Avaliação da acurácia de modelos preditivos de função tardia do enxerto renal e construção de um novo modelo

Abstract

Introduction: Brazilian studies have been reported incidences of delayed graft function (DGF) varying between 54 and 82%, 2-3 fold the incidences described by American and European cohorts. The available predictive models of FTE, based on distinct clinical realities, are not validated in our population. Objectives: To evaluate the incidence and risk factors for DGF in a cohort of deceased donor (DD) kidney transplants (KT) performed in a Brazilian region where standard criteria donors are predominant. To evaluate the accuracy of the available predictive models and, since none of these models present good accuracy, to construct a local prediction model. Methods: Retrospective cohort study including all DD KT performed between Jan 2014 and Dec 2017 in two transplant centers: Hospital Geral de Fortaleza and Hospital Universitário Walter Cantídeo. We excluded: combined transplants, preemptive transplants, KT using pulsatile machine perfusion, and graft losses or deaths in the first week after KT. The predictive DGF models tested were those described by Irish et al., Jeldres et al., Chapal et al., and Zaza et al. For the construction of the new predictive model, machine learning was used, including non-linear models besides logistic regression, and the three methods with the best performance were used. Results: 443 patients were included. Patients were predominantly men (56.7%), young adults (44.2 ± 14.7 years), mixed race (84.4%), with chronic renal disease of unknown etiology (30.2%), who remained 46.8 ± 45.2 months on dialysis. Donors had a mean age of 31 ± 12.9 years, most of them died from trauma (70.3%), 5.1% were hypertensive (HA), 0.6% were diabetic, and the final creatinine was 1.2 ± 0.6 mg/dL. Only 4.8% were expanded criteria donors. 83.1% of the grafts were perfused with HTK and the mean cold ischemia time (CIT) was 20.9 ± 4 hours. The incidence of DGF in this sample was 53%. In the multivariate logistic regression analysis, the variables independently associated with DGF were antecedents of diabetes (OR 1.922, 95% CI 1.119-3.302, p=0.018), time on dialysis (OR 1.009, 95% CI 1.004-1.014, p<0.001), donor with a history of HA (OR 2.331, 95% CI 1.247-4.355, p=0.008), donor final creatinine (OR 1.947, 95% CI 1.320-2.872, p=0.001) and CIT (OR 1.115, 95% CI 1.058-1.175, p<0.001). The predictive models of DGF available presented regular or poor discriminant power: Irish (AUC 0.686), Chapal (AUC 0.638), Jeldres (AUC 0.613), Zaza (AUC 0.591). 9

The three models with the best performance were decision tree (AUC-ROC = 0,791), neural networks (AUC-ROC = 0,886) and support vector machine (SVM) (AUC-ROC = 0,784). In the decision tree, the variables considered were: donor age, recipient body mass index (BMI), donor blood pressure, time on dialysis, cause of donor death (cerebrovascular), high dose of vasoactive drugs, donor glucose, final donor creatinine, donor diuresis and CIT. In the neural network model: donor age, CIT, time on dialysis, donor BMI, final donor creatinine, cause of donor death (cerebrovascular), recipient BMI, donor diuresis, donor blood pressure and high dose of vasoactive drugs. In the SVM model: CIT, cause of donor death (cerebrovascular), donor age, donor diuresis, donor glucose, high dose of vasoactive drugs, donor blood pressure, recipient BMI, time on dialysis, and donor BMI. The final model showed excellent discriminant power (AUC 0.922). Conclusion: the incidence of DGF in the sample was high, despite the predominance of standard criteria donors. In addition to variables classically associated with DGF, identified by logistic regression, variables related to donor maintenance were pointed out in non-linear statistical methodologies. The available predictive models had poor accuracy in predicting DGF in our population. The developed model presented excellent performance.