Nasir, Khurram

Abstract Aims Alcohol and its relationship with cardiovascular disease (CVD) remain contentious. Several observational studies have found cardioprotective associations, while Mendelian randomization studies have found null or harmful effects of moderate drinking. Biases, including abstainer misclassification and sick-quitter effects, may obscure true associations. It is widely understood that former drinkers are often sick quitters. However, removal of former drinkers from analyses addresses bias in the abstainer group but fails to account for induced survivorship among current drinkers. This study sought to determine whether adjusting for biases altered the association between alcohol and CVD by comparing conventional models with those using (i) occasional drinkers as the reference group instead of lifetime abstainers and (ii) ‘intention-to-treat’ reallocation of former drinkers based on past drinking patterns. Methods and results The Multi-Ethnic Study of Atherosclerosis is a prospective cohort study of 6814 participants aged 45–84 years between 2000 and 2002 without baseline CVD. A total of 6755 participants were included after excluding those with missing data. Alcohol consumption was self-reported and classified as lifetime abstainer, former, or current drinker [occasional (≤1 drink/week), light (2–7 drinks/week), moderate (8–14 drinks/week), heavy (>14 drinks/week)]. Outcomes included coronary artery calcium (CAC; 0, 1–100, 101–300, and >300) and clinical events [myocardial infarction (MI), stroke, cardiovascular mortality, composite major adverse cardiovascular events (MACEs), and heart failure]. Of 6755 participants, 20.6% were lifetime abstainers, 24.1% former drinkers, and 55.3% current drinkers. Adjusting for biases altered the observed relationship between alcohol and CVD, showing increased harm with light drinking, decreased protection with moderate drinking, and a dose–response relationship between alcohol and CAC, particularly CAC >300 in adjusted models. Protective associations between moderate drinking and MI [hazard ratio (HR): 0.64, 95% confidence interval (CI): 0.38–1.06] and cardiovascular mortality (HR: 0.81, 95% CI: 0.53–1.25) were attenuated and became non-significant. Light drinking shifted from null to increased risk for MI (HR: 1.46, 95% CI: 1.11–1.92), stroke (HR: 1.54, 95% CI: 1.17–2.03), MACE (HR: 1.33, 95% CI: 1.12–1.57), and heart failure (HR: 1.42, 95% CI: 1.10–1.84). Conclusion Associations between alcohol and CVD in observational studies are sensitive to methodologic variation. Results from models correcting for abstainer/sick-quitter biases may help explain discrepant results from previous studies.

Objective Meaningful assessments of how large language models (LLMs) incorporate clinical guidelines require large-scale testing over many queries. Here, we evaluate the prevalence of clinical guideline omissions and hallucinations in a large sample of diagnostic LLM outputs. Methods We used simulated case vignettes and zero-shot prompting to generate diagnostic outputs and rationales from GPT-4.1 and DeepSeek-V3. English case vignettes were created for hypercholesterolaemia and type-2 diabetes mellitus. Each vignette contained identical medical information, while sociodemographic characteristics varied in terms of sex, ethnicity and location. We calculated the prevalence of existing and hallucinated clinical guidelines in LLM outputs across disease, LLM and sociodemographic characteristics. Results We analysed a total of 12 197 LLM outputs, which quantifies three hazard areas: omissions (up to 97% for DeepSeek-V3 and 46% for GPT-4.1), hallucinations (up to 9%) and inconsistencies (guideline citation rate ranging from 0% to 78.39% across sociodemographic vignettes). Omission and hallucination rates were generally similar across vignettes with different sex or ethnicity data, yet were particularly sensitive to patient location. Discussion This study highlights significant variability in clinical guideline prediction across two different diseases, three different sociodemographic variables and two LLMs, even when the LLMs were instructed by identical prompts, establishing clinical guideline prediction in LLM outputs as a stochastic event. Conclusion The stochastic nature of LLMs creates a unique challenge for evidence generation and clinical deployment. Being able to measure and capture this stochasticity within high-quality research designs will be a prerequisite to advancing the responsible deployment of LLMs in healthcare.

