Establishment of a gradient boosting prediction model for respiratory failure after non-cardiothoracic surgery based on intraoperative indicators

Abstract

Objective To develop and validate a machine learning prediction model for postoperative respiratory failure (PRF) in patients after non-cardiothoracic surgery based on intraoperative indicators. Methods A total of 705 patients undergoing non-cardiothoracic surgery in our hospital from January 2014 to June 2019 were enrolled, and then 565 patients of them were assigned in the training set (including 128 cases of PRF) and 140 patients into the test set (35 cases of PRF). Another 164 patients undergoing non-cardiothoracic surgery at West China Hospital from May 2019 to January 2020 and Zhongshan Hospital from June 2019 to December 2019 were assigned into the validation set (41 cases of PRF). Nineteen intraoperative indicators were extracted, and 6 machine learning algorithms, such as gradient boosting model (GBM), generalize linear model (GLM), k-nearest neighbor (KNN), naive bayes (NB), neural network (NNET), and support vector machine linear (SVM) were used to develop and test the models and were verified in the validation set. The best model was screened out by comparing the performance of each model, and finally, the web page prediction model was established. Results GBM obtained the best performance, with an accuracy of 76.2%(95%CI: 69.0%~82.5%), an area under the subject curve (AUC) of 0.794 (95%CI: 0.707~0.882), an area under the precision-recall curve (AUPRC) of 0.641, and a Brier score of 0.169. Conclusion The model developed based on GBM algorithm is of higher generalization, accuracy, and clinical utility, and helps avoid overfitting. The developed web page prediction model (http://150.158.55.139) can provide a new dynamic evaluation method for PRF and quantify surgical risk.

Keywords

• postoperative respiratory failure
• prediction model
• clinical utility
• quantification of surgical risk