Risk prediction models of primary graft dysfunction in cardiac transplant patients: a need to improve?

Cardiac transplant remains the gold standard for end-stage heart failure. Over 5,000 patients now undergo transplants each year 1,2 . The leading cause of 30-day mortality after transplant is primary graft dysfunction (PGD). The incidence of PGD is difficult to determine due to a lack of a standardized definition of the disease before 2014, but some studies report up to 30% 3 . A metaanalysis found a pooled incidence of PGD between 1.6-7.7% 4 . The prevalence of PGD has increased, as reported in a 2024 update from the International Consortium on PGD 5,6 . PGD is now defined as ventricular dysfunction of a donor graft that fails to provide hemodynamic stability within 24 hours post-transplantation that cannot be attributed to any other cause 3,7 . Such secondary causes may include graft dysfunction due to pulmonary hypertension, intraoperative complications, or hyperacute rejection. It can be separated into PGD-LV, for disease affecting the LV or biventricular failure, and PGD-RV, when due to isolated RV involvement 7 . Criteria for PGD-LV are need for VAD or VA-ECO, an EF >45%, or high-dose ionotropic support in the immediate transplant period. An EF of less than 40% is considered to be suggestive of PGD in the absence of other causes 7 .The gold standard for the prevention of PGD is a cold flush of preservation fluid for myocardial protection, which helps improve tolerance to ischemic time. The 2014 consensus statement by Kobashigawa et al., supports the use of inotropes such as phosphodiesterase inhibitors and catecholamines for initial management 7 . Therapy can be escalated to use intra-aortic balloon pumps followed by initiation of mechanical circulatory support (MCS) and ultimately extracorporeal membranous oxygenation (ECMO) 7 . Plasmapheresis may be used to combat inflammatory cytokines that are thought to underly PGD 7,8 . This article examines risk factors and linear and machine learning models used to predict outcomes of PGD.Common risk factors for PGD have been grouped into procedural factors, recipient factors, and donor factors 9 .Age has been identified as a significant risk factor, which may be due to decreased tolerance for long ischemic times in the hearts of older patients 10 . Singh et al. identified an odds risk of 20% for each decade increase in age. Another risk factor of PGD is the cause of donor death. PGD rates were increased in patients who died of intracranial hemorrhage compared to traumatic death, which may be attributed to a catecholamine surge decreasing myocardial function 11 .Gender mismatch between donor and recipient may be a risk factor for PGD and had a worse survival at five years, which persisted despite size-matching organs. Left ventricular hypertrophy of donor hearts should be kept under 14mm and without associated EKG changes 12 . Hearts with LVH are more sensitive to ischemic changes. However, these results are conflicting, with some sites having positive results while others reported worse survival and incidences of graft failure [13][14][15][16] . Lastly, donor ionotropic requirements are shown to induce LV and RV dysfunction and are risk factors for PGD 17,18 .The need for pre-operative MCS is strongly linked to the development of PGD. This is thought to be due to the activation of inflammatory mediators, causing vasodilation and lowering of systemic vascular resistance 8,17 . Increased ischemic time may underly the increased incidence of PGD in MCS patients 19 . This may be due to a summative effect of blood exposure to the surfaces of bypass machines which can further exacerbate an inflammatory response.Pre-operative recipient amiodarone is commonly used in advanced heart failure patients for the treatment of arrhythmias and may be an indicator of the critical, pro-arrhythmic state of these patients. Dose and duration-dependent relationships between amiodarone use and PGD posttransplant have been identified 20,21 . However, early studies have also found that patients receiving pre-operative amiodarone had lower post-operative heart fates that were more likely to require atrial pacing without an increase in post-operative mortality 22 .Other notable recipient factors affecting PGD were recipient diabetes mellitus 23 , recipient age, and recipient re-sternotomy. Advanced glycation of end products and coronary endothelial inflammation may induce graft loss. Diabetes has been identified as a predictor of graft loss within and after the first year of transplant. Advanced recipient age is associated with PGD