Precision epigenetics provides a scalable pathway for improving coronary heart disease care globally

Coronary heart disease (CHD) is the largest cause of death in the world and in the USA [1]. Each year, approximately 8 million people die, while millions more become temporarily or permanently disabled as a result of CHD. Disturbingly, these morbid figures are only a portion of the toll that CHD extracts from mankind. Each year, in the USA alone, the cost of healthcare services and the economic effects of premature death are $219 billion annually [2]. As high as those numbers seem, they reflect only a moderate societal burden of CHD. In comparison, many developing countries, such as Uzbekistan, have far higher rates of age-standardized CHD-related death rates, yet lack the healthcare infrastructure to compensate for the additional burden [3]. This disparity in CHD burden has large consequences. From a purely economic viewpoint, the disproportional impact of CHD is perhaps felt most acutely in China, where 38% global growth in the number of CHD deaths occurred between 1990 and 2017 [3]. When combined with the effects of an aging population, this CHD epidemic markedly threatens the economic development of China as well as overall global prosperity [4,5].

Though these numbers may be startling to the unfamiliar, these grim statistics are of little surprise to many in healthcare policy who have steadily advocated the need for more scalable methods for the diagnosis, treatment and prevention of CHD. To the lay public, perhaps the most familiar response to this epidemic has been the development of new drugs. Driven by the strong relationship between serum cholesterol levels and high blood pressure to the development of CHD [6], over the past three decades, pharmaceutical companies have introduced hundreds of agents designed to lower blood pressure, reduce cholesterol levels and decrease endothelial injury. These efforts have not always been completely successful from a clinical standpoint; for example, though their ability to reduce cholesterol levels is beyond question, statins have only a marginal benefit for preventing CHD [7]. Similarly, using a more mechanical approach to addressing stenotic arteries, device manufacturers have introduced several generations of cardiac stents to the clinical market. Unfortunately, despite the tens of billions of dollars spent on these cardiac stents, they still appear to be ‘most useless’ for the treatment of CHD, with a paucity of evidence that stents are effective for those with stable CHD [8,9]. Their primary value rests in the emergent intervention clinical space, when CHD has created late-stage lesions and vascular compromise. In summary, despite herculean efforts, our current clinical armamentarium for addressing CHD has significant shortcomings.

One problem these efforts may be encountering is finding the right patient. In order to maximize clinical benefit while minimizing unwanted effects and costs, these interventions must be targeted to the appropriate individuals. Therefore, yet another substantial industry has simultaneously pursued improvements in methods for identifying those with, or likely to develop CHD. This effort has a rich history. Beginning with the discovery by Feil and Siegel nearly 100 years ago (1928) that angina was associated with changes in the electrocardiogram [10]11, a steady stream of advances in biomedical engineering has led to the development of new methods to diagnose, and occasionally monitor, CHD. Although the technology behind some of these innovations is impressive, each of these new methods has a number of drawbacks that suggest additional approaches for diagnosing CHD are needed in order to address the burden of CHD on a global scale.

We believe that recent advances in epigenetics – specifically DNA methylation – and artificial intelligence (AI) may lay the foundation for a more effective, scalable precision epigenetics approach for the diagnosis and management of CHD worldwide. In this paper we review the past decade of studies of the DNA methylation landscape of prevalent CHD, then discuss how AI circumvents a central failing of conventional methylation analyses. We compare and contrast current diagnostics for CHD with newly introduced methylation-based methods. Finally, we highlight possible trajectories for precision epigenetic tools in cardiovascular care and how they could help alleviate health disparities in both tier 1 and developing economies.

White blood cells as a biosensor of the cardiovascular system

Although often overlooked, a first step in any methylation study should be a consideration of what constitutes the ideal biological sample. For certain applications, the answer can be obvious; for example, with respect to epigenetic studies of cancer, the ideal candidate tissue for many studies is the tumor itself. Intuitively, this makes sense, because a decade of studies has shown that the DNA methylation signature of a cell helps define its behavior and that cancer cells behave differently than noncancerous cells [12]. Retrospectively, the consideration of which tissue should be used for epigenetic analysis has not always been a prominent feature of study design. In fact, in our early studies of human DNA methylation in which we described the promoters for the serotonin transporter and monoamine oxidase A, we simply used the materials around the lab that were available from our genetic studies – in those cases, lymphoblasts and whole-blood DNA [13,14].

