• Atrial Fibrillation
• Risk Assessment
• Biomarkers

The term ‘epidemic’ is increasingly used to describe the rising global prevalence of atrial fibrillation (AF). Recent estimates suggest that AF accounts for between 0.9% and 1.6% of total healthcare expenditure in the UK, forecast to rise to 4% over the next two decades.1 This trend—which is also anticipated internationally—underpins efforts to identify individuals at high risk of future AF, in addition to those with AF without manifest symptoms, in the hope of targeted prevention and early treatment. Indeed, numerous studies are currently investigating the impact of such approaches on clinical outcomes and healthcare utilisation.

The association between AF and various conditions—including hypertension, heart failure, sleep apnoea and chronic kidney disease—is well-described, highlighting that AF is often a multisystem disorder. Accordingly, the management of AF has shifted towards a holistic and integrated approach, targeting comorbidities and risk factors, itself associated with improved outcomes.2

Before the actual onset of AF, some focus has been directed toward the identification of patients at high risk of incident AF. Various clinical risk scores have been proposed, such as the simple C₂HEST score (ie, C₂: Coronary artery disease/Chronic obstructive pulmonary disease (1 point each); H: Hypertension (1 point); E: Elderly (age ≥ 75 years, 2 points); S: Systolic heart failure (2 points); and T: Thyroid disease (hyperthyroidism, 1 point)).3 More complicated clinical risk scores have also been described for incident AF prediction, including the CHARGE-AF, Framingham and HARMS₂-AF scores, as well as the CHADS₂ and CHA₂DS₂-VASc scores (although the latter two were designed for stroke risk stratification, not for prediction of incident AF).4 Unsurprisingly, more complicated clinical risk scores will improve on simple scores in risk prediction, at least statistically.

Alongside clinical risk scores, the use of biological markers (‘biomarkers’) in AF risk prediction continues to gain traction. Various blood, urine and imaging biomarkers have been linked to AF risk and AF-related complications. Again, biomarkers almost always improve on risk stratification based on clinical risk factors (statistically), but even then some validation studies with biomarker-based risk scores only report modest risk prediction (with C-indexes<0.7), for example, for AF-related complications such as stroke. Moreover, many biomarkers are non-specific, reflecting a sick patient or a sick heart.

Liu and colleagues developed a blood protein risk score—comprised of 47 proteins—for the prediction of future AF.5 Based on a one-off plasma sample measurement in a cohort of over 36 000 individuals from the UK Biobank (2450 of whom developed AF within 12 years), the protein risk score had a strong predictive performance, with a C-statistic of 0.802 versus 0.751 for the HARMS₂-AF score. While several of the identified proteins have previously been linked to AF (including N-terminal prohormone of brain natriuretic peptide (NT-proBNP), growth/differentiation factor 15, coagulation factor X, endothelin-1 and renin), the majority (85%) are novel. Of all proteins, NT-proBNP had the largest coefficient, with a C-statistic of 0.785 for a score incorporating age, sex and this biomarker alone. In comparison, an AF polygenic risk score had a C-statistic of 0.748.

The authors should be congratulated for this novel risk score which may improve the ability to predict future AF risk, potentially facilitating the prevention of—and screening for—AF. Their findings are biologically plausible, given that proteins likely better represent and characterise underlying pathological processes linked to AF development—such as inflammation, endothelial dysfunction, oxidative stress, fibrinolysis, apoptosis and calcium haemostasis—over crude clinical risk factors. Herein, we consider the potential benefits and limitations of identifying individuals at high risk of incident AF, before discussing current and future approaches to AF prediction (figure 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Considerations and approaches in the identification of individuals at high risk of future atrial fibrillation (AF). C₂HEST: C₂, Coronary artery disease/Chronic obstructive pulmonary disease; H, Hypertension; E, Elderly; S, Systolic heart failure; T, Thyroid disease. HARMS₂-AF: Hypertension, Age, Raised body mass index, Male sex, Sleep apnoea, Smoking, Alcohol.

First, it is important to emphasise that statistical significance does not equate to clinical significance. Although the C-statistics for the protein risk score and NT-proBNP alone were higher than for the clinical risk score,5 differences were modest, with debatable clinical and practical significance. We must remember that, although identification of high-risk patients may lead to benefits through targeted risk factor modification, it may also lead to overdiagnosis and overtreatment. For example, a high predicted risk of AF in an individual may lead to a 12-lead ECG, which may not demonstrate AF but instead show evidence of possible coronary artery disease, leading to a cascade of investigations and treatments which may be of no appreciable benefit, or worse, result in harm.

This is further compounded by the relative complexity of the protein risk score, which, if it were to be implemented clinically, would introduce an additional step in management. In already oversubscribed and under-resourced healthcare settings, this could result in treatment delays and increased expenditure. In addition, the prevention of AF in individuals with a high protein risk score but without associated comorbidities may prove challenging, in the absence of clear therapeutic targets.

In the present study, proteomic analysis relied on a one-off blood sample measurement, which may require specialist laboratory provisions and costs. In reality, AF risk is dynamic, fluctuating over time and with the development and treatment of comorbidities. For example, it is well-described that AF-associated stroke risk prediction is improved through repeated measures, taking into account these fluctuations.6 This observation, which almost certainly applies to AF risk prediction, is not factored into the current study, and would likely have further impacts on resource usage.

In addition, and as noted by the authors, the analytical methodology used a targeted proteomic approach of known proteins of cardiovascular interest. While this approach is common and efficient, it may introduce bias in the analysis. The use of mass-spectrometry-based approaches has been reported, often combined with high-performance liquid chromatography. These techniques may facilitate the discovery of novel AF biomarkers across the age spectrum.

Beyond clinical risk factors and proteomic models, emerging approaches may offer some solutions. Despite its relative infancy, artificial intelligence (AI) has rapidly revolutionised the field of risk prediction. For example, AI-enabled analysis of the 12-lead sinus rhythm ECG has similar value to clinical risk factor models for predicting incident AF over a 5-year time period.7 This good predictive ability of AI is somewhat offset by a lack of explainability (ie, it may not be apparent which ECG or echocardiographic feature contributes towards increased risk in an individual patient), which can prove challenging for physicians and patients to deal with. The application of AI and machine learning to other technologies—such as wearable health devices and implantable cardiac devices—may offer further promise in AF risk prediction, although requires further study. Additionally, AI may afford the opportunity to evaluate the plasma proteome dynamically to improve risk prediction. Nevertheless, one must remember that proteomics is only one of the ‘omic’ technologies, and further research is required to understand a combined multiomics approach using AI to aid interpretation. Validation in multiple cohorts is also needed, given the ethnic differences in incident AF and AF-related complications.8

Overall, the study by Liu et al provides important new insights into the prediction of AF using proteomics. This work lays the foundations for prospective studies on the additive value of proteomic screening over simpler clinical risk scores, taking into account clinical efficacy and cost-effectiveness. More broadly, the future of AF prediction may lie in machine learning and AI models that incorporate clinical risk factors and multimodality diagnostics (eg, proteomics, metabolomics, ECG and echocardiography) in order to improve performance. Until these are validated, traditional simple clinical risk scores will remain an important tool in the armamentarium of the practicing cardiologist.