Written by: Romit BhattacharyaChristopher S MarnellSo Mi Jemma ChoAniruddh P PatelYunfeng RuanSatoshi KoyamaAmanda R JowellMark TrinderSara HaidermotaKim LanneryMichael C HonigbergSeyedeh M ZekevatIda SurakkaGina M PelosoPradeep Natarajan
Find these authors on Roon
AP
Dr. Aniruddh Patel
General Cardiology · MGH
PN
Dr. Pradeep Natarajan
Preventive Cardiology · MGH
Introduction
Prior work from the 20th century concluded that >90% of the risk of first myocardial infarction (MI) could be attributed to nine modifiable risk factors – hypertension, diabetes, smoking, waist-hip ratio, dyslipidemia, alcohol intake, dietary pattern, physical activity, and psychosocial stressors1. It has served as a core piece of evidence linking social and behavioral factors to heart disease, but occurred prior to the advent of contemporary biomarkers and large-scale genomics assays, potentially discounting their importance. Indeed, a recent pooled analysis of 112 studies assessed the population attributable risks (PARs) of traditional risk factors to be far lower, and highlighted a need to include biomarkers and genetic risk scores in future analyses2.Since the 1990s when INTERHEART data were collected, the attributable risk of cardiovascular disease (CVD) to various risk factors has shifted. Global populations have witnessed dramatic increases in prevalence of cardiometabolic disease, due to increasing prevalence of obesity and cardiometabolic risk factors3–7. Simultaneously, widespread use of preventive medications such as statins and improved management of modifiable risk factors for coronary artery disease (CAD) may be acting to reduce the influence of traditional risk factors8. Additionally, contemporary biomarkers such as lipoprotein(a) (Lp(a)), high sensitivity C-reactive protein (hsCRP), and polygenic risk scores for CAD (PRSCAD) may play a larger role as traditional risk factors are better managed in the population 9, 10.In particular, understanding of the inherited basis of CAD has significantly advanced over the last two decades 11–17. Though CAD was previously observed to be enriched in families, monogenic etiologies captured as familial hypercholesterolemia (FH) variants do not sufficiently capture the extent of empiric familial aggregation. PRSCAD have been shown to refine risk of incident CAD beyond family history and are particularly impactful in refining risk for younger individuals 15, 18–20.Using data from the UK Biobank and validation data from the Mass General Brigham Biobank (MGBB), we sought to update our understanding of the contributions of traditional modifiable risk factors alongside novel and genetic risk factors in a contemporary cohort. We hypothesized that in this population of middle-aged individuals, contemporary biomarkers and genetic factors contribute a substantial portion of risk for incident CAD.
Methods
The UK Biobank (UKBB) is a population-based cohort of 502,629 participants from the ages of 40 to 69 years at baseline recruited during years 2006–2010 21. At the baseline study visit, participants provided demographic information, medical history, family history, medication use, and lifestyle information. Due to withdrawn consent, relatedness, or missing CAD status, 36,169 individuals were excluded. 15,221 individuals were excluded for lack of genetics consent or failure of quality control during sequencing. 116,374 individuals in the UKBB used in derivation and training of the PRSCAD were excluded to minimize overfitting. Participants with prevalent CAD, prior myocardial infarction or revascularization were excluded as well (n=13,989). Finally, 21,169 individuals were excluded for missing Lp(a) laboratory values. The final analysis cohort comprised 299,707 individuals (Figure 1). As a validation cohort, a subset of the MGBB – a hospital-based cohort in the United States - with genomic and biomarker data available was studied to confirm major findings. All study participants provided written informed consent. Secondary use of data for this study was approved by the Massachusetts General Hospital Institutional Review Board (protocol 2021P002228) facilitated through UKB application 7089. All data are available to registered users through the UKBB.
Baseline characteristics included sex (field code 31) and age at recruitment (field code 21022). Exposures encompassed both clinical and genetic factors. Clinical factors were defined as laboratory data, medical diagnoses, anthropometric measures, and smoking. Laboratory data were generated by UKBB at study enrollment and included hsCRP, Apolipoprotein B (ApoB), Apolipoprotein AI (ApoAI), Lp(a), total cholesterol, triglycerides, high-density lipoprotein (HDL) cholesterol, and low-density lipoprotein (LDL) cholesterol (field codes 30710, 30640, 30630, 30790, 30690, 30870, 30760, and 30780, respectively). Medical diagnoses of diabetes and hypertension were linked to the UKBB via EHR with ICD10 diagnostic codes (I10-I15). Dyslipidemia was defined as ApoB:ApoA1 ratio, with adjustment for statin use in the regression model, but not used for defining the condition. Anthropometric measures such as waist circumference and hip circumference were ascertained at study intake (field codes 21001, 48, 49). Additional information regarding the technical aspects of the serum assays is available in the Supplemental Methods section.