However, with our discoveries of the cg05575921 locus in the aryl hydrocarbon receptor repressor (AHRR) in 2011, our group came to realize that the choice of tissue for methylomic studies mattered [15]. For example, retrospectively, we discovered that the xenobiotic pathway that is regulated by AHRR is very active in white blood cells (WBCs). Because WBCs actively participate in the catabolism of toxins such as polyaromatic hydrocarbons, the AHRR locus is able to undergo marked changes in DNA methylation status, in this case from over 90% to less than 10% methylation, as a function of exposure to tobacco smoke [16]. If the gene were epigenetically silenced in WBCs, we likely would not have observed a significant signal.

Fortunately, for those seeking to develop new diagnostic tools for the assessment of CHD, WBCs are ideal tissue. The most obvious attribute is their availability: phlebotomy is routinely performed during clinical encounters, and WBC DNA is readily prepared even in relatively low-tech settings. Perhaps even more importantly, WBCs are an integral portion of the cardiovascular system. Embryologically, WBCs arise as clusters of cells attached to the endothelial cells of the aortic arch during the third week of gestation, and begin to migrate to the liver between the fourth and sixth week of gestation [17]. Beginning at about the 11th week, hemopoietic stem cells from the liver begin to colonize the bone marrow, in which they will produce WBCs for the rest of life [18]. WBCs are a heterogeneous group that can be dichotomized as being either lymphoid or myeloid. Lymphoid cells include T lymphocytes, B lymphocytes and natural killer cells. Myeloid cells include neutrophils, basophils, eosinophils and monocytes [19]. Typically, myeloid cells, particularly neutrophils, constitute the bulk of the WBC count, with lymphocytes normally constituting 20–40% of the cell count. Unlike brain cells, which are protected by the blood–brain barrier, WBCs are continuously exposed to the same metabolite/hormonal concentrations that the rest of the vascular system experiences. Therefore, from a theoretical standpoint, they are well positioned to serve as ‘cellular biosensors’ for the cardiovascular system.

For some epigenetic applications, it is critical to note that although the overall DNA methylation signature for each cell is very similar, each WBC cell type has its own unique epigenetic features. In order to compensate for that complexity and the occasional shifts in individual WBC counts that may be associated with disease, computationally intensive methods for correcting for cellular heterogeneity in genome-wide analyses have been developed [20]. Though significant, highly reproducible, cell-type-specific changes in DNA methylation in response to environmental factors have been observed [21], very few in vitro diagnostic tests that have come to market use methylation assessment approaches capable of incorporating complex cell-type corrections [22]. Therefore, although they can potentially be important, any methylation findings that are cell type specific have an added layer of complexity that can constitute a barrier to clinical implementation. Fortunately, although the use of these techniques is critical for understanding the origin of positive findings (and is absolutely essential when using DNA sources such as saliva), many investigators have reported that these correction techniques have no effect on major findings when using WBCs as a source of DNA [23,24].

Regression analyses of DNA methylation of prevalent categorical CHD

The proliferation of array-based methylation tools has led to an explosion in the number of genome-wide methylation studies conducted by those interested in the biology of CHD (Table 1). In part driven by the ease of laboratory measurement and the continuous nature of physiological variables that lends itself to analysis, the vast majority of CHD-related examinations have focused on continuous traits such as lipid levels. In contrast, analyses of the epigenetics of prevalent categorical CHD are relatively few. In fact, a search of PubMed using the terms ‘coronary heart disease’ and ‘methylation’ and ‘genome wide’ identified only 50 studies. Of these 50 studies, only four [25–28] analyzed original whole-blood DNA samples from subjects categorized for current CHD or an analogous condition using a traditional epigenome-wide approach, with three of these studies conducting epigenome-wide studies on eight or fewer CHD subjects. The remainder of the studies analyzed methylation of other types of DNA (e.g., arterial), lumped CpG methylation values into differentially methylated regions or analyzed CHD-related methylation with respect to other biomarkers associated with CHD such as lipids [29], but not prevalent categorical CHD.

