A Machine Learning Model for Post-Transplant Lymphoproliferative Disorder Using the UNOS Database

Abstract

<h3>Purpose</h3> Post Transplant Lymphoproliferative Disorder (PTLD) is a complication noted in cardiac allograft recipients. This study was undertaken to generate a risk prediction model for PTLD at 1, 3 and 5 years post heart transplant using machine learning models. The United Network for Organ Sharing (UNOS) database was used for data generation <h3>Methods</h3> The UNOS database was probed from 1987 to 2015<b>.</b> After excluding retransplants and those with incomplete information regarding malignancy, 39696 adult patients were identified out of which 884 patients had PTLD. A Random Under-sampling Boosting ("RUSBoost") technique in combination with Decision Tree models ("ensemble of models") was used to create a PTLD prediction model. Random Under-sampling technique was combined with boosting (AdaBoost M2) to enhance the prediction model. Variables with higher predictive power were used for testing and training. A 5-fold cross validation was performed to train and test them on five different subsets of data. Model performance was evaluated using average Area Under the Curve (AUC) using Receiver Operating Curves (ROC) for the 5 test subsets <h3>Results</h3> The significant continuous variables for predicting PTLD were ischemic time, recipient serum creatinine, and presence of B2, DR1,DR2 antigens in recipient serum. Of the categorical variables investigated gender, diabetes in the recipient, ethnicity, CMV status of the donor and induction with azathioprine, OKT3,thymoglobulin or cyclosporine were significant. The AUCs ranged from 0.76 to 0.79 and 0.66 to 0.7 in the testing or validation batch respectively. <h3>Conclusion</h3> This is the first report of using machine learning algorithms to derive a risk prediction model for PTLD. This model has a robust AUC for predicting PTLD at years 1, 3 and 5 posttransplant. The limitations of this study are its retrospective nature and that it needs external validation.