Background: Primary graft dysfunction (PGD) is the leading cause of early mortality after a heart transplantation (HT). Our current incidence rate of severe PGD is 15% per year, which is twice the international average. The International Consortium on PGD was established to identify contemporary risk factors for PGD and as a part of this multicenter collaboration, we trained a multicenter machine learning (ML) algorithm to predict the risk of severe PGD. In this study, we aimed to validate the AI-PGD calculator in our local contemporary population and evaluate its potential as a tool for improving outcomes following HT.
METHODS AND RESULTS: We conducted a retrospective review using electronic health records of all adult HT recipients at the Toronto General Hospital from 2020-2022 and collected the 18 variables required for AI-PGD risk score. Only patients with all 18 variables and single organ HT were included. We evaluated the performance of the AI-PGD risk score using several metrics, including the area under the receiver operating characteristic (AUCROC) curve, discriminatory power and calibration.
A total of 84 HT recipient and donor matches were included in our study as outlined in Table 1. Among recipients, male sex, the presence of durable LVAD and longer ischemic time were associated with a significantly higher risk of severe PGD. The AI-PGD model demonstrated an AUC-ROC of 77.1% [95% CI, 63.9%-90.2%] (Figure 1). The model calibration was excellent with a predicted severe PGD rate of 13.6 ± 0.08% and an observed rate of 14%. When the threshold of 8% was chosen for severe PGD, the overall sensitivity was 100% and the specificity was 18%.
Conclusion: The AI-PGD model has a superior performance in predicting severe PGD after HT in our local cohort and meets the threshold for clinical use. The use of this model in clinical practice may lead to more informed decision making and improved HT outcomes.