Abstract Background Alcohol consumption and its relationship with cardiovascular disease (CVD) remains controversial. Observational studies have reported J-shaped associations suggesting cardioprotective effects of moderate drinking, while other observational and Mendelian Randomization studies have found dose-response harm. Biases such as abstainer misclassification, sick quitter effects, and residual confounding may obscure true associations. This study aimed to reassess alcohol’s impact on CVD using novel methods to minimize bias. Purpose We evaluated the relationship between alcohol and subclinical (coronary artery calcium [CAC]) and clinical (myocardial infarction [MI], stroke, cardiovascular death, major adverse cardiovascular events [MACE], and heart failure) outcomes in the Multi-Ethnic Study of Atherosclerosis (MESA). We compared conventional models with those using (1) occasional drinkers as the reference group instead of lifetime abstainers and (2) intention-to-treat (ITT) reallocation of former drinkers based on past drinking patterns. Methods MESA is a prospective cohort study of 6,814 participants aged 45-84 years without baseline CVD. Alcohol use was classified as lifetime abstainer, former drinker, or current drinker (occasional [≤1 drink/week], light [2-7 drinks/week], moderate [8-14 drinks/week], heavy [>14 drinks/week]). We used multinomial logistic regression for CAC and Cox proportional hazards models for clinical outcomes across four models: (1) non-drinker reference, (2) lifetime abstainer reference, (3) occasional drinker reference, and (4) occasional drinker reference with ITT reallocation of former drinkers to their previous drinking levels. Results Increasing alcohol consumption correlated with a dose-response increase in CAC. Heavy drinking was consistently associated with severe CAC (>300). In fully adjusted models, light (OR: 1.23, 95% CI: 1.02-1.48), moderate (OR: 1.27, 95% CI: 1.04-1.55), and heavy drinking (OR: 1.42, 95% CI: 1.12-1.78) were associated with higher CAC scores compared to occasional drinkers. Traditional models replicated protective effects of moderate drinking for MI and cardiovascular mortality. However, these associations attenuated in adjusted models. Moderate drinking showed an attenuated, non-significant benefit for MI (HR: 0.64, 95% CI: 0.38-1.06) and cardiovascular mortality (HR: 0.81, 95% CI: 0.53-1.25) in fully adjusted models. Light drinking shifted from null to increased risk for MI (HR: 1.46, 95% CI: 1.11-1.92), stroke (HR: 1.54, 95% CI: 1.17-2.03), MACE (HR: 1.33, 95% CI: 1.12-1.57), and heart failure (HR: 1.42, 95% CI: 1.10-1.84). Conclusion Adjusting for abstainer and sick quitter biases altered the observed relationship between alcohol and CVD, showing increased harm with light drinking, attenuated protective effects of moderate drinking, and a dose-response relationship for CAC in bias-adjusted models. These findings illustrate the importance of methodology in alcohol-CVD relationships.

Background Although cardiovascular mortality (CVM) rates in the United States have been declining overall, our study evaluated whether this holds true for areas with increased social deprivation. Methods and Results We used county‐level cross‐sectional age‐adjusted CVM rates (aa‐CVM) (2000–2019) linked to the Centers for Disease Control and Prevention Social Vulnerability Index (SVI–2010). We grouped counties as per SVI (Groups I 0–0.2, II 0.21–0.4, III 0.41–0.6, IV 0.61–0.8, and V 0.81–1) and calculated the relative change in the aa‐CVM between 2000 to 2003 and 2016 to 2019. We used adjusted linear regression analyses to explore the association between a higher SVI and temporal aa‐CVM improvement; we studied this temporal change in aa‐CVM across subgroups of race, sex, and location. The median aa‐CVM rate (per 100 000) was 272.6 (interquartile range [IQR]: 237.5–311.7). The aa‐CVM was higher in men (315.6 [IQR: 273.4–363.9]) than women (221.3 [IQR: 189.6–256.7]), and in Black residents (347.2 [IQR: 301.1–391.1]; P <0.001) than White residents (258.9 [IQR: 226–299.1]; P <0.001). The aa‐CVM for SVI I (233.6 [IQR: 214.8–257.0]) was significantly lower than that of group V (323.6 [IQR: 277.2–359.2]; P <0.001). The relative reduction in CVM was significantly higher for SVI group I (32.2% [IQR: 24.2–38.4]) than group V (27.2% [IQR: 19–34.1]) counties. After multivariable adjustment, a higher SVI index was associated with lower relative improvement in the age‐adjusted CVM (model coefficient −3.11 [95% CI, −5.66 to −1.22]; P <0.001). Conclusions Socially deprived counties in the United States had higher aa‐CVM rates, and the improvement in aa‐CVM over the past 20 years was lower in these counties.