibitors, SGLT2 inhibitors, alpha glucosidase inhibitors, meglitinides, insulins, orlistat, phentermine and topiramate, naltrexone and bupropion and others). Trials comparing two non GLP-1 regimens are not eligible; the evidence among comparators enters only through trials that include a GLP-1 receptor agonist containing arm. Bariatric and metabolic surgery trials are inventoried and excluded from the networks, because surgical populations and open label designs are not exchangeable with drug trials.

Outcomes and time points

Outcomes, their definitions, effect measures, target time points, acceptable windows, importance ratings and provisional thresholds are listed in outcomes.csv (39 outcomes at version 0.1.0). Critical outcomes per population are listed in populations.csv. section outcomes gives the rules for choosing a time point and an estimand.

Other criteria

No restriction by language, publication date, publication status or setting. Non-English reports are translated. Conference abstracts and registry results are eligible as sources of data; a study reported only as an abstract is included and flagged, and its exclusion is a sensitivity analysis. Retracted studies are excluded from all syntheses and listed with the retraction notice.

Information sources

Every source, date searched, access route and record count is logged per run in search/runs/<date>/run_log.json. Manual searches are logged in search/manual/ with the date, the person and the result.

Search strategy

A single, versioned concept file (search/strategies/concepts.yaml) lists every class term, generic name, brand name and development code of the interventions and the PubMed controlled vocabulary. Code in R/search_strategy.R renders it into each source's syntax, and each run stores the exact query sent. No design, language, date or outcome filters are applied, so that the same harvest feeds the inventory of trials, observational studies, registrations and systematic reviews. Specifically:

• PubMed: title and abstract terms for every name and code, the MeSH heading and pharmacological action "Glucagon-Like Peptide-1 Receptor Agonists", drug MeSH headings and supplementary concepts. The endogenous hormone and receptor headings are excluded because they retrieve physiology without additional drug studies. Because ESearch returns at most 9,999 identifiers, the query is partitioned by Entrez date and ranges are halved until every partition is below the cap; the union is reconciled against the total count.
• Europe PMC: TITLE_ABS terms with explicit word forms (truncation is not supported inside phrases), all sources except patents.
• OpenAlex: title and abstract search with the same terms in chunks, unioned on work identifier.
• ClinicalTrials.gov: intervention field search for all names and codes, plus a general term search for the class terms, all statuses.

The strategy will be peer reviewed by an information specialist with the PRESS 2015 checklist before the baseline search (mcgowan2016press) (TO BE COMPLETED). Its sensitivity is checked against a benchmark set: the trials included in all published systematic reviews in the review inventory (section reviews). Any benchmark trial not retrieved leads to a documented correction of the concept file, logged as an amendment. The full rendered strategies are reproduced in Appendix A.

Study records

Data management

All data live in the public repository as plain text tables (UTF-8 CSV, compressed with gzip when large) described by a Frictionless Data Package (datapackage.json) that gives every column's type, constraints, primary keys and foreign keys. Validation runs on every change. Copyrighted material (full texts, supplements, abstracts) is stored in a local, non-public file store; the repository holds a manifest with source URL, retrieval date, SHA-256 checksum, size, media type and license for every file. Only content under licenses that permit redistribution (for example CC BY, CC0, public domain registry records) is redistributed.

The inventory has three linked levels:

• record: one bibliographic or registry item (article, preprint, conference abstract, erratum, retraction notice, registry entry, results posting, protocol, statistical analysis plan, regulatory document, published systematic review), with every identifier (PMID, PMCID, DOI, OpenAlex ID, Europe PMC ID, registry ID);
• study: one investigation, identified by a stable internal ID and, when available, its primary registry number; secondary registry numbers are stored as aliases;
• study_record: a many to many link with a role (primary report, secondary report, protocol, analysis plan, registry entry, results posting, regulatory review, erratum, retraction, commentary) and the evidence for the link.

Deduplication of records uses, in order: normalized DOI, PMID, PMCID, OpenAlex ID, then a fuzzy match on normalized title, first author surname, year and journal, reviewed by a person when the match is not exact. Every merge is logged. Records are linked to studies by registry numbers found in the PubMed data bank fields, the abstract, the full text, and the registry's own reference list, then by trial acronym, sponsor, sample size and arm set. Uncertain links are adjudicated by two reviewers. Overlapping observational populations (the same database and period) are linked through a cohort overlap table so that no population is counted twice.

