Risk prediction in heart failure using invasive hemodynamics

From January 2010 to December 2017, 2205 patients with heart failure with reduced ejection fraction underwent right heart catheterization at the department for cardiology at the University of Heidelberg (the cause for catheterization is listed in Table 1). Heart failure with reduced ejection fraction was defined as a left ventricular ejection fraction (LVEF) below 40% measured by echocardiography or as evaluated by left ventriculography. For 883 patients, high sensitivity Troponin T (hs-cTnT) as well as N-terminal pro-brain natriuretic peptide (NT-proBNP) prior to right heart catheterization were available (Supplementary Table 1). Both ambulatory and hospitalized patients were included. The patients on inotropes (n = 59) or intravenous diuresis were not excluded (n = 135). To calculate a more accurate staging system using cardiac biomarkers, patients without measurement of hs-cTnT and NT-proBNP were excluded from this analysis. Ischemic cardiomyopathy was defined as an impairment in left ventricular ejection fraction caused to a relevant part by coronary artery disease. Dilated cardiomyopathy (DCM) was diagnosed using echocardiographic parameters and cardiac magnetic resonance. If severe valve disease was the main cause of impaired left ventricular systolic function, the patients were classified as having valvular heart disease. Other causes of heart failure included cardiac amyloidosis, non-compaction-, hypertrophic-, chemotoxic-, and restrictive cardiomyopathy, hemosiderosis, myocarditis, muscular dystrophy and cardiomyopathy of unknown cause.

This included the patient’s medical history, cardiovascular risk factors, clinical assessment including evaluation of New York Heart Association (NYHA) class at the time of diagnosis. Further, complete laboratory workup including cardiac biomarkers (hs-cTnT, NT-proBNP), and serum creatinine, was done in all patients. Glomerular filtration rate (GFR) was calculated by using the Modification of Diet in Renal Disease formula. For hs-cTnT, cut-off value was < 14 pg/ml. Right heart catheterizations via a femoral venous approach were performed to determine CI, PAP, PVR and SvO₂ as described before [13]. Cardiac index was determined by saturation measurement according to the Fick principle. The pulmonary artery pressures, mean PAWP and mean right ventricular and right atrial (RA) pressures were measured during end expiration breath hold at baseline for at least three heart cycles. Mean PAP was calculated by Metek software (Metek GmbH, Roetgen, Germany). Pulmonary vascular resistance was calculated as (mean PAP − PAWP)/cardiac output.

For the risk stratification analysis, the combined endpoint was all-cause mortality, heart transplantation (HTX) or left ventricular assist device (LVAD) implantation. Follow-up was obtained by review of the patients’ hospital chart. If the follow-up could not be completed an inquiry to the responsible population registration was conducted. In case follow-up could not be completed, the date of the last visit was recorded as a censored event. All-cause mortality was the secondary endpoint excluding patients undergoing HTX.

Besides a new score developed in this study, previously established risk models were applied to our patient cohort. All established risk models were multivariable models for the prediction of all-cause mortality. The SHFM score was calculated by multiplying the β coefficient by the variable and summing the values as described by the authors [10]. These variables included: age, sex, NYHA class, LVEF, ischemic/non-ischemic cardiomyopathy, systolic blood pressure, use of angiotensin converting enzyme inhibitors (ACEI), use of diuretics, serum sodium concentration, hemoglobin concentration, lymphocyte count, serum uric acid concentration as well as cholesterol concentration in serum. The MAGGIC multidimensional risk score was calculated attributing points to each variable as described in the original publication [14]. Those variables included: gender, smoking status, diabetes, chronic obstructive pulmonary disease (COPD), time of heart failure diagnosis, ACEI or beta blocker use, LVEF, NYHA, creatinine concentration, body mass index (BMI), systolic blood pressure and age.

