Prediction of postoperative sepsis mortality risk based on machine learning in patients undergoing abdominal surgery

Objective To explore the feasibility of constructing prediction models of postoperative sepsis mortality risk based on machine learning in patients undergoing abdominal surgery. Methods A case-control trial was designed and conducted on the patients diagnosed with sepsis after abdominal surgery from Medical Information Mart for Intensive Care Ⅳ (MIMIC-Ⅳ) database, and 90-day mortality was defined as the primary endpoint event after hospitalization. The dataset was ramdomly split into training (70%) and test (30%) datasets according to wether diagnosed with postopertive sepsis or not. On the training dataset, logistic regression (LR), gradient boosting decision tree (GBDT), random forest (RF), support vector machine (SVM) and adaptive boosting (AdaBoost) were used to develop the prediction model for death. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, negative predictive value, accuracy and F1 score were used for model evaluation on the test dataset. Results A total of 986 patients were finally analyzed, of whom 251 patients (25.5%) died within 90 d after hospitalization. The AUC values of LR, GBDT, RF, SVM and AdaBoost prediction models were 0.852, 0.903, 0.921, 0.940 and 0.906, respectively. The model based on SVM yielded the best AUC value, higher differentiation and better prediction performance, while LR performed the worst among them. Conclusion The performances of the prediction model of postoperative sepsis mortality based on GBDTT, RF, SVM and AdaBoost are all better than that of traditional LR model, which may help to assist clinical decision making and improve adverse outcomes.