The study, led by Anastasios Mavridis and colleagues at Sahlgrenska University Hospital and the University of Gothenburg, drew on the Swedish National Diabetes Register, a national quality registry that captures the vast majority of diabetes care in the country. The researchers followed 171,917 individuals who were registered for the first time between January 2013 and December 2019, accumulating more than 1.6 million clinical registrations and up to seven years of follow-up per person. During that window, 2,734 people suffered an ischemic stroke and 428 a hemorrhagic stroke. Because Sweden links its health registries through unique personal identity numbers, the team could connect each patient’s laboratory values, prescriptions, comorbidities and stroke outcomes into a single longitudinal record of routine clinical care.

What sets the analysis apart from most real-world studies is how it handled the messy dynamics of actual treatment. In clinical practice, patients switch drugs, stop and restart, and their blood pressure, weight, kidney function and cholesterol evolve alongside those decisions. Standard observational analyses typically adjust for characteristics measured once, at baseline, an approach that can badly distort estimates when treatment and risk factors influence each other over time. The Swedish team instead used inverse probability of treatment weighting within a marginal structural model, a statistical framework designed precisely for this treatment-confounder feedback. Each patient’s exposure status was updated at every registry visit, drawing on pharmacy dispensing records to reconstruct when drugs were actually in use, and weights were recalculated to balance clinical characteristics across treated and untreated intervals.

The headline result concerns GLP-1 receptor agonists. Across weighted time-varying Cox models, use of these drugs was associated with a 54 percent lower hazard of ischemic stroke (hazard ratio 0.46, 95 percent confidence interval 0.29 to 0.72) and a 49 percent lower hazard of total stroke (hazard ratio 0.51, 95 percent confidence interval 0.33 to 0.79). For hemorrhagic stroke the point estimate pointed in the same direction but the confidence interval was wide, spanning 0.28 to 2.28, reflecting the small number of bleeding events. By contrast, SGLT2 inhibitors showed no clear association with any stroke subtype, with an ischemic stroke hazard ratio of 0.94, and DPP-4 inhibitors were likewise neutral, at 1.17 for ischemic stroke.

To place these numbers in context, the authors systematically searched PubMed for meta-analyses of both randomized trials and observational studies, screening 274 abstracts and 102 full texts. The picture that emerged explains why the field has been so confused. For SGLT2 inhibitors, meta-analyses of randomized controlled trials consistently reported no effect on stroke, with effect sizes between 0.84 and 1.30, while observational meta-analyses suggested a protective association between 0.75 and 0.87. For GLP-1 receptor agonists, both trial and observational syntheses agreed on a protective effect, roughly a 10 to 20 percent risk reduction. DPP-4 inhibitors were neutral in trials. The new Swedish findings align with the randomized trial evidence for all three classes, and they do so using a method that many earlier observational studies lacked.

Why did previous real-world studies so often disagree with the trials? The authors argue that the answer lies largely in methodology. Most prior observational work adjusted only for baseline variables, which is appropriate for estimating the effect of initial treatment assignment but poorly suited to dynamic treatment patterns. When patients frequently switch or discontinue therapy, an intention-to-treat style estimator loses information about whether prescriptions are actually filled and says little about the hazard at any given moment of drug use. By updating exposure, covariates and concomitant medications throughout follow-up, the marginal structural model approach directly targets the question of what happens while patients are on treatment, which may partly explain why the GLP-1 association observed here was stronger than the 10 to 20 percent reductions typically reported in randomized trials.

The authors are careful to caution against over-interpreting that larger magnitude. Their estimates reflect treatment use during follow-up, not the effect of starting a drug in a treatment-naive patient, and the two estimands are not directly comparable. Differences in adherence, persistence, study populations and residual confounding may all contribute. As with any observational study, unmeasured confounding remains a possibility, and the team acknowledges that substantial missing data in several clinical variables, handled through two-level multiple imputation, means the results depend on the adequacy of the imputation models. The hemorrhagic stroke analyses were also limited by small event counts, and the comparison with published meta-analyses was explicitly contextual rather than a formal evidence synthesis.

Biologically, the divergence between drug classes makes considerable sense. Beyond lowering glucose, GLP-1 receptor agonists reduce systolic blood pressure, body weight and post-meal lipaemia, improve endothelial function and dampen vascular inflammation. Experimental work suggests anti-atherosclerotic effects through reduced oxidative stress, better nitric oxide bioavailability and modulation of macrophage activity within plaques, potentially stabilizing the vulnerable lesions that cause ischemic strokes. Intriguingly, GLP-1 receptors are also expressed in the central nervous system, and preclinical models hint at direct neuroprotective properties, including reduced excitotoxicity, less neuroinflammation and diminished apoptotic signalling. SGLT2 inhibitors, by contrast, deliver their cardiovascular and renal benefits largely through osmotic diuresis, natriuresis and improved cardiac loading, mechanisms that may not target the atherosclerotic and thromboembolic pathways most relevant to ischemic stroke. Some researchers have even hypothesized that SGLT2 inhibition raises hematocrit and blood viscosity, potentially offsetting cerebrovascular gains.