Continuous data are expressed as median and 25% and 75% percentile [Q1; Q3]. The categorical variables are expressed as absolute numbers and percentages. For the management of missing data, we employed a MICE (Multiple Imputation by Chained Equations) imputation approach to create 50 multiply imputed datasets to ensure stability of the imputations. Each missing value was imputed under a predictive model using the fully conditional specification method, where each incomplete variable is modeled conditionally given the other variables. Once imputations were complete, the analyses were performed on each of these datasets separately. The final estimates and their standard errors were combined by taking the average estimates of the 50 imputed datasets. At first variables of interest were entered in univariate Cox’ proportional hazards model with the combined endpoint as a dependent variable. In a second step invasive hemodynamic parameters were entered separately in a Cox proportional hazards model adjusted for age, sex, creatinine, presence of ICM, hs-cTnT and NT-proBNP to assess the prognostic value of invasive hemodynamics. In a next step, the multiple models were also adjusted for heart rate, atrial fibrillation during the procedure and history of atrial fibrillation. The hemodynamic parameters were not entered all together in a multiple model due to a high collinearity between the different variables. For variable selection and to test a possible benefit from using a penalized model, in a third step mean PA pressure, mean RA pressure, mean PAWP, age, NT-proBNP, hs-cTnT, SVO_2, creatinine and presence of ischemic cardiomyopathy were selected as predictors in a Least Absolute Shrinkage and Selection Operator (LASSO) based Cox proportional hazards model using 50 imputed datasets. As systolic and diastolic and mean PA pressure are all highly correlated variables, only mean PA pressure was chosen. Mixed venous oxygen saturation was chosen over the CI due to its superior performance in the multiple Cox’ model. As pulmonary resistance is calculated using PA pressure and PAWP it was as well excluded. The regularization parameter (lambda) in the LASSO procedure was determined through fivefold cross-validation and was chosen as the value that minimized the average cross-validated prediction error. The combined endpoint consisting of all-cause mortality, need for heart transplant or left ventricular assist device implantation was assessed as the outcome of interest. To assess whether the LASSO model has an additional value, the same variables were entered in an unpenalized Cox’ proportional hazard model. Concordance index of both models, the penalized and unpenalized was compared. We then further developed models separately for ischemic and non-ischemic cardiomyopathy. To compare the performance of the new model with established heart failure models, binary logistic regression was calculated with event free survival at 6 months, 1 year and 2 years was calculated and then compared with the SHFM and the MAGGIC scores. The comparative performance of the predictive models (new model, SHFM, and MAGGIC) was evaluated using DeLong’s test. Since the SHFM and MAGGIC Score were created to predict all-cause mortality, the comparison was also conducted with all-cause mortality as an endpoint. The results of the binary logistic regressions and DeLong’s tests were visualized using ROC curves, with the AUC providing a measure of the overall predictive accuracy of each model. Model performance was further compared using Akaike’s Information Criteria. To further investigate in which patient cohort the new model might be most useful in, it was divided into a “high risk” and “low risk” group. This was conducted using the optimal cut-off by Youden’s Index of the receiver-operator characteristics (ROC) curve predicting 1-year event free survival. The predictive value of the SHFM and MAGGIC scores was further investigated in the subcohort of patients who died within 1-year follow-up period and were classified as “high risk”. In this subgroup, the number of correctly and falsely predicted patients was counted for each score. The differences in prediction was also expressed as an absolute risk reduction to calculate a potential “number needed to catheterize”.

Eight-hundred-and-eighty-three patients with at least moderately reduced LVEF as defined by left ventriculography or echocardiography and complete right heart catheter workout and cardiac biomarkers including hs-cTnT and NT-proBNP were identified. Eighty-eight patients were lost to follow-up as of the 1 st of January 2020. The most frequent indication for right heart catheterization was the initial workup for heart failure (n = 233, (26%)) followed by catheterization as part of HTX listing (n = 222 (25%) and a suggested valvular heart disease (n = 217 (26%)). Further details about indication for invasive measurement are given in Table 1. Three-hundred-and-thirty-three patients (38%) died during follow-up, 63 patients underwent implantation of a left ventricular assist device and 72 patients received heart transplantation. The median time until the endpoint was reached was 4.9 years (IQR 1.9 years, 75% quartile not reached). Heart failure was at least partially caused by ICM in 359 (41%) cases, by DCM in 442 (50%) cases and by valvular heart disease in 303 (34%) cases. In 292 (33%) patients DCM was the only relevant cause of heart failure, 163 were diagnosed with isolated ICM (19%) and 46 (5%) with valvular heart disease as the only relevant cause of heart failure. Sixty-one cases with other causes of heart failure included 24 (3%) cases of cardiac amyloidosis, 10 (1%) cases of hypertrophic cardiomyopathy and 9 (1%) cases of cardiomyopathy with unknown cause. Other causes of heart failure were non-compaction cardiomyopathy, vasculitis, restrictive cardiomyopathy, hemosiderosis, and autoimmune related. The patient characteristics are shown in detail in Table 1 and Table 2. The missing data are listed in Supplementary Table 3.