and mortality, likely due to increased comorbidities and increased rates of fatal infection 17,23 . Prior recipient sternotomy from congenital surgery, CAD implantation, or CABG presents a challenging dissection during transplantation and thus may increase ischemic times or increase the risk for reoperation and bleeding. It has been linked to a three-fold increase in risk of PGD 24 .Procedural factors Prolonged ischemic time during transport and surgery increases the risk of PGD 23,25,26 . Warm ischemic time refers to surgical periods or aortic clamping where blood flow is halted, while cold ischemic time refers to time spent in cold storage. Cardiopulmonary bypass time is linked to PGD due to the occurrence of ischemic reperfusion injury as well as systemic inflammatory pathway activation 27 . Lastly, heart size discrepancies predict mortality at 30 days and one year, likely due to insufficient cardiac index to support body habitus 28 .Incidence of biventricular PGD is known to be higher in DCD recipients as compared to Donation after Brain Death (DBD) recipients. Interestingly, DCD recipients with severe PGD required fewer days of mechanical support than DBD recipients suggesting that different types of donors produce different effects on graft function and recovery 29 .This paper is a qualitative review. Google Scholar and PubMed were used to search for relevant literature. Keywords used were primary graft failure, primary graft dysfunction, cardiac transplantation, and cardiac allograft recipients. All papers published on risk prediction in primary graft dysfunction after cardiac transplantation were reviewed and key information was synthesized into the results of this review. Keywords included PGD, primary graft dysfunction, risk prediction, linear/AI models, risk factors, and cardiac transplant.Six risk prediction models consisting of three linear and three machine learning-driven models were identified in the literature.Linear Models Three popular linear prediction models are RADIAL, PREDICT, and ABCE 23,30,31 (table 1A) Segovia et al., developed the first linear model in 2011 using a Spanish cohort of transplant patients to help establish a definition for PGD as well as a predictive score 23 . They found six multivariate risk factors of PGD: Right atrial pressure  10 mm Hg, recipient Age  60 years, Diabetes mellitus, Inotrope dependence, donor Age  30 years, and Length of ischemic time >240 minutes. The c-statistic between the actual and predicted PGD incidence was 0.74, demonstrating reasonable predictive ability.The PREDICTA prediction score was developed in 2019 using data from UK heart centers over three years and was compared to the RADIAL score 31 . The c-statistic was 0.704 compared to 0.547 from the RADIAL score in the validation cohort. The risk factors identified in this cohort also included diabetes and increasing donor age. Unlike the RADIAL score, they also identified preoperative MCS, prolonged cardiopulmonary bypass time, and prolonged implant time. The incidence of PGD was 38%, similar to their prior finding of 36%.Benck et al., developed the ABCE risk score in 2021 which was based on the severity of disease. Different risk factors were identified for mild to moderate PGD versus severe PGD which may suggest different mechanisms of disease 32 . Multinomial modeling was used to identify risk factors for mild/moderate and severe PGD. PGD occurred in 24% of the cohort. Prior cardiac surgery, recipient GDMT (ACEI/ARB/ARNI plus MRA), treatment with amiodarone plus a beta blocker were identified as three recipient risk factors. Three surgical factors were associated with severe PGD: prolonged ischemic time, increased RBC transfusions, and increased platelet transfusions. A machine learning model exclusively for severe PGD was developed because mild/moderate PGD was not associated with an increased risk of mortality. The ABCE model used four variables available before surgery: treatment with amiodarone plus beta blocker, previous cardiac surgery, and ischemic time. This model has a c-statistic of 0.77 compared to 0.41 when compared to the RADIAL score for severe PGD. Only 48% of patients with severe PGD survived one year, while mild/moderate PGD did not affect survival.Two models using machine learning algorithms have been developed since 2021 (table 1B). Linse et al. developed a non-linear artificial neural networks model to evaluate donor-recipient variables for PGD using a cohort of 64,964 patients using the ISHLT registry 33 . The incidence of PGD was 3.7%. Thirty-three of 77 risk variables were identified as relevant. The model had a cscore of 0.70 (95% CI: 0.68-0.72) compared with the RADIAL score which had a c-statistic of 0.53 (CI 0.52, 0.54). The most influential variables were underlying heart failure diagnosis, ischemia time, and sex mismatch, which were not among the ISHTA 34 . Renal function had a lower influence. 