Predicting Health Outcomes through Machine Learning in Pediatric Heart Transplantation: Use of National Data from the United Network for Organ Sharing

Background and Purpose: Rates of survival for pediatric solid organ transplant recipients continue to improve, yet ongoing concerns remain regarding the rates of late acute rejection (LAR) and hospitalization among pediatric heart transplant (HT) recipients. Transplant social work researchers and practitioner have sought to identify risk factors of poorpost-transplant health outcomes for pediatric HT patients. Data-driven modeling and machine learning (ML) approaches have had limited application in social work research. In the current study of pediatric HT, we leverage the availability of robust databases of longitudinal patient electronic health records to identify unique risk factors and their complex relationship with outcomes. ML approaches may aid in identifying high-risk patients beyond traditional statistical approaches. The purpose of the current study was to examine the use of ML models to predict post-transplant health outcomes for pediatric HT recipients.

Methods: For this study we used the national United Network for Organ Sharing (UNOS) data which contains pre-transplant medical information, psychosocial factors, and long-term post-transplant health outcomes of pediatric HT recipients. Various ML and deep learning (DL) models were used to predict late acute rejection, hospitalizations, and mortality at each 1-, 3-, and 5-years post-transplant. The ML models included XGBoost (XGB), support vector machine (SVM), random forests (RF), multi-layer perceptron (MLP), and adaptive boosting (AdaBoost). Variables predicting post-transplant outcomes included donor and recipient medical and psychosocial predictors. SHAP (SHapley Additive exPlanations) were calculated to estimate the importance of each variable for prediction thus increasing the interpretability of results.

Results: A total of 5,807 HT patients from 1987-2019 were identified in the national UNOS data. Logistic regression, RF, and AdaBoost models were the best performing algorithms for different prediction windows across outcomes. When considering area under the receiver operating curve (AUROC), SVM outperformed other approaches for predictions of 1-year, 5-year prediction of rejection (AUROC: 0.668 and 0.712, respectively), 1-year, 3-year, 5-year hospitalization (AUROC: 0.589, 0.595 and 0.591, respectively), and 1-year mortality (AUROC: 0.701). MLP models for 5-year prediction window for mortality achieved the highest AUROC among all tested models for all predictions (AUROC: 0.729).

Examining impact of the predictor variables in terms of mean (|SHAP Value|) across all SVM, RF, and MLP models suggested recipient variables of graft status, recipient malignancies since listing for transplant, recipient ethnicity, and graft status have higher impact on prediction. Important patient psychosocial factors predicting outcomes across the time frames included patient age, ethnicity, level of education, and gender. Female and adolescent patients were shown to be at greater risk for rejection episodes and mortality compared to male or younger patients.

Conclusions and Implications: Machine learning approaches can identify salient predictors and their complex relationship with outcomes, thereby aiding social work researchers and practitioners in identifying at-risk patients, providing a solid foundation for improved psychosocial assessment and care, risk stratification, and guide for decision-making. The current study demonstrates the value of ML approaches for modeling post-transplant health outcomes using patient-level data and informs multidisciplinary transplant teams about the future potential of these innovative approaches to improve psychosocial risk assessments.