Looking forward and backward for sudden death risk: competing risk is everywhere.

Abstract

This article refers to ‘Dynamic changes in cardiovascular and systemic parameters prior to sudden cardiac death in heart failure with reduced ejection fraction: a PARADIGM-HF analysis’ by L.E. Rohde et al., published in this issue on pages 1346–1356. The risk of sudden cardiac death (SCD) has long been known to by dynamic. For example, Solomon and colleagues demonstrated that the months after a myocardial infarction represents a transient high-risk period where the absolute rate of SCD is acutely elevated before declining to a basal and lower rate.1 Likewise, in patients with implantable cardioverter-defibrillators (ICDs), the observed distribution of ICD therapies is non-random, with clear clustering of ventricular arrhythmias.2 The reality of this dynamic SCD risk is buttressed against the typically cross-sectional nature in which we deploy risk assessment in the care of patients. Missing, therefore, is a risk assessment approach that flexibly captures dynamic changes in SCD risk over the course of a patient's lifetime. Such a strategy could have significant implications for prevention of SCD events and, possibly, maximization of ICD benefit. In this issue of the Journal, Rohde and colleagues sought to identify dynamic risk factors for SCD in patients with systolic heart failure (HF).3 The cohort included 8399 patients from the PARADIGM-HF (Prospective Comparison of ARNI with ACEI to Determine Impact on Global Mortality and Morbidity in Heart Failure) trial,4 a randomized controlled comparison of sacubitril/valsartan and enalapril in subjects with chronic, systolic HF and left ventricular ejection fraction (LVEF) ≤40%. Leveraging pre-specified serial assessment during the trial, the authors employed a multivariable model of time-updated markers to identify dynamic predictors of SCD risk. To assess the relationship between trajectories of clinical and serological markers with SCD risk, they performed a ‘look-back’ analysis, working retrospectively from the time of death. A third analysis involved a machine learning algorithm [classification and regression tree (CART) analysis] to identify dynamic predictors of risk. Cognizant of the relevance of competing risk, the authors used logistic regression models to evaluate whether a particular risk factor was differentially associated with the odds of SCD vs. competing non-SCD mortality. The authors identified several time-updated SCD risk factors including hyperbilirubinaemia, advanced New York Heart Association (NYHA) class (III/IV), hypoalbuminaemia, hypokalaemia, systolic blood pressure <100 mmHg, hyperuricaemia, and a total cholesterol <135 mg/dL. Of note, each of these seven time-updated risk factors were also significantly associated with the risk of non-SCD; in five instances with a similar magnitude of association between the two death types and in two instances (NYHA class III/IV, hypoalbuminaemia) demonstrating a stronger association with non-SCD as compared to SCD. Harmonized with the time-updated approach, look-back analysis identified temporal trends in hypoalbuminaemia, hyperuricaemia, and hyperbilirubinaemia in the months prior to both SCD and non-SCD. In a proof-of-concept machine learning approach, the authors demonstrated that incorporation of baseline markers (ICD status, natriuretic peptide level) and time-updated variables (total bilirubin, total cholesterol, NYHA class) yielded a sixfold gradient of risk across low and high-risk subsets of the cohort. While the authors should be congratulated for tackling the important challenge of capturing dynamic SCD risk, there are several principles to consider when evaluating the clinical implications of these findings. First, the risk of SCD and the risk of competing, non-SCD are closely correlated. As our group and others have shown previously,5 traditional risk markers for SCD (e.g. diabetes, advanced NYHA class, atrial fibrillation, lower LVEF) are similarly potent markers for non-SCD mortality in patients with cardiovascular disease. The principle findings of this study would suggest that this correlated risk remains strong in a time-updated paradigm. At least one explanation for this, as the authors appropriately note,3 is the interconnected pathophysiology of HF mortality and SCD risk. To that end, the expanding armamentarium of therapies that lower HF mortality6 would also be anticipated to lower SCD risk. Indeed, the lowering of SCD risk with HF pharmacotherapy has been shown consistently over the past two decades including for beta-blockers,7 renin–angiotensin–aldosterone system inhibitors8, 9 and more recently, sacubutril/valsartan.10 A second possible explanation for correlated SCD and non-SCD risk is the nature of the biomarkers assessed. Several of the markers considered, including NYHA class or serological markers of hepatic congestion (e.g. hyperbilirubinaemia), are directly implicated in the physiology or phenotyping of clinical HF. It is then not particularly surprising that these markers predict both SCD and non-SCD mortality. Looking ahead, markers