Abstract MP75: Recalibration and Additional Data Domains Leads to Modestly Improved Performance of Risk Calculators for Heart Failure Readmission

Abstract

Introduction: Heart failure (HF)-related hospitalizations are a growing public health burden. We evaluated two published risk calculators for predicting 30-day readmission after HF hospitalizations: 1) using the original coefficients, 2) updating the coefficients 3) developing a new model with additional variables and updated coefficients. Hypothesis: Recalibrating model coefficients and adding variables would improve the performance of existing 30-day readmission risk calculators. Methods: We identified 45,059 adults hospitalized for HF between 2012-2017 within Kaiser Permanente Northern California, an integrated healthcare delivery system. We used split sampling for development and validation testing. The risk calculators tested included: LACE+ Index and Yale CORE. We used logistic regression on our population to derive the recalibrated coefficients. For the model with additional variables, we included all variables used in the original models plus additional variables, including cardiovascular medication use and socioeconomic status. We used gradient boosting with k-fold cross validation to avoid overfitting. We assessed model performance using area under the curve (AUC) and calibration plots. Results: Discrimination (AUC) was poor using original models: LACE+ [0.56 (0.54-0.58)] and Yale CORE [0.55 (0.54-0.57)]. Recalibrating coefficients resulted in small improvements for LACE+ [0.58 (0.57, 0.60)] and Yale CORE [0.58 (0.57, 0.60)]. Adding variables resulted in a modest improvement for the gradient boosting model [0.61 (0.59, 0.62)]. Calibration plots (Figure 1) showed good calibration except for the Yale CORE model with the original coefficients. Conclusions: Recalibrating coefficients and incorporating prior medication and socioeconomic status led to modest, significant improvements in discrimination while maintaining good calibration. However, overall performance improvements are needed to increase the utility of these published risk calculators to predict readmission.