Association between clinically recorded alcohol consumption and initial presentation of 12 cardiovascular diseases: population based cohort study using linked health records

Objectives To investigate the association between alcohol consumption and cardiovascular disease at higher resolution by examining the initial lifetime presentation of 12 cardiac, cerebrovascular, abdominal, or peripheral vascular diseases among five categories of consumption. Design Population based cohort study of linked electronic health records covering primary care, hospital admissions, and mortality in 1997-2010 (median follow-up six years). Setting CALIBER (ClinicAl research using LInked Bespoke studies and Electronic health Records). Participants 1 937 360 adults (51% women), aged ≥30 who were free from cardiovascular disease at baseline. Main outcome measures 12 common symptomatic manifestations of cardiovascular disease, including chronic stable angina, unstable angina, acute myocardial infarction, unheralded coronary heart disease death, heart failure, sudden coronary death/cardiac arrest, transient ischaemic attack, ischaemic stroke, intracerebral and subarachnoid haemorrhage, peripheral arterial disease, and abdominal aortic aneurysm. Results 114 859 individuals received an incident cardiovascular diagnosis during follow-up. Non-drinking was associated with an increased risk of unstable angina (hazard ratio 1.33, 95% confidence interval 1.21 to 1.45), myocardial infarction (1.32, 1.24 to1.41), unheralded coronary death (1.56, 1.38 to 1.76), heart failure (1.24, 1.11 to 1.38), ischaemic stroke (1.12, 1.01 to 1.24), peripheral arterial disease (1.22, 1.13 to 1.32), and abdominal aortic aneurysm (1.32, 1.17 to 1.49) compared with moderate drinking (consumption within contemporaneous UK weekly/daily guidelines of 21/3 and 14/2 units for men and women, respectively). Heavy drinking (exceeding guidelines) conferred an increased risk of presenting with unheralded coronary death (1.21, 1.08 to 1.35), heart failure (1.22, 1.08 to 1.37), cardiac arrest (1.50, 1.26 to 1.77), transient ischaemic attack (1.11, 1.02 to 1.37), ischaemic stroke (1.33, 1.09 to 1.63), intracerebral haemorrhage (1.37, 1.16 to 1.62), and peripheral arterial disease (1.35; 1.23 to 1.48), but a lower risk of myocardial infarction (0.88, 0.79 to 1.00) or stable angina (0.93, 0.86 to 1.00). Conclusions Heterogeneous associations exist between level of alcohol consumption and the initial presentation of cardiovascular diseases. This has implications for counselling patients, public health communication, and clinical research, suggesting a more nuanced approach to the role of alcohol in prevention of cardiovascular disease is necessary. Registration clinicaltrails.gov (NCT01864031).