Exercise frequency, smoking habits, and Townsend deprivation index were assessed using survey data at study enrollment. Dietary pattern was scored in accordance with the Alternate Healthy Eating Index (AHEI) as previously described22. For normally distributed variables with <5% missing values such as ApoA1, hemoglobin A1c, hsCRP and total weekly metabolic equivalents of exercise, missing values were imputed using the R function predict() in the package ‘stats’ (base R) based on age and sex variables and then compared to the non-imputed, incomplete variable sets, to confirm similar results were not due to informative missingness.
The multi-ancestry genome-wide polygenic score for CAD (GPSmult) was used as the PRSCAD in this study. GPSmult was constructed using LDPred2, incorporating the weighted effects of over 1.2 million SNPs from 51 GWAS for CAD (>269,000 cases and >1,178,000 controls), other atherosclerotic diseases, and their risk factors from multi-ancestry cohorts, as previously described 23. Scores included individuals of African, East Asian, European, Hispanic and South Asian ancestries. See Supplemental Methods for additional details.
To determine HeFH status, whole exome sequencing data was available for all 299,707 participants. Pathogenic variants in LDLR, APOB, and PCSK9 were curated. Predicted loss of function variants in PCSK9 and APOB genes were excluded since these variants potentially cause hypobetalipoproteinemia. Genotype- and variant-level quality control (QC) methods were used. Stringent variant QC methods criteria were used (missingness < 1%, MAF > 0.1%). Additional details on the ascertainment of HeFH genotype are available in the Supplemental Methods section.
The primary outcome was incident CAD defined as first-time MI or coronary artery revascularization. Individuals were defined as having CAD on the basis of at least one of the following criteria, as previously reported22, 24: (1) MI, coronary artery bypass grafting, or coronary artery angioplasty documented in their medical history by a trained nurse; (2) Hospitalization for an ICD-10 code for acute MI (I21.0, I21.1, I21.2, I21.4, I21.9); (3) Hospitalization for an OPCS-4 coded procedure: coronary artery bypass grafting (K40.1–40.4, K41.1–41.4, K45.1–45.7); or (4) Hospitalization for an OPCS-4 coded procedure: coronary angioplasty with or without stenting (K49.1–49.2, K49.8–49.9, K50.2, K75.1–75.4, K75.8–75.9). Incident events were ascertained through electronic health records (EHR) linked to the UK Biobank, with follow-up truncated at 10 years.
Baseline characteristics were described as mean with standard deviation for parametric distributions, median and interquartile range for non-parametric distributions, and numeric with proportions for categorical variables.
Three models were built to predict the risk of incident MI: the Base Model (adjusting for age, sex, smoking status, lipid lowering therapy), Clinical Risk Factor Model (additionally adjusting for waist-to-hip ratio [WHR], hypertension, diabetes, diet, minutes of weekly exercise [METs], smoking, Townsend deprivation index [TDI], Apolipoprotein B-to-Apolioprotein A1 ratio [ApoB:ApoA1], and the novel biomarker/genetic model (additionally including high sensitivity C reactive protein [hsCRP]), lipoprotein(a) [Lp(a)], heterozygous familial hypercholesterolemia variants [HeFH], and a multiancestry polygenic risk score for CAD [PRSCAD]).
For each model, multivariate Cox-regression models were constructed to estimate the hazard ratio for incident MI. Proportional hazards were confirmed by plotting cumulative incidence and Schoenfeld residuals, and log transformations were performed as appropriate. Model fit was assessed using the Nagelkerke R2. The relative contribution to the model’s explanatory power was calculated by the percent loss in the R2 when the CRF was removed from the model (denoted ‘R2 loss’) thus estimating the entropy loss function as previously described25. In brief, each model’s full multivariate Cox proportional hazard model was first constructed. Then, a null density model that represented the effect of excluding each covariate was constructed, and the resultant entropy represents explained risk lost from permutation. Formulaically, the R2 is derived from logarithmic mean of the full minus null model for each predictor. Bootstrapping was performed 1000 times to estimate the 95% CIs.
