CoChat
Sign Up
Cited by 1Diabetes Treatment and ManagementAdvanced Causal Inference TechniquesMachine Learning in Healthcare

Emulating real-world GLP-1 efficacy in type 2 diabetes through causal learning and virtual patients

Open access
MacLellan, Calum Robert, Petkov, Hristo, McKeag, Conor, Dong, Feng, Lowe, David John, Maguire, Roma, Moschoyiannis, Sotiris, Armes, Jo +3 more·2025·PLOS Digital Health
DOI 10.1371/journal.pdig.0000927PMC PMC12279107
Open this paper in CoChat — free accountOpen PDF

Abstract

Randomized controlled trials (RCTs) remain the benchmark for assessing treatment effects but are limited to phenotypically narrow populations by design. We introduce a novel generative artificial intelligence (AI) driven emulation method that infers effect size through virtual clinical trials, which can emulate the RCT process and potentially extrapolate into wider populations. We validate the virtual trials by comparing the predicted impact of glucagon-like peptide-1 (GLP-1) agonists on HbA1c in type-2 diabetes (T2DM) with its true efficacy established in the LEAD-5 trial. Our emulation model learns treatment effects from real-world evidence data by a combined generative AI and causal learning approach. Training data comprised pre- and post-treatment outcomes for 5,476 people with T2DM. We considered three treatment arms: GLP-1 (Liraglutide), basal insulin (glargine), and placebo. After training, virtual trials were conducted by sampling 232 virtual patients per arm (according to the LEAD-5 inclusion criteria) and predicting post-treatment outcomes. We used difference-in-differences (DiD) for pairwise comparisons between arms. Our goal was to emulate LEAD-5 by demonstrating a significant DiD in post-treatment HbA1c reduction for GLP-1 compared to basal insulin and placebo. We found significant differences in HbA1c reduction for GLP-1 vs basal insulin (-1.21 mmol/mol (-0.11%); p < 0.001) and GLP-1 vs placebo (-2.58 mmol/mol (-0.24%); p < 0.001) in our virtual populations, consistent with LEAD-5 (Liraglutide vs glargine: -2.62mmol/mol (-0.24%); p = 0.0015, Liraglutide vs placebo: -11.91 mmol/mol (-1.09%); p < 0.0001). The causal AI-powered clinical trials can emulate LEAD-5 in important measurements for T2DM. Our algorithm is specialty agnostic and can explore counterfactual questions, making it suitable for further study in the generalizability of RCT results in real-world populations to support clinical decision-making and policy recommendations.

AI snapshot

Written from the paper’s full text.

TL;DR

A generative AI plus causal learning model learned treatment effects from real-world type 2 diabetes data and emulated the LEAD-5 trial by generating virtual patients and estimating post-treatment outcomes. The virtual trials reproduced the directional ranking of GLP-1 vs basal insulin and placebo for HbA1c and BMI, but effect sizes differed and counterfactual simulations showed GLP-1 was not always the most effective treatment outside LEAD-5 criteria.

Key findings

  • For HbA1c, virtual trials found significant GLP-1 reductions vs basal insulin (-1.21 mmol/mol (-0.11%); p < 0.001) and vs placebo (-2.58 mmol/mol (-0.24%); p < 0.001), matching the direction of LEAD-5 (liraglutide vs glargine: -2.62 mmol/mol (-0.24%); p = 0.0015; liraglutide vs placebo: -11.91 mmol/mol (-1.09%); p < 0.0001).
  • For BMI, virtual GLP-1 reduced BMI significantly vs basal insulin (-0.79 kg/m²; p < 0.001) and placebo (-0.61 kg/m²; p < 0.001), although BMI increased in all virtual arms; LEAD-5 reported weight reductions of -3.43 kg vs glargine and -1.39 kg vs placebo.
  • For systolic blood pressure, virtual GLP-1 reduced SBP significantly vs basal insulin (-2.99 mmHg; p < 0.001) and placebo (-2.38 mmHg; p < 0.001); LEAD-5 found a significant reduction vs glargine (-4.51 mmHg; p = 0.0001) but not vs placebo (-2.53 mmHg; p = 0.0791).
  • In counterfactual simulations on real SCI-Diabetes patients, among those who actually received basal insulin, GLP-1 was significantly less effective for HbA1c than basal insulin (+1.10 mmol/mol; p < 0.001); among those who actually received GLP-1, GLP-1 was also less effective than basal insulin (+0.64 mmol/mol; p < 0.001).