Overview of major studies on alcohol consumption and CVD
There are multiple systematic reviews and meta-analyses of the association between alcohol consumption and aggregated CVD 1-7 as well as CVD biomarkers. 4,8,9 Most have shown that moderate levels of alcohol intake are associated with a lower risk of CVD morbidity and mortality, as well as more favourable cardiovascular health profiles in general, than non-drinkers. However, there is a growing scepticism around this observation, with a series of recent commentary pieces pointing out a number of methodological shortcomings in the evidence that the U-shape is based on. [10][11][12] These include failure to have decomposed the current non-drinking group into life-long abstainers, former drinkers and those who drink on an occasional (but not weekly basis). It is known that former drinkers have an increased risk of CVD mortality 13 compared to life-long non-drinkers; therefore combining these two groups is likely to lead to the protective effects of moderate drinking being over estimated in being compared to a non-drinking group that consists of former drinkers, who may have quit for health reasons. Similarly, it has been shown that the onset of ill health is associated with a reduction in regular alcohol consumption to drinking on an occasional basis, 14 therefore combining these individuals with non-drinkers also introduces bias.
Evidence from short-term alcohol feeding interventions has shown that moderate drinking is related to higher levels of high-density lipoprotein cholesterol (HDL-C) and adiponectin, as well as lower levels of fibrinogen, for which it has been argued reflects indirect evidence of a causal association of moderate alcohol conferring a lower risk of experiencing CVD outcomes. 8 However, findings from large scale Mendelian randomisation studies as well as pharmacological trials suggest that the causal role of these biomarkers in the development of CVD is at best unclear. [15][16][17][18][19] Furthermore, a recent individual participant data Mendelian randomisation analysis of the association alcohol consumption (using the rs1229984 variant in the alcohol dehydrogenase 1B gene (ADH1B) as an instrumental variable) found that individuals with a genetic variant which predisposes them to low-level alcohol intake or non-drinking actually had a lower risk of CHD and a generally better cardiovascular health profile than those who did not carry the genetic variant. 20 This is counter to the association typically found in observational studies. However, it should be noted that in a related paper from the same consortium, that non-linear associations were found for alcohol intake and a range of biomarkers associated (but not necessarily causally related) with CVD including non-HDL cholesterol, BMI, waist circumference and C-reactive protein, but not for HDL-C, triglycerides or interleukin- 6. 21 Given differences in the association of alcohol consumption and different CVD biomarkers, it could be hypothesised that moderate alcohol consumption may be protective for some CVDs but not others. 22 However, the evidence base for specific CVD phenotypes is sparse in comparison to that of aggregated CVD outcomeswith the majority of research focussing on acute myocardial infarction or stroke (total and broad categories of ischaemic or haemorrhagic). There have been calls for further research into the association of alcohol consumption and deeper phenotypes of CVD 23 to further our understanding of the role of alcohol consumption in the development or prevention of individual CVDs in order to improve risk prediction at both an individual and population level. However, few studies are sufficiently powered to examine individual CVDs and fewer still are in a position whereby they are also able to disaggregate the current non-drinker group into non-drinkers, former drinkers and occasional drinkers. We provide an overview of research from major investigator led prospective observational studies as well as meta-analyses of the topic of alcohol consumption and specific CVDs in Table A. As noted above, there are a number of existing studies for primary endpoints such as myocardial infarction and ischaemic stroke, with meta-analyses and Mendelian randomisation studies also available for these topics. However, there is a substantially smaller evidence base for other outcomes such as heart failure, cardiac arrest/sudden coronary death, angina, transient ischaemic attack, peripheral arterial disease, abdominal aortic aneurysm, unheralded CHD death and haemorrhagic stroke subtypes (subarachnoid or intracerebral haemorrhage), and many of the investigator led cohort studies have in some manner combined different current non-drinking groups into an aggregate category. 24 In pooled analyses the association between each nondrinking category and risk of initially presenting with cardiac arrest/SCD compared to moderate drinking was nonsignificant. In in contrast to previous studies in men, we did not find a linear doseresponse association between alcohol intake and lower risk of SCD, in fact, we observed the opposite. We found that heavy drinkers have a much larger risk of initially presenting with Intracerebral haemorrhage than moderate drinkers after excluding incurrent cardiovascular conditions.

Peripheral arterial disease
Yes 59 Never, former and occasional drinkers were combined.

No No
In US male physicians, increasing alcohol consumption was associated with a lower risk of PAD; however, this was compared to a mixed reference group and We observed elevated risk of initially presenting with PAD amongst all types of nondrinker in comparison to moderate drinkers. We are also the first the report the did not specifically separate heavy drinkers from those who drink in moderation.
association between heavy drinking and PAD, finding that heavy drinkers are more likely to initially present with PAD than their moderate drinking counterparts.

Abdominal aortic aneurysm
Yes [60][61][62] In largest and most recent study, never and former drinkers separated but occasional drinkers combined with moderate.

No No
Moderate alcohol consumption was associated with a lower hazard of AAA. The association between higher doses of alcohol (> 120g and 60g per week for men and women, respectively) and risk of AAA remain unknown.
We are the first to report on the association between heavy drinking and AAA, finding no significant difference in risk of initial presentation between moderate and heavy drinkers.
a Major defined as a prospective observational study with a sample size ≥ 10,000 participants and validated clinical events. b Defined as assessing the first presentation in the absence of intercurrent other cardiovascular disease.