Additional analyses were used to corroborate findings: calculation of the C-statistic for model fit, and calculation of the Population Attributable Risk (PAR) for overall model and individual CRF explanatory power. The C-statistic was calculated for each model using the ‘concordance’ function of the ‘survival’ package in R, and compared across models for model fit. The standard error and 95% confidence interval of the c-statistic was derived from the variance provided by the function. Improvements in model performance were compared by sequentially adding variables between the (i) base model, (ii) clinical model, and (iii) novel biomarker/genetic model.
PAR models were constructed based on multivariable logistic regression models. For these models, dichotomous risk factors are treated as ‘exposed’ or ‘unexposed’, whereas continuous variables are treated as polychotomous exposures, whereby the lowest quintile is ‘unexposed’ and quintiles 2 through 5 (i.e. any elevation in risk) is considered ‘exposed’, as is standard for PAR analyses. As was done previously in the INTERHEART study and the Global Risk Factors Consortium, this analysis aligned with other large trials to enable comparison. All PAR calculations were performed using the R package ‘attribrisk’ v0.1.4 (https://rdrr.io/github/bethatkinson/attribrisk/). Additional details are included in the Supplemental Methods. 95% confidence intervals (95% CI) were computed from a grouped jackknife estimate of variance based on k=20 groups.
All analyses were completed using R 4.1.2. Associations were deemed statistically significant if p < 0.05.
Validation analyses were conducted in the MGBB, which is a consented cohort with biobanked and clinical data from the course of routine clinical care at Mass General Brigham in Boston, Massachusetts, USA26. A subset of the cohort has additionally undergone genotyping and whole exome sequencing as previously described. This cohort provides opportunity for assessing the importance of genomic contributions to disease modeling in a cohort with a higher rate of comorbidities and disease prevalence compared to UKBB. Given that data were collected observationally in course of clinical care many elements of the full model were not available so a sparser model was constructed using age, sex, smoking status, lipid lowering therapy, systolic blood pressure, diabetes, ApoB:ApoA1 ratio, presence of heterozygous familial hypercholesterolemia variants, and polygenic risk score for CAD. Nagelkerke R2, C-statistic, and PAR were calculated as described above. These methods are further detailed in Supplemental Methods.
Results:
Among the 299,707 participants, 55.6% (166,533) reported female sex at birth, and the mean age was 56.2 (standard deviation (SD) 8.08) years. 131,305 (43.8%) reported ever smoking, 18,455 (6.2%) had a history of diabetes, and median WHR was 0.87 (interquartile range [IQR]=0.80–0.93). 92.1% (275,974) identified as White, 2.2% (6,509) as Black, 2.1% (6,256) as South Asian (Indian/Pakistani/Bangladeshi), 1.3% (3,758) as other, 0.8% (2,433) as mixed, and 0.5% (1,355) as Chinese. Baseline characteristics are summarized in Table 1.
The R2 for each model increased across the (i) base model (R2=0.021 [0.020, 0.022]), (ii) clinical model (R2=0.045 [0.043, 0.046]), and (iii) novel biomarker and genetic model (R2=0.053 [0.052, 0.055]). The C-statistic for model discrimination similarly increased across the base model (C=0.65 [0.65, 0.65]), clinical model (C=0.72 [0.71, 0.72]), and novel biomarker and genetic model (C=0.73 [0.73, 0.74]). Using the PAR analysis, the clinical model explained approximately 75% (PAR 74.9% [71.8, 78.0]), while the novel biomarker and genetic model explained 84% (PAR 84.3% [82.4, 86.3]) of the risk of first MI. Results are summarized in Table 2.
The primary measure of an individual risk factor’s contribution to the models was the loss in R2 after removal of that variable from the model. After adjustment for age, sex, cholesterol medication use and ever-smoking status, hypertension (R2 loss 15.2% [14.5, 17.1]) and PRSCAD (R2 loss 12.4% [10.8, 13.3]) were the top 2 most predictive factors.
Dyslipidemia represented by ApoB:ApoA1 ratio (R2 loss 3.4% [2.6, 3.5], presence of diabetes (R2 loss 2.2% [1.5, 2.0]), hsCRP (R2 loss 1.8% [1.5, 2.0]) and Lp(a) (R2 loss 1.5% [1.2, 1.8]) were the next most highly prioritized, all with comparable metrics. Notably, while known monogenic HeFH variants are known to have high hazard ratios for coronary disease (HR 1.27 [1.06, 1.51]), due to low prevalence in the population their influence on the model was minimal (R2 loss 0.4% [−0.3, 1.4]). Summarized results and comparable assessment of individual risk factor contribution by C-statistic and PAR are available in Table 2, Figure 2 and Supplemental Table 1.
