Abstract:
AIM To construct a predictive model of acute kidney injury (AKI) for patients undergoing cardiothoracic surgery based on the random forest algorithm and analyze its predictive value.
METHODS Two hundred and twelve patients in our hospital who underwent cardiothoracic surgery from January 2018 to December 2020 were selected as the research subjects. The random number table method was used to establish the training set and the test was set at a ratio of 2:1. Random forest algorithm was used to rank the importance of predictive indicators of AKI after cardiothoracic surgery. According to the out-of-bag data error, the Akaike information criterion and the Bayesian information criterion, the ranking indexes were screened and the predictive model was constructed. The multidimensional scaling method (MDS) was used to observe the predictive ability of the model after cardiothoracic surgery and internal verification method was adopted to validate the model’s ability to predict postoperative AKI in patients undergoing cardiothoracic surgery.
RESULTS Among the 212 patients, AKI occurred in 64 cases (occurrence group) and did not occur in 148 cases (non-occurring group). The 16 indicators were ranked in importance according to the degree of decline in average accuracy and the degree of decline in average Gini index. Using the out-of-bag data error, Akaike information criterion and Bayesian information criterion, 4 variables were screened out, namely, postoperative NGAL, postoperative TIMP 2, postoperative IGFBP7 and postoperative pCr, and they were included in the model(P<0.01). It was observed that the prediction models were well differentiated by the MDS method.
CONCLUSION The random forest algorithm prediction model based on postoperative NGAL, postoperative TIMP 2, postoperative IGFBP7 and postoperative pCr can be used to predict the occurrence of AKI in patients after cardiothoracic surgery.