Study description
We included 1,937,360 anonymised patients from the CALIBER (CArdiovascular research using LInked Bespoke studies and Electronic health Records) programme. 63 The cohort used in the present study was drawn from patients registered with Clinical Practice Research Datalink (CPRD) 64 general practices in England that consented to data linkage (approximately 5% of the UK population).
We used an open cohort design, where participants joined the cohort when they met the inclusion criteria at any point between 1st January 1997 and 25th March 2010 (the last CPRD data submission).
Patients were included in cohort if they were aged ≥ 30 years, had at least one year of electronic health record data which met CPRD data quality standards, and had no record indicating any cardiovascular disease prior to study entry. Patients for who sex was not recorded and those pregnant within six months of the eligibility date were also excluded. Patients were followed up until the date of an initial presentation of one of our cardiovascular endpoints or were censored on the date of leaving the practice/last data submission from their practice. Patients who died before 1st January 2001 were excluded as cause-specific mortality data was not available for them (see study flow diagram; Figure A

Assessment of alcohol consumption in CALIBER
Self-reported alcohol consumption was collected prospectively and coded by general practitioners or practice nurses on the consultation date in CPRD. The most recent alcohol consumption record in the five years before entry into the study was used to classify participants drinking behaviour. In light of current debates on the U/J-shaped relationship observed between alcohol consumption and aggregated CVD outcomes 10 five drinking categories were defined including: (1) non-drinkers (Read 66 codes such as "tee-total" and "non-drinkers"), former drinkers (those with codes for "stopped drinking alcohol" and/or "ex-drinker"), occasional drinkers (those with codes for "drinks rarely" and/or "drinks occasionally"), current moderate drinkers (those who had a code for current alcohol consumer and an indicator of whether they drank within daily [32g or 24g of alcohol for men and women respectively] and/or weekly [168g of alcohol for men and 112g for women] recommended sensible drinking limits for the UK at the time of observation 69 ) and current heavy drinkers (defined as those who exceeded daily and/or weekly sensible drinking limits). We also utilised data fields with information entered on daily and/or weekly amount of alcohol consumed to define participants as non-drinkers, moderate drinkers (drank within daily and/or weekly guidelines) and heavy drinkers. Weekly alcohol data was available as a continuous variable, so we were able to classify consumption using standard thresholds (outlined above and in Figure B). Data on daily alcohol intake was entered using categories of: (1)  .00], for which we defined moderate drinking as anything >1 UK unit but less than 3 (women) or 7 (men) UK units. Unfortunately information on binge drinking was only available for a select minority of the cohort (~100 people) therefore a separate category for this drinking behaviour was not defined (but these patients were coded as heavy drinkers). We reclassified non-drinkers as former drinkers if they had any record of drinking recorded in their entire clinical record entered on CPRD prior to study entry (in cases whereby non-drinkers had no record of drinking before entering the study we assumed that they were not former drinkers). This resulted in 19,853 (out of 184,747; 10.7%) non-drinkers being recoded as former drinkers, a further 6,826 (3.7%) participants were reclassified through having a positive history of alcohol abuse. A flow diagram outlining our coding algorithm is presented in Figure B and the exact Read codes used to define drinking categories are presented in Table B. We provide analyses of the association of these drinking categories with cardiovascular traits as means of validating the clinically recorded alcohol consumption measure we derived using CPRD data in Figure C (with adjustments for age, age-squared and sex). Consistent with observational studies we found that higher alcohol intake was associated with increased levels of HDL-C and total cholesterol as well as elevated blood pressure. 20 We also observed that heavy drinkers were more likely to be smokers. Diabetes was more prevalent in non, former and occasional drinkers at baseline, consistent with the sick-quitter 70 and sick-non-starter 71 hypotheses. Higher alcohol intake was also associated with lower odds of prevalent diabetes and HbA1c levels which is concordant with crosssectional studies 72,73 Increased alcohol consumption was also associated with lower levels of creatinine, 74 but non-and former-drinkers did not have elevated levels compared to moderate drinkers. Non-drinkers, former drinkers and occasional drinkers all had higher BMI than moderate drinkers whereas heavy drinkers did not significantly differ from those who drank in moderation.