Methods

  • The emulation model combined a Wasserstein GAN with gradient penalty (WGAN-GP) and NOTEARS-MLP to learn a directed acyclic graph and structural causal model from real-world observational data, then generated virtual patients satisfying LEAD-5 inclusion criteria.
  • Training data comprised pre- and post-treatment outcomes for 5,476 people with T2DM (abstract); the dataset section describes SCI-Diabetes as containing 56,476 unique patients with T2DM. Pre-features were collected from 9 months before treatment and post-features within 12 months after, using median values if multiple measurements existed.
  • Virtual trials sampled 232 virtual patients per arm (placebo, basal insulin/glargine, GLP-1/liraglutide), predicted post-treatment outcomes, and used difference-in-differences for pairwise comparisons, with results computed over 60 random seeds and 95% confidence intervals obtained by bootstrapping.

Limitations

  • The paper states that virtual patients were sampled with a random-based approach according to inclusion criteria, and that sampling covariates independently may overlook real-world correlations, potentially undermining ecological validity; future work could use causal-preserving generative models or Bayesian bootstrapping.
  • The paper states that pre- and post-treatment features were based on yearly median values, which may obscure fine-grained treatment effects; the emulation assumes causal sufficiency, and unmeasured confounding remains possible during model training.
  • The paper states that SCI-Diabetes uses GLP-1 as a general drug category rather than the specific liraglutide tested in LEAD-5, that basal insulin and placebo categories are not identical to the original trial drugs, and that different units (HbA1c mmol/mol vs %, BMI vs bodyweight) and inability to replicate an identical cohort contributed to discrepancies; it also notes complete-case analysis reduced the training set. Evident from the text: the abstract and dataset section report different T2DM cohort sizes (5,476 vs 56,476), which is not reconciled.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1371/journal.pdig.0000927

Cites

12 of 12 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

12 of 14 cited sources checked · 8 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    LEAD-5 established that Liraglutide reduced HbA1c significantly compared with glargine (difference -2.62 mmol/mol (-0.24%), p = 0.0015).

    “LEAD-5 established that Liraglutide reduced HbA1c significantly compared with both glargine (-2.62mmol/mol (−0.24%) difference; p = 0.0015)”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] D Russell-Jones, 2009, Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU): a randomised controlled trial, Diabetologia source
  • Premisecitation

    LEAD-5 established that Liraglutide reduced HbA1c significantly compared with placebo (difference -11.91 mmol/mol (-1.09%), p < 0.0001).

    “and placebo (-11.91mmol/mol(−1.09%) difference; p < 0.0001)”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] D Russell-Jones, 2009, Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU): a randomised controlled trial, Diabetologia source
  • Premisecitation

    LEAD-5 demonstrated a significant reduction in SBP with GLP-1 compared with glargine (difference -4.51 mmHg, p = 0.0001).

    “LEAD-5 demonstrated a significant reduction in SBP compared with glargine (−4.51 mmHg difference; p = 0.0001)”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] D Russell-Jones, 2009, Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU): a randomised controlled trial, Diabetologia source
  • Premisecitation

    LEAD-5 demonstrated that Liraglutide produced significantly greater weight loss compared with glargine (difference -3.43 kg, p < 0.0001).

    “LEAD-5 also demonstrated that Liraglutide produced significantly greater weight loss compared with both glargine (−3.43 kg difference; p < 0.0001)”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] D Russell-Jones, 2009, Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU): a randomised controlled trial, Diabetologia source
  • Premisecitation

    LEAD-5 demonstrated that Liraglutide produced significantly greater weight loss compared with placebo (difference -1.39 kg, p = 0.0001).

    “and placebo (–1.39 kg difference; p = 0.0001)”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] D Russell-Jones, 2009, Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU): a randomised controlled trial, Diabetologia source
  • Classified as probably not a substantial claim. 12
  • May combine two claims. 17

  • Probably not a substantial claim: may not pass S1 (asserted).

    Premisecitation

    The goal in LEAD-5 was to confirm whether liraglutide, a GLP-1 receptor agonist, caused a greater reduction in HbA1c and bodyweight compared to insulin glargine and placebo in people with T2DM.