Puzzlingly, since the exhaustive 2017 review of CHD and CHD-related association studies by Fernández-Sanlés and associates, and in direct contrast to the numerous studies for incident CHD [30–32], our literature search only uncovered one additional methylome-wide study of prevalent categorical CHD or CHD analog [33]. This study, an examination of acute myocardial infarction, again by a Fernández-Sanlés-led group [26], identified 25 CpG sites associated with acute myocardial infarction. Interestingly, while the authors note that four of these sites – cg05575921 (AHRR), cg25769469 (PTCD2), cg21566642 (intergenic) and cg04988978 (MPO) – were associated with risk for incident CHD in the Framingham Heart Study (FHS) Offspring Cohort, none of these 25 CpG sites were associated with risk for incident CHD in a combined genome-wide regression/weighted gene network analysis conducted in 2021 by Si and associates using the DNA from 491 incident cases and 491 matched controls [34]. Still, the reason for the lack of additional epigenome-wide association study (EWAS) analyses of current categorical CHD is unknown, and we note that there are a number of candidate gene studies, such as those of F2RL3, that show the relationship of single-site methylation to categorical CHD [35,36]. However, a suggestion that those in the field may not believe that further regression analyses might lead to improved diagnostic tools may be implied by Fernández-Sanlés’s observation that the addition of information from the four overlapping CpG sites cited above to the standard Framingham risk score did not improve CHD risk prediction in the FHS [26].

Using machine learning & AI to identify DNA methylation predictive of CHD

Traditional regression-based EWAS approaches can be powerful tools for identifying predictors of unitary environmental exposures such as smoking. However, the summed experience of those in the field has shown that they are less well suited to identifying predictors of complex aging-associated disease states that may arise from multiple discrete environmental etiologies. There are exceptions to this rule. Sometimes, as in syndromes such as Type 2 diabetes (T2DM) – for which there is a generally accepted diagnostic biomarker, hemoglobin A1c (HbA1c), and a physiologically defined phenotype (lack of insulin secretion in the background of insulin resistance) [37] – this problem can be overcome. Even then, there are tradeoffs to the reliance on a singular biomarker. In the case of glycosylated hemoglobin level, we know there exist significant discordances between diagnoses established with HbA1c levels and those established by fasting blood glucose or by 2-h plasma glucose levels, which are also American Diabetes approved biomarkers [38], following oral glucose tolerance testing [37]. Therefore, investigators who use this approach to identify CpG sites associated with T2DM must not only be conscious of the bias in the phenotypic spectrum being selected for in their analyses, but also simultaneously take care to separate other diabetes phenotypes, such as Type 1 diabetes, from their analyses.

This key difference between T2DM and CHD may underlie the lack of robust findings using traditional regression approaches and point to a better method for detecting and quantifying the biosignature of CHD. Simply put, there is no concrete unifying physiological feature, such as insulin resistance, for CHD that can easily be assessed via laboratory testing. Instead, CHD is a clinical syndrome whose central feature is ‘insufficient supply of oxygen to the heart’ that can be produced via a number of non-overlapping pathophysiological processes including coronary artery obstruction, endothelial dysfunction, microvascular disease and vasospasm [39]. Given that we cannot directly measure this often paroxysmal oxygen supply deficit, but rather only assess outcomes or corollaries of this insufficient oxygen supply, such as a myocardial infarction or change in ECG, there is no single laboratory feature through which to classify CHD. Instead, it remains a heterogenous, latent construct whose presence is determined by a trained clinician in conjunction with patient reports, clinical examination and standardized testing [40]. The implicit understanding of this important difference between T2DM and CHD by clinical researchers may underlie the lack of published whole-blood DNA EWAS analyses of categorical CHD. Researchers may be reluctant to report a negative finding or try something that they believe will likely fail. At the same time, this reluctance to use traditional methods highlights the potential use of alternative approaches, such as machine learning (ML) and AI, for identifying epigenetic predictors for current CHD.

Broadly defined, ML is a subset of AI that encompasses the methods used to detect and learn patterns in data, while AI refers to the larger set of algorithmic approaches that allow computers to make human-like choices [41,42]. Over the past two decades, driven by advances in computing capacity, the power and complexity of ML/AI methods for producing predictive models has dramatically increased, with the results already being clinically translated [42]. A review of the US FDA database demonstrates that as of October 2022, there are 521 AI/ML-enabled devices [43]. Many of these devices, such as those used to diagnose diabetes-related retinopathy, can markedly outperform clinicians [44]. Unfortunately, the tradeoff is often that on an individual basis, the reason for prediction can be unclear.