Selection process

1. Classification. Every record is classified automatically, without exclusion, by publication type, registry identifiers and text features into candidate randomized trial, candidate nonrandomized study, systematic review, protocol or registration, and other (narrative review, editorial, preclinical, case report).
2. Title and abstract screening. Two reviewers screen every record independently against the eligibility criteria. Disagreements go to a third reviewer. Before independent screening the team screens a calibration set of 300 records and discusses every disagreement; screening proceeds when agreement on inclusion exceeds 90 percent.
3. Full text screening. Two reviewers assess every potentially eligible record independently, with reasons for exclusion recorded from a fixed list.
4. Prioritization. Records are ordered by a prioritization model implemented in R so that likely eligible records are seen first; every record is still screened by two people.

Screening decisions are stored with reviewer, date and reason in data/screening/.

Data collection

One reviewer extracts each study into the structured forms defined by the data package; a second reviewer verifies every value used in a synthesis against the source and records the page, table or figure. Discrepancies are resolved by discussion and logged. When sources disagree (publication versus registry versus regulatory review), the hierarchy is: data from the source with the most complete follow-up and the prespecified analysis, then the regulatory review, then the registry results posting, then the publication; all values are kept, the choice is recorded, and the alternative is a sensitivity analysis. Missing information is requested from investigators or sponsors with two reminders over six weeks.

Full text and supplementary material

For every record of a study classified as a candidate trial, nonrandomized study or systematic review, the pipeline retrieves, in order: the Europe PMC full text XML and supplementary files, the PMC open access package, the Unpaywall and OpenAlex best open access locations, the publisher page for supplements, the registry's attached protocol, analysis plan and consent documents, and any data or code deposit cited in a data availability statement (Zenodo, Figshare, Dryad, OSF, GitHub, Mendeley Data, Vivli, YODA, ClinicalStudyDataRequest). Repository deposits are frozen by version or commit hash. Items that cannot be retrieved openly are obtained through institutional access and marked as restricted. The manifest records each item's status: retrieved, restricted, unavailable or requested.

Data items

Study level: identifiers and all linked records; design; registration date versus start date (prospective or retrospective registration); countries and number of sites; sponsor and funding; dates; inclusion and exclusion criteria; run in; blinding; randomization ratio; planned and actual duration; estimands and their handling of intercurrent events (treatment discontinuation, rescue medication); analysis population; missing data method; conflicts of interest.

Arm level: node (molecule, route, formulation, maintenance dose, frequency), titration schedule, background therapy, lifestyle co-intervention, number randomized, analyzed and completing, and baseline characteristics (age, sex, race and ethnicity, region, weight, BMI, waist, HbA1c, diabetes duration, eGFR, blood pressure, cardiovascular disease, heart failure, background drugs).

Outcome level: definition as reported and its mapping to outcomes.csv; time point; estimand; statistic (adjusted mean change with standard error, mean and standard deviation, events and denominator, hazard ratio with confidence interval, rate with exposure); the source location; and the extractor and verifier.

Nonrandomized studies: additionally, design (new user, prevalent user, active comparator), data source and period, exposure definition and lag, follow-up start (time zero alignment), confounders adjusted for and method, handling of immortal time and competing risks, and adjusted and unadjusted estimates.

Outcomes and prioritization

Importance ratings in outcomes.csv (critical or important) follow GRADE and will be confirmed by clinical advisors and patient partners before the baseline analysis. Primary outcomes per population are listed in populations.csv.

Time points. Each outcome has a target week and an acceptable window. For continuous efficacy outcomes in type 2 diabetes the primary rule takes the trial's prespecified primary time point when it lies within the window, and otherwise the observation closest to the target week within the window. For all other windowed outcomes the observation closest to the target week is used; ties go to the later observation. Event outcomes use the longest randomized follow-up, with follow-up time extracted. Sensitivity analyses use narrow windows (26 weeks, 20 to 32; and 52 weeks, 40 to 60) and model trial duration as a covariate.