    “The goal in LEAD-5 was to confirm whether liraglutide, a glucagon-like peptide-1 (GLP-1) receptor agonist, caused a greater reduction in HbA1c and bodyweight compared to insulin glargine and placebo in people with T2DM.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [9] D Russell-Jones, 2009, Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU): a randomised controlled trial, Diabetologia source

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    We designed our emulation model to unify recent developments in machine learning-based generative modelling and causal structure (DAG) learning within a single framework.

    “we designed our emulation model to unify recent developments in machine learning-based generative modelling and causal structure (DAG) learning within a single framework.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites [11] I Gulrajani, 2017, Improved training of Wasserstein GANs, In: Advances in Neural Information Processing Systems source

      The Introduction (p. 1) describes GAN training and proposes a gradient penalty, but it does not describe an emulation model combining generative modelling with causal structure learning. Searches for DAG and causal structure learning found no passage addressing that proposed unification; the sentence appears to describe the citing authors’ own model. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [12] X Zheng, 2020, Learning sparse nonparametric DAGs., In: Proc Int Conf Artif Intell Stat source

      The Introduction (p. 1) describes a general framework for learning DAG models, supporting the causal-structure-learning component. The cited paper does not say that this framework unifies generative modelling with DAG learning, or describe the citing authors’ emulation model. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    Amongst patients who received basal insulin in practice, simulations revealed GLP-1 to be less effective than placebo for reducing HbA1c, but the result was not significant (GLP-1 vs placebo: +0.10 mmol/mol difference; p = 0.472).

    “patients belonging to the real basal insulin group (Table 3, Insulin column) benefit more from both placebo (GLP-1 vs placebo: + 0.10 mmol/mol difference) and basal insulin (GLP-1 vs insulin: + 1.10 mmol/mol difference; p < 0.001) than GLP-1, in terms of lower HbA1c, though the result with placebo is not significant (p = 0.472).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.

  • May combine two claims.

    Findingown data

    In virtual trials, BMI increased in all treatment arms, unlike LEAD-5 where bodyweight reduced.

    “BMI increased in all treatment arms, instead of reducing bodyweight as in LEAD-5.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.
    • The authors state this as an assumption: Treatment groups are exchangeable conditional on inclusion criteria, and marginal balance is sufficient to estimate the average treatment effect.

  • May combine two claims.

    Findingown data

    The SCI-Diabetes dataset contains approximately 300K individuals with diabetes, including 3K individuals with MODY recorded with certainty, and an array of variables for 56,476 unique patients with T2DM.

    “There are approximately 300K individuals with diabetes, where 3K individuals with MODY (Maturity-onset diabetes of the young) are recorded with certainty (genetic information) along with records of individuals with negative genetic test results. The dataset consists of an array of variables for clinical measurements, drug prescriptions, and demographics for 56,476 unique patients with T2DM.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Wasserstein distances are slightly elevated for HbA1c (8.29) and BMI (1.57).

    “This trend aligns with slightly elevated Wasserstein distances in Table S for these variables (e.g., 8.29 for HbA1c and 1.57 for BMI).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The relative ordering of treatment effects is preserved, and the synthetic data falls within plausible physiological ranges for all variables examined.

    “Despite these limitations, the relative ordering of treatment effects is preserved, and the synthetic data falls within plausible physiological ranges for all variables examined.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    Sampling covariates independently may overlook real-world correlations among variables, potentially undermining ecological validity.

    “We acknowledge that sampling covariates independently may overlook the real-world correlations among variables, potentially undermining ecological validity.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The main finding of LEAD-5 did not translate into phenotypically diverse populations, where Liraglutide was not the most effective choice in all patient cases.

    “the main finding of LEAD-5 did not translate into phenotypically diverse populations, where the key investigative drug (Liraglutide) was not the most effective choice in all patient cases.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.

  • May combine two claims.

    Interpretationown data

    The model's predictions match the LEAD-5 conclusions for the GLP-1 and basal insulin comparison on SBP, but not between GLP-1 and placebo.

    “Our model’s predictions therefore match the LEAD-5 conclusions for the GLP-1 and basal insulin comparison, but not between GLP-1 and placebo.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.
    • The authors state this as an assumption: Treatment groups are exchangeable conditional on inclusion criteria, and marginal balance is sufficient to estimate the average treatment effect.

  • May combine two claims.

    Interpretationown data

    The benefits of GLP-1 are only clear when intervening on phenotypically narrow cohorts; when conditions are relaxed, the conclusions can change.