The clinical implications are nuanced rather than revolutionary. The authors suggest that GLP-1 receptor agonists may reasonably be considered part of a stroke-prevention strategy in type 2 diabetes, but they are emphatic that the established benefits of SGLT2 inhibitors and DPP-4 inhibitors for heart failure, kidney disease and other outcomes must remain central to prescribing decisions. Treatment choices, they argue, should be informed by the totality of evidence for each individual patient, weighing the distinct benefit profiles of each class. Because the study drew on Sweden’s universal healthcare system, the findings are most applicable to settings with similar care structures, and the cohort, restricted to people newly registered from 2013 onward, may not fully represent those with long-standing diabetes.

For a field wrestling with how to translate landmark cardiovascular outcome trials into everyday practice, this study offers a methodological template as much as a clinical finding. By embracing the time-varying chaos of real-world care rather than freezing patients at their baseline characteristics, the Swedish team produced observational estimates that line up with the randomized evidence, resolving a discordance that has persisted across dozens of published syntheses. Future work, the authors note, should focus on how these findings can feed into personalized glucose-lowering strategies that optimize stroke prevention, a goal that becomes more attainable when trial data and real-world data finally tell the same story.

Subject of Research: Effects of SGLT2 inhibitors, GLP-1 receptor agonists, and DPP-4 inhibitors on stroke risk in type 2 diabetes

Article Title: The effect of SGLT2 inhibitors, GLP-1 receptor agonists, or DPP-4 inhibitors on stroke risk in type 2 diabetes: a nationwide longitudinal cohort study in Sweden with contextualisation analysis

Article References: The effect of SGLT2 inhibitors, GLP-1 receptor agonists, or DPP-4 inhibitors on stroke risk in type 2 diabetes: a nationwide longitudinal cohort study in Sweden with contextualisation analysis. (n.d.). https://doi.org/10.1016/j.eclinm.2026.104220

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104220

Keywords: type 2 diabetes, stroke, GLP-1 receptor agonists, SGLT2 inhibitors, DPP-4 inhibitors, Swedish National Diabetes Register, marginal structural models, ischemic stroke, hemorrhagic stroke, pharmacoepidemiology, cardiovascular outcomes, real-world evidence

The study, published in EClinicalMedicine by Yan Xie, Taeyoung Choi, and Ziyad Al-Aly of the VA St. Louis Health Care System, drew on the electronic health records of the U.S. Department of Veterans Affairs, an integrated system spanning 172 health systems, 142 hospitals, and 1,241 outpatient sites. The researchers identified 1,401,492 participant-trials among veterans who attended an in-person primary care visit between September 15, 2025, and February 28, 2026, the period when the seasonal vaccine was widely available across the VA. Within this cohort, 526,350 participant-trials involved vaccination at the index visit and 875,142 involved no vaccination at that visit or during follow-up, and the cohort collectively contributed more than 635,000 person-years of observation, with a median follow-up of roughly six months.

What sets this analysis apart from the interim estimates issued by surveillance networks in England, Canada, and the United States is its design. Most seasonal vaccine effectiveness figures come from test-negative case-control studies, which compare vaccination rates between patients who test positive and those who test negative for influenza among people seeking care for acute respiratory illness. These designs efficiently control for care-seeking and testing behavior but estimate only relative protection among tested patients. The VA team instead emulated a target trial, a framework that asks what would happen if eligible people were, hypothetically, randomized to vaccination or no vaccination at a real clinical encounter. Because vaccination occurs throughout the season while influenza activity rises and falls, the investigators did not run a single trial spanning the entire season. They instead emulated 24 sequential 7-day trials, assigning vaccination status at the start of each window and following vaccinated and unvaccinated participants concurrently so that both groups faced the same circulating viruses at the same calendar time.

This structure was engineered to blunt two well-known distortions in observational vaccine research. Immortal time bias arises when follow-up time during which a person cannot experience the outcome is misattributed to one group, and healthy-vaccinee bias arises because people who choose to get vaccinated tend to be healthier and more engaged with care than those who do not. To reduce the latter, the team restricted eligibility to veterans with a primary care visit in the previous 18 months, guaranteeing comparable engagement with the health system, and excluded participant-visits involving acute illness, recent hospitalization, end-of-life care, advanced organ failure or cancer, recent surgery, and very high predicted short-term risks of hospitalization or death. The result was a comparison between people who plausibly could have been vaccinated and people who plausibly could have remained unvaccinated at the same visit.