Estimands. The treatment policy estimand (all randomized participants, data collected regardless of adherence or rescue) is primary, consistent with ICH E9(R1) (ich2019e9r1). When a trial reports only a trial product or on treatment estimand, that result is used with a flag, and its exclusion is a sensitivity analysis. Both estimands are extracted whenever reported.

Effect measures. Mean differences in natural units for continuous outcomes (standardized mean differences only when one construct is measured with different instruments); odds ratios for binary outcomes in the model, with results presented as absolute risk differences per 1,000 using the median control group risk in the network and, for critical harms, a range of baseline risks; hazard ratios for time to event outcomes; rate ratios for recurrent events with exposure time.

Risk of bias in individual studies

• Randomized trials: RoB 2 for each result included in a synthesis, with the variants for cluster and crossover trials (sterne2019rob2). The effect of assignment is assessed for the treatment policy estimand and the effect of adherence for trial product estimands. Two reviewers assess independently and reach consensus.
• Nonrandomized studies: ROBINS-I V2 for cohort designs, recording the exact release used (robinsi2025v2); the 2016 ROBINS-I (sterne2016robinsi) with documented adaptations for designs outside the scope of the V2 release; ROBINS-E (higgins2024robinse) is not used, because these studies estimate the effects of interventions rather than exposures.
• Missing evidence: ROB-ME for each synthesis (page2023robme), informed by the comparison of registered and published trials and outcomes in the inventory.

Judgments, signaling question answers and supporting quotations are stored as data and rendered as traffic light plots.

Data synthesis

Networks and nodes

A network is defined by population (and stratum), outcome and time window. Nodes are defined at the level of molecule, route, formulation, maintenance dose and frequency (nodes.csv). Titration schedules are recorded but do not define nodes. Comparator drugs form molecule level nodes with doses pooled, unless a trial randomizes doses of a comparator against a GLP-1 receptor agonist containing node. Placebo and standard care are separate nodes; merging them is a sensitivity analysis. Background therapy shared by all arms of a trial does not define a node. New regimens are added to nodes.csv by amendment as they are found.

Connectivity rule

For each population, stratum and outcome: a network meta-analysis is performed when at least two eligible trials form a connected network of at least three nodes. When only two nodes are connected, a pairwise meta-analysis is performed with the same models. A single trial is reported descriptively. Disconnected components are analyzed separately and never bridged by assumption.

Bayesian network meta-analysis (primary)

The primary analysis is a Bayesian random effects network meta-analysis in a generalized linear model framework (dias2011tsd2; dias2018book), fitted with the R package multinma, which compiles models in Stan through rstan (phillippo2025multinma; phillippo2020mlnmr). Models outside multinma's scope (bias adjusted and dose response models) are written in Stan and fitted with cmdstanr; their Stan code is versioned in analysis/stan/ and tested by recovering known parameters from simulated networks.

For trial $j$ and arm $k$ receiving treatment $t_{jk}$,

$$ g(\theta_{jk}) = \mu_j + \delta_{jk}, \qquad \delta_{j1} = 0, \qquad \delta_{jk} \sim \mathrm{N}!\left(d_{t_{jk}} - d_{t_{j1}}, \tau^2\right), $$

with the consistency equations $d_{ab} = d_b - d_a$, a common heterogeneity variance $\tau^2$, and the covariance $\tau^2 / 2$ between relative effects in multi-arm trials. Likelihoods and links: normal with identity link for continuous outcomes from arm level means and standard errors (or the contrast level likelihood with the multi-arm covariance when only contrasts are reported); binomial with logit link for binary outcomes; normal likelihood on the log hazard ratio for time to event outcomes reported as contrasts, and binomial with complementary log log link and a log follow-up offset when only counts are reported; Poisson with log link and log exposure offset for recurrent events. Arm level and contrast level data are combined in one model when trials report differently.