90% of the variables had missing data with a mean of 42%.In 2022, an international, multicenter PGD Consortium was formed to develop a modern clinical AI-driven risk prediction model. The model underwent validation in 6 centers. This risk-scoring algorithm included 18 variables with a c-stat of 0.729 [95%CI 0.695-0.762] 35 . Newer efforts to modify and validate a new algorithm using a tree-based ML algorithm to assess the risk of severe PGD are underway with addition of 6 centers with over 1,000 heart transplants to the cohort. The preliminary PGD-AI calculator was tested on a Toronto population and compared to the RADIAL score. It demonstrated an AUC of 72.9 [59. 5-86.4]. The RADIAL score had an AUC of 0.56 [45.3-0.67.5] 36 suggestive of better discriminatory power.PGD is affected by numerous variables and remains poorly understood especially under the changing patterns of organ procurement and the more recent use of DCD hearts. Therefore, there is a critical need for high-performing risk prediction models. More recent models have begun to use machine learning algorithms instead of traditional, linear statistical methods. Classical models, such as the RADIAL, PREDICTA, and ABCE scores, may be unable to account for complex relationships between risk factors and between risk factors and outcomes 23,30,31 . Machine learning approaches allow for better fit between variables and outcomes through identifying patterns in datasets, especially in large, complete datasets. The machine learning models consistently outperformed linear prediction models. ML algorithms operate as black boxes hence require a collectively improved comprehension of their working systems.The majority of risk prediction models focus on survival after a heart transplant. There are only a few models predicting PGD and those are derived from small cohorts with little external validation. Risk prediction models have the potential to identify candidates for heart transplants earlier and promote modification of key risk factors preoperatively. Ultimately, improved current models for risk prediction are necessary before they can be implemented as a tool to assist with clinical decision-making.There is a need for high-powered predictive models that can integrate an increased number of variables from multi-institutional data improving heterogeneity in patient characteristics that affect patient outcomes in cardiac transplant patients. The main barrier to improved risk prediction is the need for complete and accurate data from representative populations. As cardiac transplant is not as common a procedure as other cardiac surgeries, it is imperative that granular data is entered accurately to reduce gaps in data continuity to improve quality. Missing data is a limitation, as it allows for selection bias and difficulty in quantifying variables with missing data. Due to the use of multicenter data, different protocols are used which may influence post-operative management and may be difficult to include in models.PGD has historically been difficult to define and future studies should use a standardized definition moving forward. The ISHLT consensus developed a standardized definition for PGD in 2014 7 . Variables identified in models developed before 2014 such as the RADIAL score may be less accurate due to ambiguities in the definition of PGD. Interestingly, Linse et al., noted that the use of pre-processed data may have affected the performance of their model, as deep networks operate best with raw data where the algorithm can identify its own relationships between the data 32 . This practice may be a better strategy in future studies.PGD remains the leading cause of early mortality and morbidity following cardiac transplant. Its risk factors are multifactorial and require improved prediction models to best match donors and recipients to improve outcomes for cardiac transplants. Limitations of current models are missing data, uneven distribution of variables in datasets, and the use of processed data. Improvement in the discriminatory ability is necessary before current models can be used to assist in clinical decision-making. The development of improved prediction tools may allow for earlier prediction which can improve prevention and treatment of PGD.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.