In contrast, most EWAS still employ traditional, generally linear-model, approaches to generate candidate methylation biomarkers [45]. These methods, which often employ cell correction techniques, have the advantage of being interpretable [20]. For example, using EWAS combined with cell correction, one can better understand whether an observed change in epigenetic signal is secondary to a change in cell type or average cell methylation. But these approaches have limitations as well. Unlike ML/AI-based approaches, they are less capable of integrating interaction effects, particularly when these effects are complex or nonlinear. Therefore investigators must often balance the advantages of the two different sets of approaches when considering the development of diagnostic algorithms.

A simple illustration of a key difference in the capacity of traditional EWAS and ML/AI approaches for creating predictors for illness is given in Figure 1. The diagnostic spectrum of T2DM can be compared with the depiction of the single-barrel saguaro cactus on the left. The defining feature of the cactus on the left is height, which is analogous to the linear defining feature of T2DM: HbA1c levels. At a certain height or cutoff point, one is classified as affected. However, all of those who have been to the desert in the southwest USA will readily realize that the vast majority of saguaro cacti are more morphologically complex. Whereas sometimes when viewed from the side, a multibranched cactus may appear to be a single tube, a fuller examination reveals a more complex branched morphology. Similarly, the biology underlying T2DM and CHD can be viewed as being the single-barreled and multi-lobed saguaro, respectively. Traditional regression approaches are ideal for capturing the biology of illnesses such as T2DM or smoking intensity that have a unifying, monodimensional biology captured by a biomarker. However, they are less capable of capturing the complexities – or more adroitly, the dimensionalities – of clinical presentations with distinct ‘lobes’ without unifying biomarkers, as depicted by the cactus on the right, for diseases such as CHD. This, in part, is why epigenetic aging indices and similar pure regression-based methods only show statistically significant, yet clinically meaningless, relationships to CHD status [46–48], while the variation in risk biology for CHD captured by the individual methylation probes is often strongly influenced by the status of other methylation markers or genotypes. As a result, there is not a linear relationship of the probe variation to the same dependent variable. However, because of their ability to accommodate nonlinear approaches, AI-based algorithms can robustly describe this multidimensional morphology. But can this theoretical capacity of AI/ML be harnessed in practice to aid the development of precision epigenetic predictors for CHD?

The answer here appears to be ‘absolutely’. In 2018 Dogan and colleagues published the first examination of the use of AI to develop DNA methylation predictors of current CHD. In this study, Dogan used existing genome-wide epigenetic and genetic data from the FHS to produce a random forest classifier that used the input of age, gender, four CpG sites and two SNPs to predict CHD status [49]. Notably, this study also introduced the use of ‘gene × methylation’ interaction effects to CHD prediction models, which can be useful for addressing the impact of genetic confounding on DNA methylation values. Overall, the in silico classifier developed by Dogan worked well, with accuracy, sensitivity and specificity of 78%, 0.75 and 0.80, respectively, in the independent FHS test set. Four years later, in a twist that attempted to gain additional power from matching gene expression data, Zhang and colleagues used AI and both genome-wide methylation and transcriptional data from this same FHS cohort to predict CHD status [50]. However, instead of using the individual values from individual CpG sites, the authors first averaged the β-values from two bins of gene promoter regions as variables in their methylation data-mining efforts. They then filtered these summarized methylation variables against gene transcription data to produce a list of 54 differentially expressed gene variables whose methylation and expression were associated with CHD. Then the methylation and expression data for each of these 54 differentially expressed genes underwent further dimensionality reduction, arriving at a set of five hub genes (ATG7, MPO, DHCR24, BACH2 and CDKN1B) to serve as predictors. The summarized methylation and gene expression data for each of these genes were then fed, together and separately, into an ML model, using LightGBM, XGBoost and random forest methods, to predict CHD status. LightGBM models worked marginally better than XGBoost and random forest models for all three types of data. A LightGBM model using methylation data only had an area under the curve (AUC), sensitivity and specificity of 0.77, 0.64, 0.83, respectively, in the test set, while a LightGBM model using only expression data had a slightly inferior performance, with an AUC, sensitivity and specificity of 0.71, 0.63, 0.83, respectively, in the test set. However, a LightGBM model that incorporated both methylation and expression data from those five genes performed best, with an AUC, sensitivity and specificity of 0.834, 0.672, 0.864. respectively, in the test set.

