IDDF2024-ABS-0009 Evaluation of the predictive model for acute bowel injury in patients following cardiothoracic surgery using the random forest algorithm

<h3>Background</h3> The present investigation involved the analysis of clinical data from heart failure patients obtained from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database. By utilizing the random number table method, a training set was formulated with a 2:1 ratio for subsequent testing. The random forest algorithm was employed to assess the significance of potential predictive indicators for ABI following cardiothoracic surgery. Internal validation was conducted to evaluate the predictive capacity of the model for postoperative ABI occurrence in this patient population. <h3>Methods</h3> Among the 903 patients included in the study, 163 patients were classified in the occurrence group (experiencing ABI), while 740 patients belonged to the non-occurring group (not experiencing ABI). The importance ranking of the 13 indicators was determined by assessing the decrease in average accuracy. Through the utilization of out-of-bag data error, Akaike information criterion, and Bayesian information criterion, a total of six variables were identified as significant and included in the predictive model (P&lt;0.01). These variables, namely postoperative NGAL, postoperative TIMP 2, postoperative IGFBP7, and postoperative pCr, underwent scrupulous screening procedures. The efficacy of the prediction models was evident through the application of the MDS method. <h3>Results</h3> Among the 903 patients included in the study, 163 patients were classified in the occurrence group (experiencing ABI), while 740 patients belonged to the non-occurring group (not experiencing ABI). The importance ranking of the 13 indicators was determined by assessing the decrease in average accuracy. Through the utilization of out-of-bag data error, Akaike information criterion, and Bayesian information criterion, a total of six variables were identified as significant and included in the predictive model (P&lt;0.01). These variables, namely postoperative NGAL, postoperative TIMP 2, postoperative IGFBP7, and postoperative pCr, underwent scrupulous screening procedures. The efficacy of the prediction models was evident through the application of the MDS method. <h3>Conclusions</h3> The utilization of the random forest algorithm prediction model that incorporates postoperative NGAL, postoperative TIMP 2, postoperative IGFBP7, and postoperative pCr enables accurate prediction of ABI occurrence in patients following cardiothoracic surgery.