Risk prediction has been used in the primary prevention of cardiovascular disease for >3 decades. Contemporary cardiovascular risk assessment relies on multivariable models, which integrate established cardiovascular risk factors and have evolved over time from the Framingham Risk Model to the pooled cohort equations to the PREVENT (Predicting Risk of CVD Events) equations. Recent scientific (ie, genomics, proteomics, metabolomics) and methodologic (ie, artificial intelligence) advances have led to a proliferation of novel models, biomarkers, and tools for potential use in risk prediction. In parallel, the growing armamentarium of preventive therapies, some with considerable cost, underscores the need for more accurate and precise risk assessment to prioritize those at highest risk who will derive the greatest absolute benefit. Accompanying the considerable enthusiasm for the potential of newer approaches to improve risk prediction is the need for rigorous evaluation and assessment of their performance (ie, accuracy, precision, incremental performance when added to contemporary multivariable risk models or established risk factors) and clinical utility (ie, actionability, scalability, generalizability) before adoption in clinical practice. Additional considerations in risk tool evaluation include reproducibility, cost–value considerations (including impact on downstream health care costs), and implications for health equity. This scientific statement defines a standardized framework for general considerations in risk prediction, statistical assessment of predictive utility, and critical appraisal of clinical utility and readiness. This scientific statement is intended to support clinicians, researchers, and policymakers in how best to evaluate current and emerging risk prediction tools and ultimately improve the prevention of cardiovascular disease in diverse populations.
Cardiovascular disease (CVD) remains the leading cause of morbidity and mortality in the United States, with estimates that >45 million adults will have CVD by 2050 if current trends continue. Effective preventive strategies are needed to address these projected increases in CVD prevalence. A foundational part of the contemporary clinical approach to the primary prevention of CVD is risk assessment with use of multivariable risk prediction models that integrate established cardiovascular risk factors. This risk-based paradigm is widely endorsed across multiple American Heart Association/American College of Cardiology clinical practice guidelines (Table 1). These models have advanced considerably over the past 3 decades, from the development of the Framingham Risk Model to predict coronary heart disease to the pooled cohort equations (PCEs) to predict atherosclerotic CVD (ASCVD; coronary heart disease and stroke) to the PREVENT (Predicting Risk of CVD Events) equations to predict total CVD (ASCVD and heart failure [HF]). Still, great enthusiasm remains to further improve accuracy, precision, and clinical utility of risk prediction models beyond those that rely on established or traditional risk factors alone. Moreover, the emergence of novel but costly therapies, such as glucagon-like peptide-1 receptor agonists, sodium-glucose cotransporter 2 inhibitors, nonsteroidal mineralocorticoid receptor antagonists, and proprotein convertase subtilisin/kexin type 9 inhibitors, underscores the need for accurate risk models to consider risk-based prioritization to target those at highest risk who may derive the greatest absolute benefit and maximize the cost-effectiveness of these therapies.
AACVPR indicates American Association of Cardiovascular and Pulmonary Rehabilitation; AANP, American Association of Nurse Practitioners; AAPA, American Academy of Physician Associates; ABC, Association of Black Cardiologists; ACC, American College of Cardiology; ACCP, American College of Chest Physicians; ACPM, American College of Preventive Medicine; ACS, American College of Surgeons; ADA, American Diabetes Association; AGS, American Geriatrics Society; AHA, American Heart Association; AMA, American Medical Association; AMSSM, American Medical Society for Sports Medicine; APhA, American Pharmacists Association; ASCVD, atherosclerotic cardiovascular disease; ASNC, American Society of Nuclear Cardiology; ASPC, American Society for Preventive Cardiology; BP, blood pressure; CCD, chronic coronary disease; CHARGE-AF, Cohorts for Heart and Aging Research in Genomic Epidemiology for Atrial Fibrillation; CVD, cardiovascular disease; DBP, diastolic blood pressure; HRS, Heart Rhythm Society; MACE, major adverse cardiovascular event; NLA, National Lipid Association; NMA, National Medical Association; PACES, Pediatric and Congenital Electrophysiology Society; PCE, pooled cohort equation; PCNA, Preventive Cardiovascular Nurses Association; PREVENT, Predicting Risk of CVD Events; SBP, systolic blood pressure; SCA, Society of Cardiovascular Anesthesiologists; SCCT, Society of Cardiovascular Computed Tomography; SCMR, Society for Cardiovascular Magnetic Resonance; SGIM, Society for General Internal Medicine; and SVM, Society for Vascular Medicine.
Scientific, technologic, and methodologic advances have led to a proliferation of novel biomarkers and tools for risk prediction of CVD, with a growing number of scientific statements highlighting the importance of innovative approaches to assess risk (eg, noncoding RNAs, polygenic risk scores [PRS], artificial intelligence [AI]). However, published studies often have variable rigor in their reporting of the predictive and clinical utility of such novel approaches to assess CVD risk. In addition, some risk tools may warrant unique considerations in determining their clinical utility and readiness for clinical implementation, such as variation across omics platforms and assays or evaluation of bias with AI tools that rely on data collected for clinical care in electronic health records (EHRs).
This scientific statement provides an updated framework for the critical appraisal of the predictive and clinical utility of novel biomarkers, models, and tools and builds on the American Heart Association scientific statement “Criteria for Evaluation of Novel Markers of Cardiovascular Risk.” Herein, we review general considerations for CVD risk prediction; statistical assessment of accuracy, precision, and incremental value relative to current risk models used in clinical practice; clinical utility; and additional considerations before implementation in practice (Figure 1). Although many of the principles outlined are also applicable to risk prediction of recurrent CVD events (ie, secondary prevention) or non-CVD events, the focus of this scientific statement is evaluation of novel tools for prediction of incident CVD for primary prevention, where risk assessment has the greatest impact on clinical practice and is used in contemporary guidelines to inform recommendations for preventive therapies. The intent is to support clinicians, researchers, and policymakers in evaluating current and emerging risk prediction strategies and ultimately improving CVD prevention in diverse populations.
Development of a risk prediction model or validation of a new biomarker or tool typically proceeds in steps from concept to statistical evaluation to clinical utility. A biomarker, as defined by the National Institutes of Health, is “a characteristic that is objectively measured and evaluated as an indicator of normal biological processes, pathogenic processes, or pharmacological responses to a therapeutic intervention,” and thus is not limited to blood tests. Before the predictive utility of a new biomarker can be evaluated, the distribution must be demonstrated to differ significantly among individuals who eventually are affected by the outcome of interest compared with those who do not develop the outcome. This is best accomplished with a prospective study design to assure that the individuals in the study were at essentially similar risk of the outcome at baseline and did not already have the disease or outcome of interest. If data are prospectively collected in an observational research cohort, relevant variables are more likely to be systematically collected in every individual, which will minimize missing data and bias in model development.
Table 2 outlines the key questions to consider when developing a risk prediction tool, which includes a priori clarifying the clinical question of interest and defining the intended population of interest, as a test may predict well in one clinical setting but not in a different setting, population, or indication. The process used to evaluate differences across groups (eg, race and ethnicity as a social construct, genetic ancestry), which could include assessment of generalizability across different subgroups, should be decided on in advance. The relevant time frame of interest should also be defined in the context of the clinical question. A disease that develops over a short time period can be studied in a few days or weeks, whereas many important outcomes, such as incident ASCVD or HF, usually take years to develop and require longer-term studies to understand the magnitude of association with biomarkers or risk models. The time horizon selected can also inform the approach for considering competing risk of non-CVD events, as adjusting for competing risk has less impact on risk estimation with shorter follow-up compared with longer follow-up times.
