Objectives: Acute kidney injury after liver transplant results from several interconnected factors related to graft, recipient, intraoperative, and postoperative events. The random decision forest model enables an appreciation of each factor's contribution, which may be helpful in setting up a preventive strategy. This study aimed to evaluate the importance of covariates at different times (pretransplant, end of surgery, postoperative day 7) with a random forest permutation algorithm.
Materials and methods: We used a retrospective singlecenter cohort of patients, without preoperative renal failure, who underwent primary liver transplants from deceased donors (N =1104). Significant covariates for stage 2-3 acute kidney injurywere included in a random forest model, and features importance was evaluated with mean decrease accuracy and Gini index.
Results: Stage 2-3 acute kidney injury occurred in 200 patients (18.1%) and was associated with lower patient survival, even after exclusion of early graftloss. At univariate analysis, recipient factors (serum creatinine level, Model for End-Stage Liver Disease score, body weight, body mass index), graft variables (weight, macrosteatosis), intraoperative factors (number of red blood cells, duration of surgery, cold ischemia time), and postoperative event (graft dysfunction) were associated with kidney failure. The pretransplant model found that macrosteatosis and graft weight contributed to acute kidney injury. The postoperative model indicated that graft dysfunction and the number of intraoperative packed red blood cells were ranked as the 2 most essential factors in posttransplant renal failure.
Conclusions: The application of a random forestfeature identified graft dysfunction, even transient and reversible, and the number of intraoperative packed red blood cells as the 2 most crucial contributors to acute kidney injury,thus indicating that prevention of graft dysfunction and bleeding are key points to limit the risk of renal failure after liver transplant.