PAR analyses were conducted as secondary analyses for ease of interpretation and comparison with prior work. Using a PAR framework, while the results were largely similar, small changes in prioritization were found in that PRSCAD had the highest PAR (32.5% [35.5–41.2]) followed by hypertension (PAR 32.5% [31.2–33.9]). WHR (PAR 18.3% [13.9–22.8]), hsCRP (PAR 16.1% [12.8–19.4]), and ApoB:ApoA1 (PAR 14.3% [11.7–16.8]) were next most highly prioritized. Broadly, there was agreement between the most highly prioritized factors between R2 loss and PAR methods. These results are summarized in Table 2 and in Figure 3.
Supplemental analyses were conducted in the MGBB, for which the baseline characteristics and comorbidities are summarized in Supplemental Table 2. Supplemental analyses used a parallel structure to the primary analyses comparing model fit and discrimination using R2, C-statistic and PAR. From the base model to the clinical model to the contemporary/genetic model, all three metrics increased where the full contemporary/genetic model yielded R2 of 0.392 (95% CI 0.368–0.419), C-statistic 0.758 (95% CI 0.745–0.775) and PAR of 98.4% (95% CI 97.7–99.0). Individual risk factors followed a similar rank order to the primary analyses with systolic blood pressure (R2 24.8% (95% CI 12.2–39.2)), followed by PRSCAD (R2 14.8% (95% CI 10.14–20.9)), then followed by cardiometabolic risk factors (BMI, diabetes, dyslipidemia) summarizd in Supplemental Table 3. By C-statistic and PAR, PRSCAD was prioritized as the most important risk factor, followed by BMI, summarized in Supplemental Table 3.
Discussion
In this large cohort study, including contemporary biomarkers and genetic information while accounting for the use of preventive medications, we show that among individuals without preexisting CAD, genetic risk factors, inflammatory risk and Lp(a) play a significant and previously underappreciated role in incident CAD. Importantly, this study updates our understanding of the relative importance of traditional and contemporary risk factors.
Adding traditional risk factors to the base model of age, sex, lipid-lowering therapy and ever-smoking status improved both model fit as measured by R2, and model discrimination as measured by C-statistic. Additionally, adding multi-ancestry CADPRS, hsCRP, Lp(a) and HeFH variants further improved prediction accuracy and explanatory power by all metrics (R2, C-statistic and PAR). The total PAR of the full model was 84.3% of the risk of incident MI in this cohort.
While prior work estimated 80–90% of risk of incident MI is due to modifiable lifestyle factors alone, these data were generated prior to the inclusion of large-scale genomic sequencing efforts in population-scale biobanks and have needed updating. Using a “leave-one-out” paradigm, sequentially removing variables from the full model was used to calculate the relative contribution of risk factors to the overall predictive power of the model. Our analysis shows that the two most powerful risk factors were hypertension and PRSCAD. This is due to the fact that both have high hazard ratios for incident MI (and are strongly causative risk factors), and also due to their high prevalence among the study population. Hypertension was prevalent among 30% of the population and a high polygenic risk is typically defined as a percentile risk score greater than the 80th percentile, though of course both of these risk factors exist along a continuum.
The influence of secular trends of increasing rates of obesity, diabetes and a cardiometabolic risk picture are also emphasized by the results of the PAR analysis which weights population prevalence highly. While polygenic risk has only recently been studied at a larger scale and was thus not integrated into prior analyses, our findings are consistent with worldwide trends that hypertension has emerged as the most prevalent cardiovascular disease risk factor, largely doubling in prevalence since the INTERHEART study27. Similarly, next to PRSCAD and hypertension, the next most important risk factors were WHR, hsCRP, ApoB:A1 ratio and Lp(a), which implicates a combined cardiometabolic and inflammatory phenotype responsible for risk in contemporary populations that has followed global trends.
