Machine learning enhanced prediction of deep sternal wound infection after surgical myocardial revascularization
Abstract
Background
Deep sternal wound infection following surgical myocardial revascularization is a potentially devastating complication. This study aims to improve risk prediction of deep sternal wound infection with machine learning algorithms.
Methods
This single-center retrospective study contains data from 5221 consecutive patients who underwent surgical myocardial revascularization between 2007 and 2022. Two machine learning algorithms (Extreme Gradient Boosting and Deep Neural Network) were trained with perioperative parameters and validated to detect deep sternal wound infection. Their predictive accuracy was then compared to conventional statistical modelling in terms of multivariable logistic regression. Shapley Additive Explanations was applied to the Extreme Gradient Boosting model to determine the importance of each contributing feature to the occurrence of deep sternal wound infection.
Results
The overall incidence of deep sternal wound infection was 3.4 % and 54.7 % occurred within 15 days after surgery. The predictive accuracy of the applied machine learning models was identical (AUC: 0.851, p = 0.982) whereas both, Extreme Gradient Boosting (AUC: 0.851, p = 0.031) and Deep Neural Network (AUC: 0.851, p = 0.017), outperformed the multivariable logistic regression model (AUC: 0.796). According to the Shapley Additive Explanations, the five most important predictive features were body mass index, red blood cell transfusions, pleural effusion requiring pleurocentesis, lower preoperative hemoglobin levels and concomitant peripheral artery disease.