Secondary Logo

Journal Logo

Predictions and Outcome

Estimation of morbidity risk factors in intensive care unit: a Bayesian discriminant approach: 028

Scolletta, S.; Giomarelli, P.; Cevenini, G.; Biagioli, B.

Author Information
European Journal of Anaesthesiology: June 2004 - Volume 21 - Issue - p 14

Introduction: Most risk stratification scores are based on unmodified preoperative patient variables. However, there are a number of intra- and postoperative physiologic variables that might influence operative morbidity. Higgins and colleagues evaluated the relative contribution of preoperative condition, operating room events, and physiologic measurements at Intensive Care Unit (ICU) admission to outcome [1]. Our goals were: 1) to evaluate the validity of Cleveland Clinic Scores in our patients population; and 2) to develop a Bayesian risk model able to predict the independent risk factors of morbidity after coronary artery bypass graft (CABG).

Method: We selected over 80 perioperative variables that could potentially be associated with postoperative morbidity, and analysed the influence of each risk factor on outcome at ICU discharge. 1,090 consecutive adult patients underwent CABG were analysed. A training data set (740 patients, who had surgery between January 2000 and December 2001) was used for designing a Bayesian probability model. A testing data set (350 patients, operated on between January and December 2002) was employed for evaluating the prognostic power of the Bayesian approach and comparing it with Cleveland Clinic Scores. The classification performances of both Cleveland Clinic Scores and Bayesian model were assessed by ROC curve analysis.

Results: The in-ICU morbidity rate was 20.5% (223/1090). The Cleveland Clinic Scores had a good discriminant power, and the comparison of two ROC areas gave no statistical significance (P < 0.48). At Bayesian discriminant analysis, stepwise feature selection procedure stopped with 12 variables able to discriminate morbidity: 1) postoperative decreased oxygen delivery and 2) need for inotropic support after surgery were the most significant predictive variables. The Bayesian model showed an excellent ability to recognize morbidity risk (sensitivity and specificity = 72%), and the comparison of areas enclosed by ROC curves between Bayesian and Cleveland Clinic Scores showed high statistical significance (P < 0.001).

Conclusions: Most prediction rules use sequential numerical risk scoring to quantify the patient's prognosis, and when appropriately used may represent an advanced form of audit. Cleveland Clinic Score seemed to be a suitable tool for predicting morbidity in our CABG patients. The Bayesian approach identified the major clinical predictors of morbidity and was a better-designed risk prediction score for discriminating patient's subsequent progress.

Reference:

1 Higgins TL, Estafanous FG, Loop FD, et al. ICU admission score for predicting morbidity and mortality risk after coronary artery bypass grafting. Ann Thorac Surg 1997; 64: 1050-1080.
© 2004 European Society of Anaesthesiology