An evidence-based data science perspective on the prediction of heart failure readmissions

The prevention of unplanned 30-day readmissions of patients discharged with a diagnosis of heart failure (HF) remains a profound challenge among hospital enterprises. Despite the many models and indices developed to predict which HF patients will readmit for any unplanned cause within 30 days, predictive success has been meager. Using simulations of HF readmission models and the diagnostics most often used to evaluate them (C-statistics, ROC curves), we demonstrate common factors that have contributed to the lack of predictive success among studies. We reveal a greater need for precision and alternative metrics such as partial C-statistics and precision-recall curves and demonstrate via simulations how those tools can be used to better gauge predictive success. We suggest how studies can improve their applicability to hospitals and call for a greater understanding of the uncertainty underlying 30-day all-cause HF readmission. Finally, using insights from sampling theory, we suggest a novel uncertainty-based perspective for predicting readmissions and non-readmissions.