The Colorado Heart Failure Acuity Risk Model: A Mortality Model for Waitlisted Cardiac Transplant Patients

Abstract

Background: Currently, there is no mathematical model used nationally to determine the medical urgency of patients on the heart transplant waitlist in the United States. While the current organ distribution system accounts for many patient factors, a truly objective model is needed to more reliably stratify patients by their medical acuity.

Objectives: The aim of the study was to develop risk scores (Colorado Heart failure Acuity Risk Model [CHARM] score) to predict mortality in adults waitlisted for heart transplant.

Methods: Risk scores were based on multivariable logistic regression models with mortality endpoints at 90 days, 180 days, 1 year, and 2 years. The models included serology data and patient history variables from waitlisted patients (N = 4,176) within the Scientific Registry of Transplant Recipients database from January 1, 2017, to September 2, 2023.

Results: The CHARM score included serum markers (brain natriuretic peptide, creatinine, sodium, aspartate aminotransferase, albumin, total bilirubin) and clinical variables (history of cardiac surgery, prior transplant, willingness to accept an hepatitis C virus positive heart, use of extracorporeal membrane oxygenation, use of mechanical life support, implantation of a cardiac defibrillator, and ventilator support prior to transplant). Sample holdout-validation for the models yielded average area under the curves of 0.825 (90-day), 0.805 (180-day), 0.779 (1-year), and 0.766 (2-year). Risk indices for all models were 99% correlated with observed mortality rates.

Conclusions: The CHARM score provides reliable calibration and prediction, offering an objective system for identifying critically ill patients on the heart transplant waitlist. The CHARM score will be useful in the era of continuous distribution to standardize organ allocation.

Keywords: failure; heart; mortality; prediction; transplant; waitlist.

Graphical abstract

Central Illustration

CHARM (Colorado Heart Failure Acuity Risk Model): A Mortality Model for Waitlisted Cardiac Transplant Patients Cardiac surgery refers to underwent coronary artery bypass grafting, valve replacement or repair, surgery to repair a congenital disease, and left ventricular remodeling. AST = aspartate aminotransferase; AUC = area under the curve; BNP = B-type natriuretic peptide; ECMO = extracorporeal membrane oxygenation; HCV = hepatitis C virus; SGOT = glutamic-oxaloacetic transaminase.

Figure 1

Correlation Heatmap of Independent Variables Pearson’s correlation coefficients were calculated using all patients (N = 4,176) for the 13 independent variables used to construct the 4 models (90 days, 180 days, 1 year, and 2 years). This is used to provide a measure of collinearity. Blue indicates a positive correlation, and red indicates a negative correlation. The color saturation and circle area increase as the correlation coefficients increase in magnitude. No highly significant correlations were observed. BNP = B-type natriuretic peptide; ECMO = extracorporeal membrane oxygenation; SGOT = glutamic-oxaloacetic transaminase.

Figure 2

Tiered-Risk Index System Patient mortality probabilities were calculated for all patients in the cohort (N = 4,176) and are provided as a function of observed patient mortality rate per tier for the (A) 90-day, (B) 180-day, (C) 1-year, and (D) 2-year models. The goodness-of-fit was calculated with 6 tiers using the Hosmer-Lemeshow methodology and was greater than 0.99 for all 4 models. The observed morality rate (OBS), number of patients per group (Tier N), and the number of waitlist patient deaths (Events N) are reported for each tier and model. CHARM = Colorado Heart failure Acuity Risk Model.

Figure 3

Area Under the Receiver Operating Characteristic Curves Using Sample Hold-Out Validation Logistic regression was performed on the training set (N = 2,088) to calculate each model’s coefficients. Predictions were made using the test set (N = 2,088) for the (A) 90-day, (B) 180-day, (C) 1-year, and (D) 2-year models. Upper and lower 95% CIs were provided for each model in brackets. AUC = area under the curve.

Figure 4

Time-Dependent Univariable Analysis of Tiered-Risk System We performed time-dependent univariable analysis to assess each tier and each model. All log-transformed HRs are referenced to tier 1 for the (A) 90-day, (B) 180-day, (C) 1-year, and (D) 2-year models. Upper and lower 95% CIs were provided for each model.