Developing predictive models for postoperative complications in cardiac patients

Student thesis: Master's Thesis

Abstract

A number of cardiac preoperative risk stratification tools have been developed to predict mortality, and less often mortality and morbidity in surgery. Depending on their severity, postoperative complications can have a significant impact on patients' quality of life, hospital length of stay, and healthcare costs and resource usage. Nevertheless, mortality often remains the main 'key performance indicator' used in surgery and is the most commonly reported outcome when evaluating cardiac risk scores.In this thesis, cardiac data in Golden Jubilee National Hospital was analysed to develop predictive models for postoperative complications in patients undergoing coronary artery bypass graft (CABG), valve, and combined valve and CABG surgery.All patients undergoing cardiac surgery, recorded in the Golden Jubilee National Hospital CaTHI database between 1st April, 2012 and 31st March, 2016, were analysed.Three outcomes were investigated: (a) if the patient had postoperative complications, (b) if the patient had severe postoperative complications, and (c) the level of postoperative complications. For each outcome, prediction models were developed, using logistic regression (a, b) and ordinal logistic regression (c). The performance of the models was measured, using receiver operating characteristic (ROC) curves (a, b) and confusion matrices (c), and compared with the performance of the logistic EuroSCORE predicting each outcome.Of 3700 admissions, 59.7% had CABG, 26.4% valve, and 13.9% combined CABG and valve surgery. Overall, 48.65% of the patients had postoperative complications, with the prevalence of mild complications being 7.05%, moderate 36.65%, and severe complications being 4.95%.For the model (a) predicting postoperative complications, the area under the ROC curve (AUC) was 0.636 with the sensitivity of 65.7% and specificity of 54.6%. For the model (b) predicting severe postoperative complications, the AUC of the local model was 0.685, with the sensitivity of 86.9% and specificity of 46.8%. The model (c) predicting the level of postoperative complications resulted with the confusion matrix, where the accuracy for predicting no complications was 58%, mild 92%, moderate 63% and severe complications 95%.Being the most accurate based on AUC, the local model predicting severe complications included eight variables: age, sex, diabetes, left ventricular function, previous cardiac surgery, hypertension, active endocarditis and previous myocardial infarction.The variables associated with severe complications and the local model predicting severe complications could help the clinicians identify which patients are more likely to have severe complications in order to allocate resources accordingly.
Date of Award1 Oct 2017
LanguageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde

Cite this