In this study, we developed and validated a multivariate preoperative risk prediction model for the risk of prolonged ICU stay. The prediction model uses 12 readily available preoperative patient and disease characteristics assigning an independent weight to each to provide quantitative information about risk. The prediction model demonstrated relatively strong discriminatory ability (area under ROC curve = 0.72), with no significant deviation from perfect fit. This prediction model, which makes use of routinely available preoperative data, can serve the clinician and the patient by providing a simple method to assess accurately the risk of prolonged ICU stay following CABG surgery. Also, by accounting for patient variability, the model may provide an estimate for required resources in CABG surgery and help efforts to control costs and avoid bed shortages in ICU.
Because of the ever-present shortage of beds within hospitals, especially within ICU, considerable effort needs to be expended in resource planning and allocation. One solution to increase efficiency in the field of cardiac surgical ICUs is to plan the operations to use available resources in an optimal fashion . Prediction of postoperative ICU stay would facilitate decisions to allocate resources and to plan weekly schedules for CABG operations. When ICU bed availability is an issue, patients with high risk of staying for a prolonged stay could be electively scheduled for surgery in a series rather than parallel. While scheduling the operative list, a case mix of patients that includes patients with probability of prolonged ICU stay and those who are likely to have uncomplicated recovery could potentially avoid the possibility of blocking beds in ICU.
Existing risk stratification models for mortality following CABG have been shown to be good predictors of prolonged ICU stay. Lawrence and co-workers concluded that the Parsonnet score was a good predictor of ICU stay <24 hours, which could help cardiothoracic units when resources are limited to a few ICU beds . Nilsson and associates also found the EuroSCORE to be a useful predictor of ICU stays greater than two days in open heart surgery . However, our study found that, compared to a specifically designed prediction model for prolonged ICU stay, the Parsonnet, and both versions of the EuroSCORE were not reliable predictors, with a tendency to under-predict.
Several studies have identified risk factors for prolonged ICU stay with varying definitions. Wong and colleagues examined 885 CABG patients and defined prolonged ICU stay as greater than 48 hours. Also, unlike our study, they examined post-operative factors. The risk factors identified included increased age, female sex, pre-operative myocardial infarction, post-operative use of intra-aortic balloon pump, inotropes, bleeding, atrial arrhythmia and renal insufficiency . Michalopoulos and co-workers used the same definition as Wong, and included perioperative factors such as blood use and inotrope support in their final logistic regression model, with only age and ejection fraction identified as preoperative risk factors . Other postoperative factors identified as predictors of prolonged ICU stay have included elevated Troponin- T levels  and pulmonary artery blood temperature greater than 36.4 degrees C on admission to the ICU . Inclusion of peri- or post-operative factors in our study, however, would have limited the usefulness of the prediction model in aiding resource management prior to surgery.
Christakis and colleagues  analysed preoperative risk factors for prolonged ICU stay in 889 consecutive patients undergoing isolated CABG between 1990 and 1992. Using the same definition as in our study, 6.8% of patients had stays in ICU of greater than 3 days. Only two preoperative risk factors however could be identified, with both recent myocardial infarction and current smoking increasing the risk of prolonged ICU stay. Our study also found an association between smoking and prolonged intensive care, however, a history of, or recent, myocardial infarction was not identified as a risk factor.
More recently, Janssen and associates  published a preoperative prediction model for prolonged ICU stay defined as greater than 3 days. The analysis included 888 contemporary CABG patients, of which 104 stayed in ICU for more than 3 days. They presented a logistic regression equation to predict prolonged ICU stay, which included the following variables: lung disease, no-sinus rhythm, no mild valve pathology, prior surgery, priority, and on-pump surgery. These factors are quite different from those identified in our study, except for priority and on-pump surgery. A reason for this might be due to the sample size, with only 104 outcomes compared to the 457 prolonged ICU stay patients in our analysis. Inclusion of on-pump surgery as a preoperative factor is based on the fact that in most circumstances, the use of CPB is a preoperative planned approach and not necessarily an intra-operative decision, except in a minority of cases. There may be some concern that on-pump surgery would be identified as a risk factor purely due to selection bias with off-pump surgery being performed in lower-risk cases. However, Bucerius and colleagues concluded, in an analysis which included over 700 off-pump CABG, that avoiding cardiopulmonary bypass could optimize patient outcome with respect to prolonged ICU stay .
There are limitations to this study which need to be considered. One such limitation is that it is based on data from one institution, and therefore, subject to the efforts of local practices and case mix. Although we have validated the prediction model on external data between April 2003 and March 2004, this model still requires validation from other institutions. It is also unclear how useful this tool might actually be in aiding resource management, compared to a clinicians own estimates of risk for an individual patient. It has been shown though that clinicians tend to overestimate the probability of mortality and prolonged ICU stay . The application of this model to other cardiac surgery procedures is also needed as these other non-CABG procedures could have a significant impact on ICU bed availability.
In conclusion, clinicians may use the prediction model contained within this paper to aid in resource management within the ICU. The logistic version of the model can be easily programmed into appropriate software resident on desktops and hand-held computers. Alternatively, the clinical risk assessment tool could be used and provided on small pocket-sized laminated cards providing simple and easy approximations of the risk of prolonged ICU.