Statistical refinement went further. The researchers applied marginal structural models with inverse-probability weighting, estimating propensity scores from a comprehensive covariate set that included demographics, comorbidities, frailty indices, laboratory values, healthcare utilization, vaccination histories across five prior influenza seasons, COVID-19 vaccination records, and algorithmically selected high-dimensional variables drawn from diagnoses, medications, laboratory results, and procedures. Veterans in the unvaccinated group who later received the vaccine were censored, with time-varying inverse probability of censoring weights correcting for the informative nature of that departure. Outcomes were strictly defined by laboratory confirmation: an influenza-associated emergency department or urgent care visit within one day of a positive test, or a hospitalization within 30 days of a positive test.

The headline numbers were sobering but far from discouraging. Vaccination was associated with a vaccine effectiveness of 22.25 percent against laboratory-confirmed influenza-associated emergency department visits, 31.84 percent against influenza-associated hospitalizations, and 21.95 percent against the composite of the two outcomes. In absolute terms, vaccination prevented approximately five emergency department visits and one hospitalization per 10,000 vaccinated persons over the follow-up period. Protection appeared consistent across time windows, with effectiveness against the composite outcome of roughly 21 to 24 percent during the intervals from 31 to 120 days after vaccination, and it held across prespecified subgroups defined by age, split at 65 years, and by immunocompromised status.

To gauge how much the subclade K mismatch mattered, the researchers ran the identical design on the 2024–2025 season. The contrast was clear: effectiveness in the prior season reached 31.03 percent against emergency department visits, 48.57 percent against hospitalizations, and 32.15 percent against the composite outcome. The ratio between seasons worked out to roughly 0.7, meaning relative effectiveness in 2025–2026 was about 30 percent lower than in the year before, a statistically significant decline consistent with the antigenic drift of subclade K. Yet the direction of benefit remained firmly positive, echoing interim estimates from the CDC VISION network, which reported about 30 percent protection against adult influenza-associated hospitalization, and from England and Canada, which found meaningful if reduced protection against medically attended H3N2 illness.

The authors deployed an unusually thorough robustness program. Thirteen sensitivity analyses varied the follow-up horizon, cohort exclusion criteria, outcome definitions, target populations, model specifications, weight truncation thresholds, grace periods, and landmark exposure definitions, and the estimates held steady throughout. Negative outcome control analyses provided an additional check for residual confounding: vaccination showed no association with all-cause hospitalization within 14 days of follow-up or all-cause death within the same window, exactly the null results expected if the design were not systematically biased. These controls cannot prove the absence of confounding, but they strengthen confidence that the observed protection reflects a genuine vaccine effect rather than an artifact of who chooses to be vaccinated.

Limitations remain. The VA population skews older, whiter, and more male than the general United States population, which may limit generalizability. Outcomes required laboratory confirmation within VA data, so infections tested or treated elsewhere were likely missed, meaning the absolute reductions probably represent a lower bound on the vaccine’s real-world impact. Vaccinations received outside the VA system may also have been incompletely captured, a form of misclassification that would tend to bias effectiveness estimates downward. The study did not evaluate influenza-associated death, because event counts were too small for adequate statistical power, nor did it estimate effectiveness against specific viral subtypes separately.

The public health message, however, comes through clearly. Even in a season when the predominant virus drifted away from the vaccine strain after strain selection, vaccination continued to reduce emergency department visits and hospitalizations attributable to laboratory-confirmed influenza. Given that seasonal influenza causes an estimated 9.4 million to 51 million illnesses, 120,000 to 710,000 hospitalizations, and 6,300 to 52,000 deaths annually in the United States, absolute reductions of even a few events per 10,000 people translate into substantial averted morbidity across the tens of millions of doses administered each year. The findings argue against letting an imperfect antigenic match undermine vaccination campaigns, and they make the case for timely, season-specific effectiveness monitoring using complementary designs. As the authors conclude, the 2025–2026 vaccine was less effective than its predecessor by roughly 30 percent in relative terms, but it still did its most important job: keeping veterans out of emergency departments and hospital beds during a challenging influenza season.

Subject of Research: Real-world effectiveness of the 2025–2026 seasonal influenza vaccine against influenza-associated emergency department visits and hospitalizations among U.S. veterans during an H3N2 subclade K-dominated season

Article Title: Effectiveness of the 2025–2026 seasonal influenza vaccine among U.S. veterans: an observational study

Article References: Xie, Y., Choi, T., & Al-Aly, Z. (2026). Effectiveness of the 2025–2026 seasonal influenza vaccine among U.S. veterans: an observational study. eClinicalMedicine, 100, Article 104183. https://doi.org/10.1016/j.eclinm.2026.104183

Image Credits: AI Generated

DOI: 10.1016/j.eclinm.2026.104183

Keywords: influenza vaccine, vaccine effectiveness, H3N2 subclade K, target trial emulation, Veterans Affairs, antigenic drift, hospitalization, emergency department, observational study, marginal structural models, 2025–2026 season, public health