Extending our observations beyond the primary care setting to a validation cohort in a hospital-based biobank allowed us to verify major findings in a population enhanced for comorbidities. In the MGBB we were able to verify that even using sparse models, the inclusion of polygenic risk to baseline clinical cardiometabolic variables added substantial explanatory and discriminative power. Among individual risk factors systolic blood pressure and PRS were prioritized by R2 analyses, followed by BMI, diabetes and dyslipidemia. However, the PAR analyses prioritized PRS first, followed by BMI, diabetes, dyslipidemia and then systolic blood pressure, perhaps reinforcing that secular trends in worsening adiposity, diabetes and dyslipidemia are of increasing importance – especially among hospitalized patients. The magnitude of the contribution of individual risk factors to the model differed between the UKBB and MGBB – which may be expected due to differences in cohort demographics and different phenotyping fidelity in a hospital-based biobank versus a research cohort with standardized data collection methods.
Our results highlight a need for simultaneous focus on the management of traditional risk factors and increasing our treatment armamentarium for contemporary risk factors. While challenges exist to optimal blood pressure management including increasing comorbidity burden, imprecision in measurement methods, and limited access to lifestyle intervention programs, renewed efforts at broad-scale implementation are urgently needed 27–29. WHR is similarly modifiable, and with an expanding arsenal of effective therapies for obesity such as incretin-mimetics, these findings highlight the importance for cardiologists to familiarize themselves with these therapies. Additionally, WHR likely represents, at least in part, integrated risks of lifestyle factors like nutrition and exercise that can be modified. In regards to Lp(a), while its level is 90% genetically predicted from the LPA gene locus, it is likely to become highly modifiable as monoclonal antibody and RNA-based therapies are in promising Phase 3 trials30, 31. Additionally, each of these factors – hypertension, obesity, dyslipidemia and Lp(a) – have a relationship to inflammatory risk that is being increasingly understood and is now modifiable – through both dietary and lifestyle interventions as well as pharmaceutical interventions as demonstrated by trials such as COLCOT, LoDoCo2, and CANTOS32–34. Furthermore, while genotypes are nonmodifiable (leaving aside genetic editing), these findings highlight their important contribution to contemporary CAD risk. Given its sizeable contribution to CAD risk, while healthy lifestyle intervention has been shown to mitigate high genetic risk in observational studies, prospective studies to identify strategies to best assess PRSCAD-associated CAD risk are needed11, 35.
The above findings add to the current body of literature regarding global cardiovascular disease risk. Recently, the Global Cardiovascular Risk Consortium pooled data from more than 1.5 million subjects, and found 5 traditional CVD risk factors (BMI, systolic blood pressure, non-HDL-C, current smoking, and diabetes) had a PAR of 57.2% for incident CAD and 22.2% of all-cause mortality. The authors were unable to assess risks of genetic variants or contemporary biomarkers, and specifically identify these areas as possible variables to explain the residual unexplained risk in their study2. In the Prospective Urban Rural Epidemiology (PURE) study, 14 traditional risk factors were assessed across 21 countries including behavioral (tobacco use, alcohol consumption, diet quality, physical activity, sodium intake), metabolic (hypertension, dysglycemia, diabetes, non-HDL-C, and WHR), socioeconomic (education level, depression), and other risk factors (grip strength, air pollution, ambient pollution)36. Approximately 70% of the risk of incident CAD was explained by these factors, with hypertension having the largest single PAR (22.3%) – consistent with our findings. However, genetics and contemporary biomarkers were not included in the analysis.
Our study had several strengths and weaknesses. First, an important limitation of our study is that our study cohorts largely comprise of individuals of European ancestry. In this study, we demonstrate robustness across two continents but validation across communities and contexts is warranted. Second, lifestyle risk factors in this study were ascertained through self-report. Results that violated our expectations – such as unhealthy diet and physical activity being non-significant contributors to model prediction could perhaps be attributed to this self-report recall-based data collection. Lastly, the well-documented “healthy volunteer” selection bias of the UK Biobank alters risk factor distributions but whether this differentially impacts some risk factors and their clinical effects over others is not well understood37.
Conclusions
In this prospective cohort study of 299,707 individuals, the addition of genetic risk factors and contemporary biomarkers to explanatory models for CAD shows previously underappreciated importance of contemporary CRFs such as PRS, hsCRP, and lipoprotein(a) alongside traditional CRFs such as hypertension, dyslipidemia and presence of diabetes. This reflects an increasingly recognized interplay between germline genetics and secular trends in cardiometabolic health.
PN
Dr. Pradeep Natarajan
Cardiology · Preventive CardiologyMassachusetts General Hospitalpradeep