Six-hundred-and-eighty-six patients (78%) were classified as either NYHA class III or IV at baseline and 589 (67%) patients showed heart failure symptoms according to INTERMACS stage VI or lower. The hemodynamic parameters support the notion of a severe reduced LVEF with a mean CI of 2.05 [1.73; 2.40] l/min/m² and mean PAWP of 22.00 mmHg [17.00; 28.00]. Mean systolic PA pressure was elevated (32.00 mmHg [25.00; 40.00]) as well as mean PVR (183 dyn*s/cm⁵ [117; 249]) [15]. Cardiac biomarkers were elevated with mean NT-proBNP of 6.237 pg/ml [2575; 13337] and mean hs-cTnT of 33 pg/ml [18; 64]. One-hundred-and-forty-three patients had cardiac resynchronization therapy (16%). All heart failure parameters are listed in Table 2.

A multiple Cox’ proportional hazard model using invasive hemodynamics adjusted for age, creatinine, cardiac biomarkers, sex, and presence of ICM and the combined endpoint of all-cause mortality, heart transplantation or left ventricular assist device implantation as dependent variable revealed a statistically significant predictive value for all invasive parameters. The systolic PA pressure showed to be associated with a two percent increase in the risk of reaching the combined endpoint per unit increase in mmHg (Cox proportional hazards ratio 1.02 [1.02; 1.03], p ≤ 0.001). Mean RA and PA pressures showed similar results (Cox’ proportional hazard ratio 1.03 [1.02; 1.03] and 1.03 [1.02; 1.04], p ≤ 0.001 in both) as well as diastolic PA pressure and mean PAWP (Cox’ proportional hazard ratio 1.01 [1.00; 1.01] and 1.03 [1.01; 1.04], p ≤ 0.001 in both). Pulmonary arterial pressure > 20 mmHg showed a hazard ratio of 2.05 [1.38; 3.06], RA pressure > 15 mmHg a hazard ratio of 1.96 [1.57; 2.44] and PAWP > 15 mmHg was associated with a hazard ratio of 1.77 [1.29; 2.44] (p < 0.01 for all). Mixed venous oxygen saturation was associated with a four percent decrease in reaching the combined endpoint per % increase (Cox proportional hazards ratio 0.96 [0.95; 0.97], p < 0.001) and the Cardiac Index was associated with a hazard ratio of 0.67 [(0.56; 0.80], p < 0.001). Pulmonary vascular resistance was a significant predictor of the endpoint in the overall cohort and both subgroups (HR 1.0014, 1.002, and 1.0013 respectively, p < 0.001 for all). In the subgroup of patients with non-ischemic cardiomyopathy all invasive parameters showed a significant association with the combined endpoint whereas diastolic and mean PA pressure and mean PAWP were not associated in the subgroup of patients with ischemic cardiomyopathy. Results of the multiple Cox’ regression are given in detail in Table 3. Results of the univariate Cox’ proportional hazards model are listed in detail in Supplementary Table 4. We additionally calculated the model excluding patients on inotropic therapy or with diagnosis of amyloidosis to generate a more homogeneous patient cohort (Supplementary Table 6). The models were also further adjusted for heart rate and presence of atrial fibrillation during the catheter (Supplementary Table 7) and history of atrial fibrillation (Supplementary Table 8). All findings were basically replicated in the cohorts (ICMP/non-ICMP) and invasive hemodynamics remained a strong predictor for the primary outcome.

