The Journal of Heart and Lung Transplantation
International Society for Heart and Lung Transplantation.
(2)| Volume 41, ISSUE 4, SUPPLEMENT , S11-S12, April 2022

Using Machine Learning to Develop a Contemporary Primary Graft Dysfunction Prediction Model: The International Consortium on PGD


      Primary Graft Dysfunction (PGD) is a leading cause of death early after heart transplantation (HT). Available PGD risk scores are limited to single centres and fail to capture contemporary risk factors. The International Consortium on PGD is an innovative multicenter collaboration comprising of data from 14 centres in the United States, Canada and Europe (Figure 1). This consortium is the largest prospective registry on PGD to-date. The purpose of this study is to derive a novel prognostic algorithm for PGD that incorporates variables relevant to the current HT era using machine learning (ML).


      The International Consortium on PGD includes data on baseline demographics of adult donors and recipients, detailed pre and post-HT hemodynamics, and the type and duration of MCS required before and after HT. The primary outcome includes ISHLT-defined severe PGD within the first 24 hours of HT. Logistic regression will be used to quantify the risk of developing PGD among recipients. An ML algorithm will be used to identify and assess prognostic factors. We will apply resampling methods (i.e. cross validation and bootstrap) to tune the algorithm and assess its performance


      To date, we have collected preliminary data on 2,764 single-organ HT recipients from 2010 - 2020 (75% of the anticipated final sample size) including 106 distinct clinical variables. Based on preliminary results, 215 (7.8%) recipients met criteria for severe PGD. Based on our previous analyses, clinically relevant predictors of outcome will also include recipient and donor age, history of diabetes mellitus and preoperative use of durable LVAD in addition to those identified by the ML algorithm. Our prediction model performance will be assessed in terms of the discriminatory power and calibration of the final model. The model will be derived on 80% and validated on the remaining 20% of the cohort.