Predictably disappointing: Modeling risk of right heart failure in LVAD recipients

Compared with older devices, the latest generation of left ventricular assist devices (LVADs) continues to reduce the incidence of many complications including mechanical failure, pump thrombosis, and stroke. Short-term survival of 88% at 12 months with LVAD support now approaches that of heart transplantation. The incidence of right heart failure (RHF) in approximately one-third of LVAD patients, however, remains stubbornly high and continues to be a leading cause of morbidity and mortality in this population (Figure 1).1–5 Considerable attention has been given by a number of investigators over past decades to developing RHF clinical prediction models that could be used to identify high-risk patients to improve patient selection and lessen the incidence of this too common and serious complication.6–10 Unfortunately, despite strong performance in their derivation cohorts, none of these prediction models have lived up to their initial hype, with universally modest performance in external validation cohorts.11 Therefore, much to our dismay as heart failure cardiologists, none have found their way into routine clinical practice or consensus guidelines. Incidence of early right heart failure in major left ventricular assist device trials. Incidence of early right heart failure in major left ventricular assist device trials. One criticism of prior RHF prediction models is that they were derived from older cohorts that did not include contemporary devices and surgical techniques. To address these concerns, Soliman et al.12 recently used the EUROMACS registry to derive a new RHF risk score. Risk factors included in the model were well known to the mechanical circulatory support community and included INTERMACS profile, use of multiple inotropes, severe echocardiographic right ventricular (RV) dysfunction, RAP: pulmonary capillary wedge pressure (PCWP) ratio, and haemoglobin. This score, derived on a larger, more contemporary cohort, was simple to calculate, and performed modestly well with regards to discrimination and calibration on a random split sample from within the EUROMACS registry, raising the question of whether this model would be different than those that came before. Could it become relevant at the bedside, or will it too be relegated largely to discussion at the podium? In this issue of the Journal, Rivas-Lasarte and colleagues present a rigorous external validation of the EUROMACS risk score in a cohort from three centres in the USA, comprising 662 patients over a 10-year period.13 There are several salutary aspects of this study. Firstly, outcomes were able to be adjudicated using a definition identical to that of the EUROMACS score. While there is a robust ongoing debate about the best definition of RHF, maintaining consistency in outcome definition from model derivation to validation is critical for assessing model performance. Secondly, the authors also assessed calibration, which is a critical but often overlooked aspect of model performance. Finally, the authors assessed the net benefit of using this model with decision curve analysis, which quantifies the actual increase or decrease in utility generated if the model were used to motivate clinical decisions. Unfortunately for the many clinicians who continue to grapple with the question of which RV is ‘good enough’ for univentricular left-sided support, but perhaps unsurprisingly, the investigators found that the EUROMACS score did not discriminate well between patients who did or did not develop RHF. The c-statistic of 0.64, compared with 0.70 in the original cohort, represents a 30% loss of discriminatory ability. Furthermore, while the cohort in this analysis was evenly divided into low, intermediate, and high predicted RHF risk according to the EUROMACS score, the incidence of RHF in each of these risk groups was higher than that of the derivation cohort. This miscalibration suggests a systematic underestimation of RHF risk when applying the EUROMACS score to this cohort. What accounts for this miscalibration? There are signs that patients in this cohort were sicker than those in the EUROMACS cohort, as evidenced by a higher incidence of preoperative mechanical circulatory support use and a higher right atrial pressure (RAP) compared with the original derivation cohort. There is also a suggestion that regional practice variation may have artificially inflated the incidence of RHF adjudicated in these US centres. In particular, the use of postoperative iNO beyond 48 h, a component of the composite RHF outcome, occurred in 17% of patients in the US validation cohort, but only IN 1% of patients in the European derivation cohort. In contrast, right ventricular assist device (RVAD) use, for which there is likely less intra-centre variation, was more similar (11% vs. 7%). Indeed, the EUROMACS model was