that are more proximal to the causal pathobiology of ventricular arrhythmias (e.g. ventricular scar phenotyping, electrical repolarization heterogeneity)11, 12 are likely to more specifically identify arrhythmic SCD risk. What, then, are the implications of identifying dynamic SCD risk? For patients with HF, identification of dynamic SCD risk – with sufficient lead-time – could lead to earlier identification of at-risk patients and potential deployment of relevant therapies including HF optimization, earlier consideration of advanced HF therapies, and deployment of anti-arrhythmic strategies. This study's findings3 would suggest that trajectories of worsening clinical HF (e.g. advancing NYHA class) or worsening HF status (e.g. worsening hepatic congestion reflected by hyperbilirubinaemia) may warrant such consideration, though the ‘lead-time’ of these changes prior to an SCD event is unclear. Moving beyond time-updated clinical variables or serological markers – which may not be particularly practical to measure – there is expanding interest in the use of multi-sensor algorithms incorporated into implantable devices to identify incipient risk. One such multi-sensor algorithm13 was able to identify HF hospitalization with a 30-day lead-time. Future work employing multi-sensor algorithms to identify SCD risk are warranted. Identification of dynamic risk – both SCD and non-SCD – may also serve to maximize the survival benefit of preventative therapies such as the ICD. Previous randomized trials of ICD therapy in patients at transient and high ‘absolute risk’ of SCD – such as immediately after myocardial infarction1 – have failed to demonstrate survival benefit despite significant lowering of SCD risk.14, 15 As our group and others have demonstrated, the survival benefit of ICD therapy is a function of both the absolute and proportional risk of SCD.16, 17 We congratulate the authors for their thoughtful consideration of competing risk in this study,3 and would highlight that given the co-association of the SCD risk factors identified with non-SCD mortality, we would not anticipate that those identified at highest risk of SCD in this study would necessarily be more likely to benefit from ICD therapy. As we have proposed previously,18 for any given absolute risk of SCD, the proportional risk of SCD serves as a ‘rheostat’ for ICD survival benefit (Figure 1A).19, 20 Indeed, when assessed in prior randomized ICD trials, increasing absolute SCD rates were associated with attenuated survival benefit as these patients had an even greater absolute risk of non-SCD mortality, thereby lowering their proportional risk of SCD (Figure 1B).16, 17 As the absolute rates of SCD have declined in patients with HF,21 future clinical decision-making tools, which incorporate both dynamic absolute and proportional risk of SCD, could help best identify patients most likely to benefit from ICD therapy. Finally, the dynamic risk paradigm has important implications for clinical trial design and staged decision-making. For example, for patients identified to be at increased absolute risk of non-SCD mortality (i.e. low proportional risk of SCD), an initial therapeutic strategy of mitigating HF mortality risk may be most appropriate. After this initial therapeutic deployment, for patients in whom non-SCD risk can be significantly lowered (i.e. increasing proportional risk of SCD), an ICD may serve as a meaningful tool to extend survival. Conversely, in patients in whom competing risk cannot be lowered, an accelerated consideration for advanced HF therapies may be more appropriate. Incorporating dynamic reassessment of risk could also serve as an important feature of future adaptive clinical trial designs for SCD prevention.22 Taken together, this study from Rohde and colleagues represents an important step in reformulating our paradigm for SCD risk assessment from static to dynamic.3 The study highlights the closely intertwined risk of SCD and non-SCD in patients with HF – even when considering dynamic trajectories of risk. By explicitly considering competing risk, the authors have helpfully framed the role of dynamic assessment of cause-specific risk which could ultimately help sharpen our assessment of absolute and proportional SCD risk over time. Given the constraints of the contemporary SCD paradigm – which is cross-sectional and focused on absolute risk – it is exactly this flexible and dynamic evaluation of cause-specific risk that we see as fundamental to SCD risk stratification, identification, and prevention. Conflict of interest: N.A.C. reports no disclosures relevant to the content of this manuscript. W.C.L. is a consultant to Medtronic with research grant, clinical endpoint committee member for CardioMems trials (Abbott, Baim Institute), NT-proBNP (Siemens and Beckman Coulter, Baim Institute) and SOVLE-CRT (EBR Systems Inc), steering committee member for Cardiac Dimensions and Respircardia. The University of Washington CoMotion holds the copyrights for the Seattle Heart Failure Model and the Seattle Proportional Risk Model and has received license fees from various companies.