# Predicting Simultaneous Heart Kidney Allocation and Posttransplant Adverse Kidney Outcomes

## TL;DR

This study develops a machine learning model to predict the need for simultaneous heart-kidney transplants or poor kidney outcomes after heart transplants, aiming to improve decision-making for patients.

## Contribution

A novel random forest model is introduced to predict adverse kidney outcomes or the need for SHKT after heart transplantation using OPTN data.

## Key findings

- 13.4% of heart transplant recipients required SHKT or had adverse kidney outcomes within a year.
- The model showed high specificity and negative predictive value but moderate sensitivity and positive predictive value.
- The model's c-statistics (0.849–0.899) indicate strong ability to differentiate between outcomes.

## Abstract

For individuals with both end-stage heart failure and end-stage kidney disease or persistent acute kidney injury (AKI), simultaneous heart-kidney transplantation (SHKT) emerges as a viable treatment option, potentially yielding superior survival rates compared with heart transplantation (HT) alone. Nevertheless, accurately forecasting kidney recovery following HT in patients with moderate kidney failure poses challenges, thereby complicating the decision-making process for SHKT.

This study employed a random forest (RF) machine learning algorithm, using 15 variables with the highest feature importance scores in the Organ Procurement and Transplantation Network (OPTN) data in which we analyzed a retrospective cohort of adult HT recipients from October 18, 2018 to December 31, 2020 in the US, with a follow-up for at least 1 year. The algorithm’s goal was to predict a composite binary outcome with a calculated probability. An adverse outcome included the need for SHKT or adverse kidney outcomes within the first-year posttransplant (defined as end-stage kidney disease requiring chronic dialysis, glomerular filtration rate (GFR) ≤ 20 ml/min per 1.73 m2 or listing for retransplant). The model underwent both internal and external validation.

Of the 6579 patients in the study cohort, 13.4% received SHKT or experienced adverse kidney outcomes within a year following HT (n = 880). The RF model demonstrated a high specificity (0.941–0.955) and negative predictive value (0.940–0.955). However, it exhibited a moderate level of sensitivity (0.605–0.694) and positive predictive value (0.604–0.680). The concordance (c)-statistics ranged between 0.849 and 0.899, indicating effective class differentiation.

This tool supplements, not replace, clinical judgment in addressing the complexities of SHKT decision-making at the time of waitlisting.