modestly better at predicting the need for RVAD than the need for prolonged iNO. The inclusion of decision curve analysis helps to further highlight the limited clinical utility of the EUROMACS RHF prediction model. Whenever a clinical prediction model is developed, there must be a clear risk-sensitive decision in mind. In the case of RHF prediction models, there are several potential decisions that are relevant, like the decision to proceed with LVAD implant at all, or the decision to initiate early RV temporary mechanical circulatory support. The critical threshold of predicted RHF risk that would trigger a different treatment approach varies for each of these decisions. For example, one might withhold LVAD entirely from a 75-year-old patient ineligible for heart transplantation only if the predicted risk of RHF were very high, since few alternative therapies exist for such a patient. On the other hand, when deciding to use a temporary therapy to prevent RHF, one may do so at a much lower threshold of predicted RHF risk, since the risks of the therapy are much lower and the potential benefit greater. In the case of the EUROMACS model, Rivas-Lasarte et al. found that the model was useful in this cohort only if this critical threshold was between 30% and 50%, and even then, only modestly. The model may also be harmful if the critical threshold is higher than 50% by under-predicting risk among high-risk individuals. This study confirms what many of us have known for years: predicting RHF after LVAD implant remains difficult and we do not seem to be any closer now than we were 25 years ago. Then and now, those who are most clinically compromised are those that are most likely to struggle.10,14 With this frustrating lack of progress, one may wonder where the field should go next. Is there any way to improve RHF prediction? One could consider adding intraoperative data to current RHF prediction models, as Soliman et al.12 did in their postoperative EUROMACS score. This is appealing in that intraoperative events like duration of cardiopulmonary bypass, bleeding, hypoxia, acute lung injury, and hypotension likely increase RHF risk and may not be correlated to preoperative variables. However, while this may help a model perform better on paper, it negates its utility to guide preoperative planning and clinical decision-making. Some may advocate for continuing the search for the next best preoperative covariate, like a novel haemodynamic or imaging parameter, to enhance future model performance. However, it is hard to imagine that there exists some as yet unknown metric strongly associated with RHF and sufficiently uncorrelated to the dozens of existing variables to meaningfully improve prediction accuracy. It remains our estimation that the complex and heterogeneous nature of post-LVAD RHF (Figure 2) is such that no ‘score’ will ever adequately discriminate RHF risk in this population. Moreover, a recent pilot investigation utilizing pressure–volume analysis, demonstrated a heterogeneous impact of postoperative LVAD flow augmentation on RV performance.15 This interventricular dependent effect of LVAD–RV interactions is yet another variable that cannot be accounted for in RHF modelling. Nonetheless, even if RHF prediction models are not sufficiently accurate to improve clinical decision-making, perhaps they can still be used to study prevention or treatment of RHF, where significant gaps in knowledge remain. For example, while early use of RV mechanical circulatory support for severe RV dysfunction is favoured,16 the efficacy of pre-emptive temporary RV mechanical circulatory support (MCS) is unknown. Similarly, the value of tricuspid valve repair, either surgically or percutaneously, needs further clarification. And finally, more research is needed on medications like pulmonary vasodilators and digoxin, as current data are limited and inconclusive.17,–19 In each of these cases, model miscalibration is less of a concern, and RHF prediction models with even modest discriminatory power could still enrich study populations with high-risk patients and improve clinical trial efficiency. With sustained effort in these directions, perhaps RHF prediction models may yet deliver on their promise to reduce the morbidity associated with RHF, even if we continue to have difficulty predicting who will get it. Factors contributing to right heart failure. Factors contributing to right heart failure. Conflict of interest: For Dr. Kiernan, Medtronic: Advisory board, consultant and speaking honoraria. This editorial refers to ‘Prediction of Right Heart Failure after Left Ventricular Assist Implantation: External Validation of the EUROMACS Right-Sided Heart Failure Risk Score’ by Rivas-Lasarte et al. The opinions expressed in this article are not necessarily those of the Editors of the European Heart Journal : Acute Cardiovascular Care or of the European Society of Cardiology.