Priors. Trial baselines $\mu_j \sim \mathrm{N}(0, 100^2)$ on the linear predictor scale. Relative effects: $d \sim \mathrm{N}(0, 10^2)$ on the log odds, log hazard and log rate scales, and $\mathrm{N}(0, (10 s_y)^2)$ on mean difference scales, where $s_y$ is a reference standard deviation of the outcome fixed in analysis/priors.csv before analysis (for example 1.0 percentage point for HbA1c and 7 percentage points for percent weight change). Heterogeneity: for binary outcomes, the log normal predictive distribution for $\tau^2$ from Turner and colleagues for the matching outcome type and pharmacological versus placebo or control comparisons (turner2015priors); for continuous outcomes, $\tau \sim \text{half-}\mathrm{N}(0, (0.5 s_y)^2)$, a weakly informative choice consistent with the empirical distributions of Rhodes and colleagues (rhodes2015priors). All selected values are recorded in analysis/priors.csv before the baseline analysis. Prior predictive checks confirm that the priors give plausible outcome values. Sensitivity analyses use vague priors ($\tau \sim \mathrm{U}(0, 5 s_y)$ or half normal with scale $5$ on log scales).

Computation and diagnostics. Four chains with 2,000 warmup and 2,000 sampling iterations each (increased when needed), fixed seeds recorded in the output. Convergence requires rank normalized $\hat{R} < 1.01$, bulk and tail effective sample sizes of at least 400 for every reported parameter (vehtari2021rhat), no divergent transitions after warmup (adapt_delta raised up to 0.99, then reparameterization), and E-BFMI above 0.3. Monte Carlo standard errors are reported. Model fit is assessed by total residual deviance against the number of data points, leverage plots, posterior predictive checks and the DIC; random and fixed effect models are compared by DIC. Fitted objects are not committed; posterior summaries and draws of the quantities reported are, with the checksums of their inputs.

Frequentist network meta-analysis (secondary)

The same networks are fitted with netmeta (balduzzi2023netmeta; rucker2012graph) using the graph theoretical random effects model with a common heterogeneity variance estimated by REML, reporting 95 percent confidence intervals and 95 percent prediction intervals (riley2011prediction). Binary outcomes with sparse events are also analyzed with the Mantel-Haenszel network method (netmetabin), which needs no continuity correction (efthimiou2019mh). The frequentist analysis provides the contribution matrix used for CINeMA (papakonstantinou2018contribution) and a check of robustness to the modeling framework. Differences between frameworks larger than the outcome threshold are reported and investigated.

Zero events and sparse data

The binomial likelihood handles zero cells without correction. Trials with no events in any arm contribute no information to relative effects on the odds scale; they are reported, and a sensitivity analysis on the risk difference scale includes them. Continuity corrections are not used.

Missing statistics and data conversions

Standard deviations are derived from standard errors, confidence intervals, t statistics or P values when not reported; medians and ranges or interquartile ranges are converted with the methods of Wan and colleagues (wan2014sd) and flagged; when no dispersion can be derived, the median standard deviation of the same outcome and time window in trials of the same population is imputed and its exclusion is a sensitivity analysis. Values read from figures are digitized twice independently and flagged. Adjusted (ANCOVA or mixed model) estimates are preferred to unadjusted ones. Change from baseline is preferred to final values; they are not mixed for standardized mean differences.

Transitivity

The distributions of the effect modifiers in effect_modifiers.csv are tabulated and plotted across comparisons for each network before any treatment effect is examined. Imbalance in a modifier across comparisons is addressed by network meta-regression (below), by restriction, or, when neither is possible, by downgrading for indirectness and incoherence.

Heterogeneity

Heterogeneity is reported as $\tau$ with its credible interval and as 95 percent prediction intervals for every comparison, interpreted against the outcome thresholds. $I^2$ is reported for the frequentist model as a secondary descriptive measure.

Incoherence

Global incoherence: comparison of the consistency model with an unrelated mean effects (inconsistency) model by DIC and residual deviance per data point (dias2011tsd2), and the design by treatment interaction model for the frequentist analysis (higgins2012designbytreatment) with the net heat plot (krahn2013netheat). Local incoherence: node splitting for every comparison with direct and indirect evidence (dias2010nodesplit) and the separate indirect from direct evidence (SIDE) method in netmeta. A nonsignificant test is not taken as evidence of coherence in sparse networks; incoherence is judged by the size of the discrepancy relative to the outcome threshold. Incoherence is investigated for data errors, effect modifiers and design differences before any change to the analysis.

