Dados do Trabalho

Título

PREDICTIVE MODEL FOR ACUTE KIDNEY ALLOGRAFT REJECTION: A MACHINE LEARNING ANALYSIS

Introdução

Kidney transplantation is the best treatment for patients in final stage of Chronic Kidney Disease as it prolongs the survival of individuals and reduces costs to the public health service. However, there are few studies related to graft survival and rejection. The objective was to create a predictive model for 30 days post-transplant rejection using machine learning techniques, elucidating variables with greater predictive potential.

Material e Método

Retrospective study with 1255 patients transplanted from living and deceased donors at a reference public hospital in Brazil between 01/13/2010 and 09/15/2020. Recipient, donor, transplantation and postoperative period data were collected from physical and electronic records. We randomly split the data into derivation (training - 80%) and validation (test - 20%) datasets. Five supervised ML algorithms were developed with this subset of variables in the training set: Simple Logistic Regression, Lasso, Multilayer Perceptron, XGBoost, Light GBM.

Resultados

There were 147 (12.48%) cases of graft rejection within 30 days after kidney transplantation. The best model was XGBoost (Accuracy: 0.839; ROC AUC: 0.715, Precision: 0.900). The model showed that deceased donor transplantation, glomerulopathy as underlying disease and donor’s use of vasoactive drugs had more than 20% importance as rejection risk factors. The variables with the greatest predictive values were Thymoglobulin Induction and Delayed Graft Function.

Discussão e Conclusões

Immunosuppression induction by Thymoglobulin acts as the main protective factor against acute renal rejection and Delayed Graft Function (DGF) acts as the main outcome inductor factor. The increased use of expanded criteria donors, a greater ratio between donor and recipient age and higher final values of donor creatinine operate as important DGF-inducing factors, corroborating with the hypothesis of inadequate donor hemodynamic maintenance as one of the factors related to the high incidence of DGF and the use of vasoactive drugs. We observed higher rates of DGF in Brazil compared to international cohorts, which could be due to a better donor care offered to patients in foreign transplant services. We fitted a machine learning model to predict 30 days graft rejection after kidney transplantation that reaches a higher accuracy and precision. It is concludable that Machine Learning models could contribute to predicting kidney survival using non-traditional approaches.

Palavras Chave

Kidney Transplantation; Graft Rejection; Machine Learning; Chronic Renal Insufficiency; Allografts