In the Framingham Heart Study performed on 6233 subjects (mean age 54 years, 54% females), 8% of women and 8.7% of men exhibited mild renal insufficiency, which was defined according to the serum creatinine value (120 to 256 μmol/l, i.e., 1,4 to 3,0 mg/dl in women and 136 to 265 μmol/l, i.e., 1,5 to 3,0 mg/dl in men) [7]. The data analysis of 18790 patients in the HOT (Hypertension Optimal Treatment) Study revealed that the impaired renal function was a predictor of increased cardiovascular morbidity and mortality and that patients suffering from renal failure exhibited a higher cardiovascular risk than patients with an intact renal function [8]. In patients with chronic heart failure, the renal function is a prognostic risk value [9, 10] which can be regarded as a predictor of mortality in this patient group [11]. Multiple models for preoperative risk evaluation in patients undergoing heart surgery also confirmed the significance of the renal function as a predictor of mortality. In these models, acute renal failure, the necessity of dialysis and serum creatinine, in form of a categorical value, were applied as risk criteria [1].
The serum creatinine level is influenced by many factors which are independent of the glomerular filtration rate: tubular secretion and reabsorption, endogenous production, variable intake, extrarenal elimination and interference, caused by the laboratory diagnostic techniques and medicaments used [12, 13]. Since the assessment of the renal function, based on the determination of serum creatinine, is associated with several limitations [13, 14] and the measurement of creatinine clearance by urine collection is rather time-consuming, several formulas estimating the renal function from serum creatinine, body weight, age and sex, as well as ethnic features, have been developed. All these formulas exhibit certain limitations. The most commonly used equation for estimating creatinine clearance, e.g., in the Medicare programme and in the transplantation waiting lists in the USA [15], is the Cockcroft-Gault formula. Although this formula also does not provide absolutely accurate results (e.g., in elderly patients) and it may over- or underestimate the true renal function [12, 16–19], several studies on cardiac insufficiency and renal impairment have shown a good correlation between the creatinine clearance values calculated according to Cockcroft and Gault and the measured glomerular filtration rate [20–24]. Because of this broad acceptance of the Cockcroft-Gault formula, we have decided to use it in our model.
The initial steps in our present study concentrated on the proper selection of a threshold creatinine clearance value. It was apparent that in all patient groups with a CC value lower than 55 ml/min the observed mortalities were higher than the ones predicted by the standard EuroSCORE model. This indicates a poor modelling of the renal impairment variable in the EuroSCORE model, which is defined by a binary serum creatinine variable, i.e. SC level above 200 μmol/l (Table 4).
The determined CC threshold of 55 ml/min is in accordance with many findings of other authors. The large HOT Study defined a CC value of 60 ml/min as a criterion for the impaired renal function [8]. Hillege and co-workers [11] divided their population of 1906 patients suffering from chronic heart failure into four groups according to the CC values estimated with the Cockcroft-Gault formula. The following intervals of CC values were used: <44, 44–58, 59–76 and >76 ml/min. The overall mortality (calculated according to Kaplan-Meyer) in the four groups studied was 36.5%, 24.8%, 17.6% and 13.7% whereby significant differences were determined between the first two and last two groups.
Because of the differences in the risk profile between the patient population examined in the EuroSCORE study and our own (see Table 1), we have developed several own institutional score models by means of logistic regression analysis. To be able to secure comparability with other institutions and because of the broad acceptance of the EuroSCORE model, we have used its 18 determinants as the basis of our model as well. The predictive power of the score model can be improved by readapting the EuroSCORE regression model to the large patient population in our institution (Table 3). Furthermore, we have shown that the variable selection of the EuroSCORE can be significantly improved by better capturing the major risk factor impaired renal function. We found either the continuous or the binarized preoperative creatinine clearance value is an easy to assess measurement which encoded renal function much better than the binarized serum creatinine value as called for in the EuroSCORE procedure.
We showed this by significance analysis of the binarized CC<55 value with p < 0.001 using Monte-Carlo methods. Furthermore we demonstrated the improvement of the predictive power by calculating the area gain under the ROC curve. The final experiments examined the individual contribution of each EuroSCORE variable to the ROC area. Only one variable ("age") showed a marginal ROC area contribution (0.0291) larger than one standard deviation of the best ROC value (0.0018). All other 17 well established predictors would be doubted when judged by ROC area improvement only. This exhibits the difficulty of the uncertainty measures (s.d.) of the ROC area in comparison to model differences. To circumvent this hurdle we evaluated the rank ordering of the predictors contribution. This can be done in two ways: either the marginal contribution measured by leaving the predictor out of the regression, or by employing only the isolated predictor. The resulting numbers and rankings are certainly varying since in the set of 18 parameter the information is partially provided by other variables.
The rank numbers expose the superiority of the proposed creatinine clearance value compared to the standard EuroSCORE choice. While the serum creatinine variable ranks number 14, the CC replacement would gain rank 5 within the (modified) set of EuroSCORE predictors. The reverse view gives even more favourable figures: the creatinine clearance value (binary or continuous) surpasses age in the top rank as a single variable predictor.
The risk model based estimation of the expected mortalities, the grouping (EM) and comparison with the total observed mortality (OM) by the two parameters, NLS and RAMQ, represent effective analytical tools in assessing the potential further influences for mortality (occurrence of preoperative disease, choice of surgical procedure, etc.). Using Monte-Carlo methods for testing the significance of deviations as well as ranking of predictor variable are a valuable addition to conventional statistical methods. These analytical methods give us the opportunity to better study, e.g., the effects of the renal function, diabetes and body mass index on the outcome of patients undergoing heart surgery.
On the basis of our results we summarize that the renal function is an important determinant of 30-days mortality in cardiac surgery. This risk factor is not well captured in the standard EuroSCORE model. Creatinine clearance calculated according to the Cockcroft and Gault equation should be considered in the preoperative assessment of the renal function instead of serum creatinine. This procedure results in a significant improvement of the risk estimation.