A weakness of each of these prior studies is the reliance on out-of-date methylation (Illumina 450K) and transcription (Affymetrix Human Exon 1.0 ST) arrays that are cumbersome to employ in a clinical setting. Therefore we have recently developed a more robust, AI-driven, clinically implementable predictor for CHD. Notably, this effort incorporated the data from three independent CHD cohorts totaling 449 CHD subjects and 2067 controls, with translation of the array-based SNP and CpG assessments into rapid, clinically implementable, precise fluorescent primer hydrolyzable probe (TaqMan^®) and methylation-sensitive digital PCR (MSdPCR) assays [51]. Table 2 summarizes the performance of an ensemble algorithm that incorporates the input from ten SNPs and six MSdPCRs (targeting cg03725309, cg12586707, cg04988978, cg17901584, cg21161138 and cg12655112) to predict current CHD status. Like the Zhang method, almost all of the sensitivity is provided by the methylation component, and two of the six MSdPCR assays in the PrecisionCHD™ test target methylation at genes (cg04988978 [MPO] and cg17901584 [DHCR24]) identified by Zhang et al. [50].

These new developments are just the beginning of precision epigenetic approaches for CHD diagnostics. Although this review focuses on the impact of advances in AI and methylation technologies, their ability to predict CHD status still is constrained by the relative lack of robustly characterized clinical datasets for algorithm training. It is often noted that algorithms are only as good as the data used to train them. Unfortunately, there is a distinct paucity of readily available sets of well-characterized subjects, both CHD cases and controls, for algorithm training. In particular, people of color and women are distinctly under-represented in available datasets. Because the clinical phenomenology of CHD in these groups can differ from those of the White FHS subjects, the performance of these tools may be less for these individuals, and the lack of understanding of these tools in non-White cohorts undercuts efforts to democratize these tools throughout the world. Larger, inclusive, well-defined datasets would ensure the robustness of these tools for all people.

Integrating precision epigenetic tools into the clinic

The introduction of new diagnostic tools into clinical practice is often slow. Some of the reasons for this are easily understood; for example, the practice of medicine itself is algorithmic, with any shift in decision-making patterns using new technologies requiring considerable education of both new and existing providers. Advances in technology tend to be viewed with a skepticism rooted in the principles of beneficence and a desire to adjust practice techniques in a methodical manner, employing robust due diligence on behalf of patient safety. But even if the new diagnostic tool is superior and the biology of the underlying test is readily understood, the rate of change can be glacial. For example, the use of HbA1c to guide the treatment of diabetes was first proposed in 1976 by Cerami and colleagues [52], but it was not adopted as a standard method for diagnosing diabetes by the American Diabetic Association until 2010 [38]. Because many of the new epigenetic tools being introduced are considerably more complex than the HbA1c test, it is likely that adoption of precision epigenetic tools may be less swift than desired, despite the potential societal and healthcare infrastructure benefits.

Part of the reason for a slower than desired speed of adoption is the already existing diverse landscape of current cardiac testing technologies. According to guidelines of the American Heart Association/American College of Cardiology, the diagnosis of CHD should only be made by a trained clinician in conjunction with a comprehensive clinical evaluation and recent exercise or cardiac imaging study [40]. The choice of the exercise or cardiac imaging study is at the discretion of the clinician, ideally guided by the pre-test likelihood of the patient having CHD using established methods such as the estimate provided by the Combined Diamond/Forrester Coronary Artery Surgery Study scale [40]. These choices include exercise ECG, exercise echocardiogram (ECHO), coronary artery calcification, pharmacological stress ECHO, exercise myocardial perfusion imaging, coronary computerized tomographic angiography (CCTA) and conventional angiography [40].

How does the performance of these new epigenetic tools compare with that of those already in use for the evaluation of stable CHD? Table 3 provides a summary of the performance characteristics of each of the clinically available tests. At first glance, the least sensitive test is the exercise ECG, while the most sensitive test is CCTA. However, there is an important caveat emptor. With the exception of PrecisionCHD, the performance metric comparator for all the other tests is the angiogram. However, it is well established that up to 50% of all patients undergoing angiography do not have obstructive coronary artery disease, with a relatively higher prevalence of ‘no obstruction’ findings in women (65%) than in men (32%) [53–56]. Instead, many if not most of those individuals have a form of ischemic CHD referred to as ‘ischemia with no obstructive coronary artery disease’ (INOCA), previously termed ‘syndrome X’; an expert panel from the European Association on Percutaneous Interventions has concluded that “up to 70% of those undergoing coronary angiography because of angina and evidence of myocardial ischemia do not have obstructive coronary arteries but have demonstrable ischemia, i.e., INOCA” [57]. They further concluded that INOCA has a “comparable incidence of adverse events as well as impaired quality of life as obstructive coronary artery disease (CAD)”. Hence it is vitally important that testing measures recognize this frequently fatal form of CHD.