Network meta-regression and subgroups

Prespecified covariates per population: those in effect_modifiers.csv (baseline HbA1c, BMI, background therapy, diabetes duration, established cardiovascular disease, eGFR, age, sex, proportion of Asian participants, trial duration, lifestyle co-intervention intensity, estimand, year, funding, blinding). Each is fitted as a common interaction across treatments within a class, with exchangeable class interactions as an alternative, when at least ten trials inform the network. Subgroup results reported within trials (for example by baseline BMI) are extracted for multilevel network meta-regression should individual participant data become available (phillippo2020mlnmr).

Additional models

• Drug level network: doses of each molecule pooled into one node.
• Class effects model: exchangeable effects within the classes in node_classes.csv, using multinma's class effect structure.
• Dose response: Emax and linear dose response models within molecule (model based network meta-analysis (mawdsley2016mbnma)), written in Stan.
• Component network meta-analysis for fixed ratio combinations of insulin and a GLP-1 receptor agonist, with additive and interaction models (rucker2020cnma).

Sensitivity analyses

Prespecified for every primary network: excluding trials at high risk of bias for the outcome; excluding trials reported only as abstracts or registry entries; excluding trials with imputed standard deviations; excluding trial product estimands; narrow time windows; fixed effect model; vague priors; merging placebo and standard care; excluding withdrawn or discontinued agents; excluding trials in exclusively East Asian populations; excluding industry funded trials (descriptive).

Ranking

Surface under the cumulative ranking curve (Bayesian) and P scores (frequentist) are reported as secondary summaries (salanti2011sucra; rucker2015pscore), always with the absolute effects and their certainty; rankings are never used alone to draw conclusions.

Synthesis of nonrandomized studies for harms

For each harm in outcomes.csv marked for nonrandomized evidence, adjusted relative effects are pooled by random effects meta-analysis, separately by design (cohort, case control, self controlled) and by comparator (active versus non user), using REML with the Hartung, Knapp, Sidik and Jonkman adjustment and a Bayesian model with the same heterogeneity priors as above. Where data sources overlap, one estimate per source, outcome and comparison is chosen by a prespecified hierarchy: active comparator new user design, then the longest follow-up, then the largest sample. Results are presented beside the randomized evidence and never pooled with it in the primary analysis. A bias adjusted synthesis of randomized and nonrandomized evidence is an exploratory sensitivity analysis with bias priors published before it is run.

Meta-biases

Risk of bias due to missing evidence is judged with ROB-ME (page2023robme), using: registered trials without results; outcomes registered but not reported; comparison adjusted funnel plots when at least ten trials inform a comparison with a common reference (chaimani2013graphical); and discrepancies between registered and published primary outcomes.

Confidence in cumulative evidence

Confidence in each network estimate of every critical outcome is assessed with CINeMA in its six domains: within study bias, reporting bias, indirectness, imprecision, heterogeneity and incoherence (nikolakopoulou2020cinema; papakonstantinou2020cinema). Within study bias and indirectness are judged from each estimate's study contributions; imprecision and heterogeneity compare the Bayesian credible and prediction intervals with the range of little or no difference defined by the outcome threshold in outcomes.csv; incoherence uses node splitting and the inconsistency model. Thresholds are provisional at version 0.1.0 and will be fixed by the clinical advisors and patient partners, with their sources, before any analysis. Every judgment and its reason are stored as data.

Conclusions follow the GRADE minimally contextualized framework (brignardello2020minimally; brignardello2018grade; puhan2014grade), which classifies each intervention against the reference by the size of its effect relative to the threshold and by certainty. Nonrandomized harms evidence is rated with GRADE using ROBINS-I, starting at high certainty and rating down for risk of bias (schunemann2019grade18).

Living review methods

Cadence. Searches run automatically every week. New records are screened monthly. Analyses are refitted and a new version released every quarter.

Triggers for an earlier release. Within four weeks of any of: results of a trial with at least 1,000 participants or of a dedicated cardiovascular, kidney or heart failure outcome trial; the first trial of a new node in a primary network; a regulatory safety communication about the class from FDA, EMA, MHRA, Health Canada, PMDA or TGA; a retraction or major correction of an included study.

Versions. Each release is a tagged, immutable snapshot with a version number (year and quarter, for example 2027.1), the search date, the data and code commit, software versions, seeds, a change log separating new evidence, corrections and methods changes, a PRISMA-LSR flow diagram with cumulative and update specific counts, and a DOI from Zenodo. A release with no new eligible evidence is still published as a dated notice.

