ESCORE DE RISCO DE INJÚRIA RENAL AGUDA APÓS TRANSPLANTE HEPÁTICO POR APLICAÇÃO DE REDE NEURAL ARTIFICIAL

Abstract

Introduction: The multifactorial origin of acute kidney injury (AKI) after liver transplantation (LT) makes it complex to predict which candidate for the procedure presents an increased risk of this complication, but the significant impact of AKI on the prognosis of these patients highlights the need of the construction of a effective prediction model applicable in clinical practice for the occurrence of this complication. Objective: To develop a new risk score for the onset of acute AKI after LT by application of an artificial neural network model. Methodology: Data were collected from one 145 patients submitted to decesead donor LT, including demographic data and comorbidities of the recipient, clinical characteristics of the donor and graft, intraoperative information and laboratory tests. The primary outcome was postoperative AKI according to the International Club of Ascites criteria. By logistic regression, the predictors were identified and inputed in the artificial neural network algorithm, then accuracy of the artificial neural network and logistic regression models were tested. A scoring system based on the β coefficient values of the predictors was conducted, and a final prognostic score was determined and categorized into risk groups in the artificial neural network model. Results: The incidence of AKI was 60.6% (n = 88 / 145) and the following predictors of AKI onset were identified by logistic regression: MELD score ≥ 25, previous kidney dysfunction, grafts from extended donors criteria, intraoperative arterial hypotension , massive blood transfusion and levels of serum lactate ≥ 2 mmol/l at the end of surgery. These six independent variables were incorporated in the artificial neural network model, and the AUROC was best for artificial neural network (0.81) than for logistic regression model (0.71). There was satisfactory agreement in the artificial neural network model between predictions and actual AKI onset events (HLχ2 of 5.57, p = 0.612). The six predictors received weighted points for the risk score construction, and according to the artificial neural network model the cutoff values for AKI risk stratification were: 0 to 6 (low), 7 to 15 (moderate) and 16 to 22 (high), with significant difference of AKI incidence between the risk groups. Conclusions: The present new score by artificial neural network application is a new instrument for identification of recipients at risk of post-LT AKI, and this score is readily available at the end of the surgery, and would be a decision tool for prophylactic or early therapeutic procedures for postoperative AKI.