Mean PA pressure, mean RA pressure, mean PAWP, age, NT-proBNP, hs-cTnT, SVO_2, creatinine as well presence of ischemic cardiomyopathy, age and sex were entered as independent variables into a Cox’ proportional hazards model. Following the LASSO regression, the coefficients for mean PAWP and sex were reduced to zero, indicating that these variables did not significantly contribute to the model’s predictive power in the context of other included variables. Subsequently, the same set of variables was utilized to construct a standard (unpenalized) Cox proportional hazards model. C statistic in both models regulated and unregulated showed no relevant difference in C-statistic in all imputed data sets (C-statistic 0.67 for both models). Therefore, lambda tuning with consecutive shrinkage of variables did not result in an overall improvement of event free survival prediction. We therefore decided to keep all variables as predictors for a new model.

An overview of the different parameters used in the established and new models is given in supplementary Table 2. To compare the effect of the new model including mean PA pressure, mean RA pressure, mean PAWP, age, NT-proBNP, hs-cTnT, SVO_2, creatinine as well presence of ischemic cardiomyopathy, and sex with established risk scores, all three models were entered separately into a binary logistic regression model with the combined endpoint as well as all-cause mortality as a dependent variable. The newly created model showed slightly better performance for all three timepoints, 6 months, 12 months and 24 months with an area under the curve of 0.77, 0.79, and 0.77, for the combined endpoint respectively and an area under the curve of 0.77, 0.76, and 0.77 for all-cause mortality. Model performance measured by AIC was lowest in the new model for both endpoints and all timepoints except for overall survival at the six months follow-up. Akaike’s Information Criteria results are given in detail in the Supplementary Table 5. The results of the binary logistic regression are illustrated in Figs. 1a, b, 2a, b and 3a, b.

Bild vergrößern

Bild vergrößern

Bild vergrößern

Event free survival at one year was 74% in this cohort. The SHFM predicted 92%, the MAGGIC score 73% and our model 72%. This corresponds to false prediction of outcome in 163 patients using the SHFM and 8 patients using the MAGGIC score.

Using Youden’s Index for an optimal cut-off of 0.33 to predict 1-year event free mortality, the new model was divided into a “high risk” and “low risk” group. Event free survival is further illustrated in Fig. 4. One-hundred-and-forty-two patients out of 172 (82%) “high risk” patients died within 1 year. This cohort was characterized by an elevated mean NT-proBNP (12,438 pg/ml [5912; 23603]), low CI (1.77 l/min [1.51; 2.10]) and low SVO2 (48% [44; 52]) when compared to the overall cohort and rather moderately elevated hs-cTnT of 55.00 pg/ml [28.00, 89.50]. In this sub cohort, the predicted 1-year mortality using the SHFM was 13% and 30% using the MAGGIC score while the actual one-year mortality was 66%, as predicted by the new model. Therefore, in this “high-risk” cohort the SHFM would have underestimated a fatal outcome in 76 patients and the MAGGIC score in 51 patients. In the “low risk” subgroup no patients died within the first year. Predicted mortality according to SHFM was 11% and 28% using the MAGGIC Score while the new model predicted a one-year mortality of 0.

Bild vergrößern

This study was conducted in a single-centre retrospective approach and data was not complete. There was no external validation cohort. The patients without complete NTproBNP were excluded in this study, which could contribute to a selection bias. Another limitation is the heterogenous patient population we used. Both ambulatory and hospitalized patients were included as well as patients on inotropes and at very different stages of heart failure. This is also the explanation for the big differences in medication between patients. The vastly different indications for invasive measurement are another cause of bias. However, considering the positive results for hemodynamic measurement underlines the utility of right heart catheterization in a broad patient population. Although there are limitations, such as the reliance on Fick’s principle for invasive hemodynamic measurements, the study provides a foundation for potential advancements. While acknowledging the need for further validation in a separate cohort, the newly developed model holds promise for improving our understanding of cardiac dynamics and patient outcomes. It must be stated that the SFHM and MAGGIC are universal models that can be applied to both patients with HFrEF and HFpEF and have been sufficiently validated in external cohorts. However, the main goal of these models is to predict all-cause mortality. Predicting hospitalization with these models must be done with caution. In addition, invasive measurement is a procedure with potential risks and complications which heart failure models do not carry. The fact that invasive hemodynamic parameters performed better in patients with non-ischemic cardiomyopathy could potentially be attributed to a Type II error, and therefore, by chance. However, the findings were consistent for multiple parameters adjusted for different potential confounders. Still, as mentioned above, further external validation is needed.