Statistical approach to updating. Every release refits all models to the cumulative data with the original priors; posteriors are never recycled as priors. No sequential testing is performed, intervals are estimates rather than tests, and no release is triggered by statistical significance (simmonds2017lsr). A conclusion is considered changed when the GRADE minimally contextualized category or the certainty level of an estimate changes, and release notes highlight such changes.

Human gates. Automation retrieves, deduplicates, classifies and prioritizes; inclusion, extraction, risk of bias, certainty and conclusions are human decisions (thomas2017lsr).

Retirement. The review moves from living to periodic updating when, for four consecutive releases, no conclusion changes and no ongoing trial that could change a conclusion is registered, or when funding ends; the decision and its reasons are published.

Comparison with published systematic reviews

Systematic reviews and meta-analyses of GLP-1 receptor agonist containing drugs are identified from the same harvest (by publication type, title and abstract classification and screening) and by searching PROSPERO and OSF for registered reviews. For each review we retrieve the full text, supplements, shared data and code, and the protocol (PROSPERO, OSF, a journal protocol, or the methods section when no protocol exists). We extract scope (population, interventions, comparators, outcomes, designs), search dates, synthesis methods, certainty methods and main results; appraise methodological quality with AMSTAR 2 (shea2017amstar2) and risk of bias with ROBIS (whiting2016robis); record discrepancies between protocol and publication; and quantify the overlap of included trials with the corrected covered area (pieper2014cca). Trials included by any review serve as the benchmark set for our search (section search). For estimates that overlap in scope, we compare direction, size and certainty with ours and explain differences.

Reporting and dissemination

• Paper. A classic article for each population or group of populations, following PRISMA 2020, PRISMA-NMA and PRISMA-LSR, generated from the same analysis outputs.
• Website. A static website on GitHub Pages built in TypeScript, with an evidence map of the inventory, every network with its interactive plots (network graph, forest plots, league tables, CINeMA profiles, funnel and contribution plots), the study pages with all linked documents, the review inventory, release notes and downloads. Interactive plots are produced in R with the ggextreme package and embedded in the site. Every figure has a text and table alternative, and the site targets WCAG 2.2 level AA.
• Data and code. All code (R) and all data produced by the review are public in the repository under the licenses in section open, with a versioned JSON interface for the website described by JSON Schemas.

Reproducibility and open science

The repository is organized as an R package (functions documented and tested) with an renv lockfile and a pinned CmdStan version. Every analysis is run from scripts in pipeline/. Searches run on continuous integration; model fitting runs on a workstation or self hosted runner, never on shared hosted runners. Inputs are checksummed, and every output records the inputs, code commit and seed that produced it.

Licenses: code under GPL 3 or later; data produced by the review (inventory tables, screening decisions, extracted data, assessments, results) under CC BY 4.0; third party content keeps its own license and is redistributed only where that license permits.

Use of automation and large language models

Automated steps (searches, deduplication, classification, prioritization, retrieval) are deterministic code in this repository. Large language models may assist with code, with drafting, and as an advisory, non binding additional reader during screening and extraction. They never replace a human decision. When used in screening or extraction, their agreement with human reviewers is measured and reported. Each release names the tools and versions used (TO BE COMPLETED before registration).

Patient and public involvement

People living with type 2 diabetes, obesity and other conditions studied will be invited to review outcome importance, thresholds and plain language summaries, and to comment on the website (TO BE COMPLETED).

Ethics

The review uses published and publicly registered aggregate data and does not require ethics approval. Individual participant data, if ever obtained, will be used under the data provider's agreement and never published.

Timeline

TO BE COMPLETED after the team is named: registration; PRESS review; baseline search (search code is validated, and a baseline harvest on 2026-10-12 tested the pipeline); calibration and screening; extraction; baseline analysis; first release.

References {.unnumbered}

Appendix A. Search strategies {.unnumbered}

The rendered strategy for each source is regenerated from search/strategies/concepts.yaml and stored with every run in search/runs/<date>/<source>_query.txt. The strategies of the run that precedes registration will be pasted here verbatim.