Using machine learning for predicting severe postoperative complications after cardiac surgery

Citation

Lapp L, Young D, Kavanagh K, Bouamrane M & Schraag S (2018) Using machine learning for predicting severe postoperative complications after cardiac surgery. Journal of Cardiothoracic and Vascular Anesthesia, 32, pp. S84-S85. https://www.sciencedirect.com/science/article/abs/pii/S1053077018307924; https://doi.org/10.1053/j.jvca.2018.08.156

Abstract
Introduction: Complications after cardiac surgery are becoming more prevalent, being a major contributing factor in patients’ quality of life after surgery, delayed discharges, and resource usage, particularly due to aging population, all of which can be improved by optimising perioperative medicine patient care pathways, using data analytics and predictive modelling. Method: All patients in Golden Jubilee National Hospital undergoing cardiac surgery between 1 st April 2012 and 31st March 2016, reported in clinical audit dataset CaTHI, were analysed. The database included preoperative variables and outcomes such as death, hospital length of stay, and if patient had complications. Three machine learning methods were explored in order to develop a predictive model for severe postoperative complica- tions after coronary artery bypass graft (CABG) and/or valve surgery. These methods were random forest (RF), naïve Bayes (NB) and support vector machines (SVM). The performance was assessed using receiver operating characteristic (ROC) curves, and was compared with our logistic regression (LR) model developed in a previous study . Results: Out of 3700 records, 59.65% of the patients had a CABG surgery, 26.49% had valve surgery, and 13.86% had a combined CABG and valve surgery. Of all admissions, 4.95% had severe postoperative complications. Twenty-five preoperative variables were analysed to find their association with severe postoperative complications, including patient characteristics, comorbidities and information about surgery. RF had the best discriminative ability (AUC ¼ 0.717), compared to LR, NB and SVM (Graph 1). LR and SVM had by far the highest sensitivity (86.9% and 82.0%, respec- tively), compared to RF and NB. RF the highest specificity (77.2%), followed by NB, LR and SVM. Discussion: Overall, RF model has the highest discriminative ability compared to other models. A predictive model could help identifying patients who will potentially have severe complications. This study is a starting point for developing a decision support tool for use in clinical practice.

Keywords
Predictive modelling, Machine Learning, Cardiac surgery

Journal
Journal of Cardiothoracic and Vascular Anesthesia: Volume 32

People

Professor Matt-Mouley Bouamrane

Professor in Health/Social Informatics, Computing Science