Machine-Learning Tool for Predicting ESRD After Heart Transplant

Background: Post-heart transplant end-stage renal disease (ESRD) increases risk of mortality. Simultaneous heart-kidney transplantation decreases this risk. Candidates with estimated glomerular filtration rate (eGFR) between 30 and 60 mL/min/1.73 m2 are ineligible for simultaneous listing and are at risk of ESRD. A model to predict post-transplant ESRD is necessary to guide eligibility in this population. Methods: 13,939 adult heart-alone recipients from the Scientific Registry of Transplant Recipients with CKD-EPI eGFR between 30 and 60 mL/min/1.73 m2 at listing were selected. Patients on dialysis or missing initial eGFRs were excluded. ESRD was defined as eGFR less than 20 mL/min/1.73 m2, chronic dialysis requirement, or renal transplant. Waitlist registration data was incorporated in an Extreme Gradient Boosting machine learning model with 10-fold cross validation to predict ESRD within one year of heart-transplant. Results: The model predicted ESRD with a C-statistic of 0.66 [Figure 1]. At the Youden’s Index threshold, the sensitivity and specificity were 0.80 and 0.42, respectively. The negative predictive value was 0.98. Important predictors were eGFR, Cytomegalovirus status, cardiac output, weight, and African American race. Conclusion: To our knowledge, this is the first model to exclusively utilize data available at waitlist registration to predict ESRD in patients ineligible for simultaneous heart-kidney transplantation. It builds upon the work of Mete et al. which utilized time of transplant data to predict ESRD in candidates with initial eGFR between 0 and 60 mL/min/1.73 m2. Our model’s ability to predict ESRD is limited. Further characterization of biomarkers and cardiac function may be necessary to predict post-transplant ESRD. Funding: NIDDK Support, Private Foundation SupportReceiver Operating Characteristics Curve of Post-Heart Transplant ESRD Machine Learning Prediction Tool