    “The benefits of GLP-1 are therefore only clear when we intervene on phenotypically narrow cohorts: when we relax these conditions (as we do here), the conclusions are free to change.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Our model was designed to isolate causal effects by controlling for known covariates and enforcing d-separation in the DAG.

    “Our model was designed to isolate causal effects by controlling for known covariates and enforcing d-separation in the DAG.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The preserved dataset is sufficient for training the structural model and conducting virtual trial emulation across the three main comparators relevant to the LEAD-5 replication.

    “We therefore consider the preserved dataset sufficient for training the structural model and conducting virtual trial emulation across the three main comparators relevant to the LEAD-5 replication.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    To match the group sizes in LEAD-5, 232 virtual patients were randomly assigned to one of three intervention groups (placebo, glargine, or GLP-1).

    “To match the group sizes in LEAD-5, we randomly assigned n = 232 virtual patients to one of our three intervention groups (i.e., placebo, glargine, or GLP-1).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    It is not possible to replicate a completely identical cohort due to different settings between the trial and real-world clinical practice.

    “it is not possible to replicate a completely identical cohort due to different settings between the trial and real-world clinical practice.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    Monitoring diagnostic metrics ensured that the model learned a plausible causal structure and generated realistic virtual patient data suitable for downstream trial emulation.

    “These diagnostics ensured that the model not only learned a plausible causal structure but also generated realistic virtual patient data suitable for downstream trial emulation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Pre- and post-treatment features were based on yearly median values, which may have obscured fine-grained treatment effects and introduced errors into the learning process.

    “Our pre- and post-treatment features were based on yearly median values, which may have obscured fine-grained treatment effects and introduced errors into the learning process.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Virtual populations were randomly assigned to one of three LEAD-5 treatment arms: placebo, basal insulin (glargine), or GLP-1 (liraglutide).

    “the virtual populations were first generated according to the trial inclusion criteria of LEAD-5, and then randomly assigned to one of three LEAD-5 treatment arms: placebo, basal insulin (glargine), or GLP-1 (liraglutide).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Random sampling of virtual patients eliminates confounding and prevents bias in the estimation of drug effects, enabling meaningful group comparisons.

    “As with treatment assignment in an RCT, our sampling ensures that we eliminate any confounding and thus prevent bias in the estimation of the drug effects (since all virtual patients have similar conditions prior to treatment), enabling meaningful group comparisons.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The relative treatment effects were preserved, with GLP-1 showing more favourable outcomes than basal insulin and placebo, consistent with the directional findings of LEAD-5.

    “the relative treatment effects were preserved, with GLP-1 showing more favourable outcomes than basal insulin and placebo, consistent with the directional findings of LEAD-5.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.
    • The authors state this as an assumption: Treatment groups are exchangeable conditional on inclusion criteria, and marginal balance is sufficient to estimate the average treatment effect.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductcitation

    This study (ID: GSH22DI002) was approved by the Research and Innovation Peer Review Committee and the Local Privacy Advisory Committee at NHS Greater Glasgow and Clyde, UK, together with the Safe Haven platform team that provided access to the retrospective data.

    “This study (ID: GSH22DI002) was approved by the Research and Innovation Peer Review Committee and the Local Privacy Advisory Committee at NHS Greater Glasgow and Clyde, UK, together with the Safe Haven platform team that provided access to the retrospective data [10].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [10]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Study design and conductreasoning

    We employed a causal structure learning approach to adjust for observed confounders and simulate treatment assignment under randomized conditions.

    “To mitigate these, we employed a causal structure learning approach to adjust for observed confounders and simulate treatment assignment under randomized conditions.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    We used d-separation on the DAG to identify confounders and adjust for them, mimicking the balance achieved in randomized trials.

    “Using d-separation on the DAG, we identify confounders (common causes of treatment and outcome) that must be controlled.”
    • Also stated in: § Background in causal learning · plain
      “This adjustment mimics the balance achieved in randomized trials, making treatment and control groups comparable.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Interpretationreasoning

    Our method provides a flexible and scalable framework to explore efficacy, safety, and heterogeneity of treatment effects beyond diabetes, potentially accelerating evidence generation in a wide range of clinical areas.

    “By enabling counterfactual reasoning and treatment effect estimation in synthetic populations, our method provides a flexible and scalable framework to explore efficacy, safety, and heterogeneity of treatment effects beyond diabetes, potentially accelerating evidence generation in a wide range of clinical areas.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The observed discrepancies likely stem from the model's tendency to regularise around typical values, possibly limiting its capacity to capture outlier cases.

