Thesis

Development of prediction models for postoperative complications following cardiac surgery

Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2022
Thesis identifier
  • T16481
Person Identifier (Local)
  • 201787114
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • While postoperative mortality in cardiac surgery has reduced in the past twenty years, due to changes in patient population undergoing open-heart surgery, postoperative complications are becoming more common. With the development of perioperative medicine, data-driven perioperative risk prediction models are now an integral component for decision-making about the type of treatment that is most suitable for the patient, for communicating risk of surgery, and for auditing purposes. However, the currently developed prediction models focus on mortality, rather than postoperative complications. In this thesis, the problem of postoperative complications in cardiac surgery is investigated by analysing cardiac patient data in Golden Jubilee National Hospital to predict (1) severe postoperative complications, (2) acute kidney injury and (3) delirium. Furthermore, cardiac anaesthetists and surgeons were involved in explorative interviews about current challenges in cardiac surgery, and a study to define and classify postoperative complications in cardiac surgery. Patients undergoing coronary artery bypass graft (CABG), valve and combined CABG and valve surgeries in Golden Jubilee National Hospital between 1st April 2012 and 31st December 2018 were analysed. The prevalence of severe complications, acute kidney injury and delirium for this patient population was 5.91%, 18.93% and 12.47%, respectively. Two types of models were developed: (1) preoperative models using data that was available before surgery; and (2) hourly prediction models that used both preoperative data and laboratory results recorded in the intensive care unit. Out of all preoperative modelling experiments (1), random forest predicting severe postoperative complications had the highest performance, with the area under the receiver operating characteristic curve (AUC) of 0.713, sensitivity of 0.562 and specificity of 0.748. When predicting the onset of acute kidney injury on an hourly basis in intensive care (2), BARTm achieved the highest mean AUC of 0.850 with sensitivity of 0.821 and specificity of 0.741. For hourly delirium prediction (3), support vector machine achieved the highest mean AUC of 0.941, sensitivity of 0.907 and specificity of 0.870. This thesis shows that using routinely collected medical data can be used to develop both preoperative and hourly ICU predictive models for postoperative complications, such as acute kidney injury and delirium. Such prediction models could help with clinical decision making, communication about risk, research in complications and auditing.
Advisor / supervisor
  • Schraag, Stefan
  • Young, David
  • Roper, Marc
  • Kavanagh, Kimberley
  • Bouamrane, Matt-Mouley
Resource Type
DOI
Funder

Relations

Items