Comparison of artificial neural network, extreme gradient boosting and logistic regression for predicting intraoperative transfusion in repeat cesarean delivery

Abstract

Objective To construct prediction models for intraoperative transfusion in repeat cesarean delivery using logistic regression, extreme gradient boosting (XGB) and artificial neural network (ANN) based on the clinical data with a large sample size and compare their prediction performance. Methods From October, 2015 to October, 2017, a total of 2 525 women underwent repeat cesarean delivery in our hospital. We analyzed the preoperative and intraoperative data of the women, and the variables with possible clinical significance were selected to construct the prediction models for intraoperative transfusion using logistic regression, XGB and ANN. The area under the receiver operating characteristic curve (AUROC), precision, recall (sensitivity) and F1 score of the 3 models were compared. Results Among the 2 525 women included in this study, 332 (13.1%) received intraoperative transfusion in repeat cesarean delivery. The final model was constructed based on 5 important predictors, including preoperative hemoglobin, operation time, uterine atony, placenta previa and ASA classification. The AUROC of the algorithms of logistic regression, XGB and ANN were 0.958, 0.959, and 0.956, respectively, and their F1, precision and recall did not differ significantly. To compare their prediction performance, further predictive verification was carried out on training samples and test samples, for which XGB yielded AUROC of 0.904 and 0.886, respectively, significantly higher than those of logistic regression (0.868 and 0.878) and ANN (0.882 and 0.884). The precision, recall and F1 of XGB were also better than those of logistic regression and ANN. Conclusion Preoperative hemoglobin, operation time, uterine atony, placenta previa and ASA classification are predictors of intraoperative transfusion in repeat cesarean delivery. Logistic regression, XGB and ANN can all be used for predicting intraoperative transfusion in repeat cesarean delivery, but XGB has higher prediction accuracy than logistic regression and ANN.

Keywords

• repeat cesarean delivery
• intraoperative transfusion
• machine learning