    “These discrepancies likely stem from the model’s tendency to regularise around typical values, possibly limiting its capacity to capture outlier cases.”
    Quote in the paperNot a substantial claim
25 of 25

  • The authors state this as an assumption.

    All relevant confounders have been measured (causal sufficiency), and no hidden confounding applies to the model training phase.

    “Our emulation assumes causal sufficiency, where all relevant confounders have been measured.”
    “we acknowledge that the assumption of no hidden confounding applies to the model training phase, where causal structure and parameters are learned from observational data.”

  • The authors state this as an assumption.

    Treatment groups are exchangeable conditional on inclusion criteria, and marginal balance is sufficient to estimate the average treatment effect.

    “where treatment groups are assumed to be exchangeable conditional on inclusion criteria, and marginal balance is sufficient to estimate the average treatment effect.”

  • The paper relies on this without support.

    Proton pump inhibitors (PPIs) and statins do not directly affect glycemic outcomes such as HbA1c.

    “These medications are commonly prescribed in the T2DM population but are not known to directly affect glycemic outcomes such as HbA1c.”

  • The authors state this as an assumption.

    All drugs within each class cause similar effects to a given patient's features.

    “under the assumption that all drugs within each class cause similar effects to a given patients’ features.”

  • The authors state this as an assumption.

    Demographic variables and pre-features follow independent normal distributions.

    “We assume independent normal distributions for the demographic variable and pre-features”
5 of 5

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Virtual SBP estimate falls outside its reported 95% confidence interval

    Major

    For GLP-1 versus glargine, the virtual SBP difference is −2.99 mmHg, but its reported 95% confidence interval runs from −2.62 to −1.89 mmHg and excludes that estimate.

  • Counterfactual HbA1c confidence interval conflicts with significance claim

    Major

    For the placebo patient group's GLP-1-versus-insulin comparison, the reported 95% confidence interval extends from −2.92 to +2.42 and includes zero, yet the same comparison is reported as p < 0.001.

  • Virtual effect sizes are described as exaggerated despite smaller reported differences

    Minor

    The discussion says virtual effect sizes were exaggerated in most cases, but both HbA1c comparisons shown in Table 2 are smaller in magnitude than their LEAD-5 counterparts. Both directly comparable SBP differences in the same table are also smaller in magnitude.

3 of 3

Cites

Flagged first, then by citations

3 of the 15 references couldn’t be matched to a known paper.

Research works printed but not linked: 1

  1. I Goodfellow, 2014, Generative adversarial nets, 27

Cited by

Sorted by citations

Related

Authors
MacLellan, C. R., Petkov, H., McKeag, C., Dong, F., Lowe, D. J., Maguire, R., Moschoyiannis, S., Armes, J., Skene, S., Finlinson, A., & Sainsbury, C.
Affiliations
  • University of Strathclyde (GB)
  • NHS Greater Glasgow and Clyde (GB)
  • Queen Elizabeth University Hospital (GB)
  • University of Surrey (GB)
Corresponding author
Dong, Feng
ORCID
Venue
PLOS Digital Health
Publisher
Public Library of Science (PLoS)
Published
2025-07-21
Volume
4
Issue
7
Pages
e0000927
ISSN
2767-3170 (ISSN-L)
DOI
10.1371/journal.pdig.0000927
PMC
PMC12279107
OpenAlex
W4412530564
Semantic Scholar
6165ee6e4fdb36d306ce28d5c4d4bb39a4d166c9
Cited by
1
References
15
Citations by year
2026: 1
FWCI
0.61
Type
journal article
Language
en
Open access
Licence
Locations
  • PLOS Digital Health (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Strathprints: The University of Strathclyde institutional repository (University of Strathclyde) (acceptedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
Medical Research Council (MR/X005925/1)
Topics
  • Health Sciences › Medicine › Endocrinology, Diabetes and Metabolism › Diabetes Treatment and Management
  • Physical Sciences › Mathematics › Statistics and Probability › Advanced Causal Inference Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
Keywords
GLP-1 receptor agonists, liraglutide, type 2 diabetes, HbA1c reduction, virtual clinical trials, generative AI, causal inference, real-world evidence, basal insulin, virtual patients

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-13.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-13.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review