Detecting INOCA is a tall task. INOCA is thought to result from a variety of pathologies, including endothelial dysfunction, microvascular dysfunction, vasospasm and non-flow-limiting obstruction (10–30%), alone in or combination [56–58]. Specialized physiological tests, such as spasm testing, may elicit anginal symptoms in those with INOCA [57,58]. However, the exact sensitivity of any of these secondary tests for INOCA is unknown. Critically, many of these subjects thought to have INOCA have rest ECG abnormalities, with an additional 30% of those with angina and normal coronary arteries on angiogram having positive exercise ECG findings [56]. Because the anatomically based methods such as CCTA are uninformative in the absence of obstruction [59], this new and evolving understanding of the CHD spectrum suggests that physiologically based methods such as exercise ECG and PrecisionCHD (which was trained on all presentations of CHD, not just obstructive CHD), may actually capture more CHD presentations than the values in Table 3 would suggest.

Still, performance is not the only metric that needs to be considered when choosing tests for CHD. A general rule in medicine is that the more invasive the test, the larger the side effect profile. This certainly applies to the CHD tests listed in Table 3. Exercise ECG and ECHO can be tiring, but rarely have serious side effects. Similarly, PrecisionCHD simply requires a blood draw. However, each of the imaging methods requires the use of contrast dyes. Though controversial, the contrast dyes used for angiographic techniques can result in high rates of kidney damage [60]. In particular, in a study of 1196 subjects undergoing coronary angiography, for those with serum baseline creatines of 1.5 mg/dl or higher, the rate of acute kidney injury was 33% in those with diabetes and 12% in those without diabetes [61]. Less well constrained are the effects of exposure to ionizing radiation. The average CCTA delivers 2.88 ± 0.85 mSv (~30 chest x-rays’ worth), while the average coronary angiogram delivers 5.61 ± 0.55 mSv of exposure [62]. Because the rate of cancer from repeated x-ray exposure and the mortal peril from acute kidney damage are not trivial, clinicians must also factor the side effects of the testing approach into consideration.

An increasingly necessary consideration is the cost of testing. The upfront costs of PET scans ($4637) and angiography ($9438) as delineated in Table 3 are considerable and, given the projections of the expanding patient base in CHD, bordering on unsustainable. There are also additional unseen costs. As noted above, one is that of possible contrast dye-induced kidney damage, but a less calculable but potentially greater unseen cost is that of providing the infrastructure to support the purchase and maintenance of specialized technologies such as PET. Often, particularly at large academic institutions, the cost of these instruments is subsidized through federal grants, with the regulation and inspection of all PET facilities requiring additional federal infrastructure. Angiography – while there have been improvements in approach, technique and procedural safety that are numerous and notable – continues to be an invasive procedure that subjects a patient to some measure of discomfort and does run risks of extremely serious complications that can result in extensive morbidity and mortality.

Finally, a major consideration in choosing a test is the value of the information for patient management. Prior to 2008, the conventional wisdom among many cardiologists was that angiography was the preferred method of CHD evaluation because of the potential for revascularization to be performed during the angiographic procedure. However, the Clinical Outcomes Utilizing Revascularization and Aggressive Drug Evaluation (COURAGE) trial, published in 2008, which was followed by the International Study of Comparative Health Effectiveness with Medical and Invasive Approaches (ISCHEMIA) trial in 2020, convincingly showed that percutaneous coronary interventions (PCIs) do not improve the rate of myocardial infarctions or survival over standard medical therapy alone in those with stable CHD [63,64]. Although some cardiologists still feel that there may be benefits for PCI in certain circumstances, and there is evidence that PCI may improve anginal symptoms in the short term [65], the weight of the scientific evidence is that although imaging approaches can show obstructive disease, the initial treatment choice for all patients shown to have stable CHD should be standard medical therapy [66]. Still, if the results from COURAGE, ISCHEMIA and other trials are correct, the use of standard medical therapy for almost all obstructive presentations is an important point for two reasons. First, it makes the ability of other methods to detect non-obstructive CHD much more impactful. Second, it eliminates the argument that imaging methods aid in choosing the correct treatment method.