Figure B -Alcohol coding algorithm in CALIBER
Higher levels of alcohol intake were also associated with higher levels of gamma-glutamyl transferase. 75

Figure C -Association of clinically recorded alcohol consumption categories and cardiovascular related traits
Overview of codes/data sources used to define each cardiovascular endpoint 15

Overview of codes/data sources used to define each cardiovascular endpoint
As outlined in the main text we used multiple endpoints based on the first recorded diagnosis of the 12 most common symptomatic CVD manifestations, including: chronic stable angina (SA), unstable angina (UA), acute myocardial infarction (MI), unheralded coronary heart disease death (UCD), heart failure (HF), cardiac arrest/sudden cardiac death (SCD), transient ischaemic attack (TIA), ischaemic and haemorrhagic (subarachnoid and intracerebral haemorrhages) stroke, peripheral arterial disease (PAD), and abdominal aortic aneurysm (AAA). We were interested in only the first occurrence of any CVD, so subsequent events (e.g. ischaemic stroke occurring after TIA) were not analysed.
Diagnoses could occur in primary or secondary care, or at death, and were defined/validated using multiple sources; including a combination of symptoms, diagnoses (including the use of additional information from ECG findings and troponin values) and medication prescriptions. There is an extensive literature demonstrating that CPRD patients are representative of the UK population in terms of age, gender, ethnicity and overall mortality as well as the validity of risk factor and disease endpoints defined using multiple electronic health record (EHR) sources. [76][77][78][79][80][81][82][83][84] We have further confidence in the validity of linked EHR in having seen in this study, and others, 85-89 that associations were broadly similar when restricting analysis to data sources; each of which will have different health care practitioners who are responsible for entering diagnoses.
Details of the data sources and relevant codes used to define each endpoint are outlined in Table C.

Multiple imputation
Multiple imputation was carried out using the mi command in the statistical package Stata, to replace missing values in exposure and risk factor variables under a missing at random assumption. Imputation models were estimated separately for men and women and included: 6. The Nelson-Aalen hazard and the event status for each endpoint analysed in the data. 90 Non-normally distributed variables were log-transformed for imputation and exponentiated back to their original scale for analysis. Five multiply imputed datasets were generated, and Cox proportional hazard models were fitted to each dataset. Coefficients were combined using Rubin's rules. 91 We checked whether the imputations were plausible by comparing plots of the distribution of observed and imputed values.  3  3  3  4  4   4  3  3  3  3   3  3  3  3  3   2  2  2  2  2   1  2  2  2  3  1  1  1  1  2  1  0  1

Supplementary analyses
We also calculated hazard ratios and 95% confidence intervals for the association of alcohol consumption and different CVDs adjusting for age and sex only.
In secondary analyses, we additionally adjusted for systolic blood pressure, diabetes mellitus status and body mass index which were not included as standard due to concerns of overadjustment bias given it is likely that they lie on the causal pathway between alcohol consumption and CVDs. 92 We tested for effect modification by gender via including interactions between drinking category and sex for each endpoint.
In sensitivity analyses we re-analysed data (1) ignoring diagnoses obtained using primary care data, (2) limited to fatal endpoints only, and (3) using information collected after 2004 when financial incentives were introduced for recording patient data on alcohol consumption. We also compared findings obtained using imputed and complete case methods. We carried out a series of post-hoc analyses within subgroups defined by smoking status, BMI and diabetes.
Our referent category in all models was moderate drinkers. 93 All analyses, here and in the main paper, were conducted using Stata v14.