In light of these and other developments, we believe that a careful consideration of the performance, side effects and treatment insight considerations of each of the testing options points to a bright future for precision epigenetic approaches for CHD. First, the sensitivity and specificity of the only clinically available test is certainly comparable, if not superior, to most of the current testing options when INOCA is included as part of the CHD spectrum. Second, precision epigenetic approaches simply require a blood draw; there is no invasive procedure, no contrast dye, no ionizing radiation. Third, not only does the mapping of the indices as shown by the Zhang [50] and Dogan [49] studies provide personalized insight into the etiological drivers of CHD for each patient, but in addition, the dynamic nature of DNA methylation allows the efficacy of treatment to be assessed [67]. Fourth, and perhaps most crucially, precision epigenetic approaches are scalable. Scalability is increasingly becoming a critical factor for providing affordable healthcare in both mature and developing economies. Simply put, given the increasing challenges of caring for our aging populations, diagnostic approaches with low infrastructure requirements and whose costs can benefit from increased volume have a critical advantage. Exercise ECGs have that advantage [68]. One commonly used diagnostic treadmill (GE Mac 5000 Stress System, GE Medical Systems, WI, USA) is available in the USA for under $6000. Furthermore, thanks to the advances of AI, the interpretation of exercise ECG testing may be automated in the future [69,70]. Precision epigenetic approaches utilizing MSdPCR are also inherently scalable; the only specialized piece of laboratory equipment needed to conduct this test is a digital PCR machine, and these are becoming increasingly affordable. There are already five manufacturers of these devices in the US market (Bio-Rad, Qiagen, Roche, Stilla and ThermoFisher) and at least three others seeking to develop similar platforms [71]. One platform is the QX-200 Droplet Digital PCR System, already FDA cleared, with the global market for dPCR expected to grow by 24% annually to reach $2 billion by 2029 [71,72]. Hence, because PCR-based testing approaches are normally easy to conduct and very affordable, there is considerable reason to believe that with time and competition, precision epigenetic testing methods can be democratized to a large share of the global population. For example, unlike sequencing-based tests that can list for thousands of dollars, PrecisionCHD retails at $849 in the USA (Table 3). With scaling, even in the US market, it is highly likely that the actual price for these and similar dPCR tests will decrease.

How would these and other advancements improve our current approach to CHD diagnostics? Although preliminary, we believe that in the near future, the following assessment algorithm, adapted from the recent review by Joseph et al. of the European Society for Cardiology and the American College of Cardiology guidelines, will come to fruition (Figure 2) [73]. In brief, for those being evaluated in the context of a likely stable CHD presentation, we believe that precision epigenetic testing will become a primary testing option for those at low and intermediate pre-test probability, and a valuable companion diagnostic for those high-risk cases for whom angiography is chosen. Our expectation, with this scalable characteristic in mind, is that a greater incidence of disease will be found earlier in the pathophysiological cascade of CHD across a larger subset of populations, when options for medical management could be brought to bear earlier and more often. Indeed, if the value proposition for epigenetics outlined in Table 3 comes to fruition, precision epigenetic CHD testing may also evolve into a frequently used adjunctive test even if a different primary testing modality is selected. However, it should be emphasized that the assessment and management of CHD is a broad undertaking with numerous stakeholders, including cardiologists, primary care practitioners, payors and, most importantly, patients. Therefore any changes will need careful vetting before implementation.

Conclusion

Without a doubt, the future in precision epigenetics for the diagnosis and management of CHD is bright. But we are still in its infancy. Real-world data collected from tens of thousands of patients analyzed exhaustively by those with a variety of skill sets and perspectives will be necessary before any clear pronouncements can be made responsibly. Furthermore, the implementation of precision epigenetics will require considerable education of providers. But the weight of a decade of steady scientific progress suggests that a bold new dawn for better, more affordable cardiovascular care is here.

Future perspective

The current financial and technical challenges of assessing CHD are formidable. However, the steady evolution and convergence of three distinct intellectual domains – namely epigenetics, AI and digital PCR – has recently led to powerful new methods for assessing CHD status. However, these areas of inquiry are not static and are instead continuing to grow exponentially, while their costs of implementation in real-world settings – as exemplified by AI-guided, integrated genetic–epigenetic tools – continue to decrease. Given the power of these and similar precision epigenetic approaches for assessing conditions such as CHD, we believe that trends indicate that these and other similar easy-to-implement clinical tools will become increasingly prevalent, if not dominant, technologies for the assessment, diagnosis and management of CHD in both mature and developing economies.