Renal insufficiency after cardiac surgery is associated with increased mortality, morbidity, and intensive care unit (ICU) stay (1,2). The preoperative predictors of postoperative dialysis have been identified in large studies that encompassed up to 43,000 patients (1,2). In contrast, research into strategies to prevent renal insufficiency has been limited. An important reason for this discrepancy is the infrequent incidence of postoperative dialysis (1% to 2%) in this patient population (1,2). Clinical trials therefore require large sample sizes to detect changes in dialysis rates with adequate power.
Surrogate measures of perioperative renal dysfunction are therefore needed for the preliminary evaluation of new renal-protective therapies using feasible sample sizes. The ideal surrogate measure should be easily measured and correlate with patient-relevant outcomes. It should also have distributional properties suited to parametric statistical tests, which are generally more powerful than nonparametric equivalents (3). If a therapy beneficially alters an ideal surrogate measure, there is justification for then studying its effects on patient-relevant outcomes (dialysis, mortality, and length of stay) in an adequately powered larger study. The surrogate measures presently in use include urinary protein markers, serum creatinine (Cr) concentration, and Cr clearance (CrCl) (measured and calculated).
Urinary levels of α1-microglobulin, β2-microglobulin, and N-acetyl-β-d-glucosaminidase are sensitive markers of real tubular injury. However, there is no evidence that perioperative increases of these markers are associated with postoperative morbidity or mortality (4). Furthermore, antifibrinolytics (e.g., ε-aminocaproic acid, tranexamic acid, and aprotinin) cause microglobulinuria independent of renal injury (5).
Serum Cr concentration has been used as a measure of overall renal function. It is easily measured and has a small interassay coefficient of variation (<5% at our institution). However, Cr concentrations are affected by factors aside from renal function, namely, sex, age, and muscle mass (6). Investigators have usually either used Cr concentration as a continuous variable or dichotomized it to define clinically significant changes in renal function. Definitions of clinically significant perioperative changes in Cr concentration have been largely arbitrary. The most common definition, an increase more than 44.2 μmol/L (0.5 mg/dL) over 48 to 72 h (7,8), was obtained from the contrast-induced nephropathy literature (9). The significance of a 44.2 μmol/L perioperative increase in serum Cr is unknown.
Measured CrCl has been used to estimate glomerular filtration rate (GFR), but its accuracy and precision have generally been low (10). Other disadvantages include increased personnel time and costs that are needed for accurate collection of timed urine samples. The shorter collection times used in perioperative trials (11) may further magnify inaccuracies in urine volume measurement.
GFR can also be estimated by calculated CrCl. The latter estimates GFR through an equation that accounts for an individual’s age, sex, weight, and serum Cr (12). It has advantages over both measured Cl and serum Cr. Calculated Cl estimates GFR and measured Cl in critically ill patients admitted to medical and cardiac surgical ICUs (13,14), even during conditions of changing renal function (14). Furthermore, calculated Cl does not require the collection of cumbersome timed urine samples.
Cr concentrations and calculated CrCl are convenient, reproducible, and inexpensive surrogate mea-sures. However, the prognostic significance and distributional properties of perioperative changes in these measures remain unclear. We therefore evaluated the surrogate measures in 2000 patients undergoing coronary artery bypass grafting, valve surgery, or both. We also sought to identify the appropriate perioperative changes in Cr and CrCl for defining clinically significant changes in renal function. We limited postoperative measurements of renal function to the 72 h after surgery, a previously defined period (15) that would capture changes in function attributable to surgery itself. We used the percentage change in CrCl to better relate the perioperative change in Cl to preoperative renal function.
Preoperative, intraoperative, and postoperative information is prospectively collected for patients undergoing cardiac surgery at the Toronto General Hospital (TGH) (Toronto, ON, Canada). Documented information includes weight (kg), postoperative dialysis, in-hospital mortality, and ICU length of stay. This database has previously been described (16). Variables are defined in Appendix 1. We identified all individuals who underwent coronary artery bypass grafting, valve surgery, or both between May 1, 1999, and August 31, 2000. Patients were excluded if they required dialysis before surgery or died during surgery.
Assuming a dialysis rate of 0.75%, the sample size of 2000 was estimated to provide sufficient outcomes (death or dialysis) for logistic regression modeling (i.e., 10 outcomes per independent variable) (17) and 95% confidence intervals (CI) more precise than ±0.10 during receiver operating characteristic (ROC) curve analysis. CIs were calculated by using the methods of Hanley and McNeil (18), assuming a ROC curve area of 0.90. The sample size was also deemed to be feasible for manual retrieval of laboratory data. We randomly selected 2000 individuals from the 3077 patients identified previously, by using a computer-generated random-number table (SAS 8.02; SAS Institute, Cary, NC).
The Cr concentrations of patients undergoing cardiac surgery at TGH are routinely measured before surgery (within 30 days) and then daily after surgery until discharge or the fifth postoperative day. After obtaining approval from the TGH Research Ethics Board, we manually reviewed medical charts to obtain these results. The preoperative Cr concentration (Crpreop) was defined as the value recorded closest to surgery but not on the day of the procedure itself. The maximal 72-h Cr concentration (Cr72hmax) was defined as the largest postoperative measurement between Days 0 and 3. All in-hospital measurements were performed at Toronto Medical Laboratories (Toronto, ON, Canada) with the Bayer Advia 1650 autoanalyzer (Bayer Inc., Tarrytown, NY), which has an interassay coefficient of variation <5%. Of Crpreop measurements, 97% were measured at Toronto Medical Laboratories. The overall results were not qualitatively altered when we excluded the 3% with preoperative measurements at facilities other than Toronto Medical Laboratories. Preoperative measurements were performed on the day before surgery for 63% (n = 1267) of the study sample. The duration between the preoperative measurement and surgery ranged from 1 to 27 days (mean, 5 days; 25th percentile, 1 day; 75th percentile, 11 days).
The 72-h change in Cr concentration (ΔCr72h) was defined as the difference between preoperative and postoperative concentrations (ΔCr72h = Cr72hmax − Crpreop). Preoperative (CrClpreop) and smallest 72-h postoperative (CrCl72h) CrCl were calculated with the Cockcroft-Gault equation (12):
- Men: CrCl = [(140 − age) × weight × 1.2] ÷ serum Cr
- Women: CrCl = [(140 − age) × weight] ÷ serum Cr
Units for age, weight, and Cr were years, kilograms, and micromoles per liter, respectively. The Cockcroft-Gault equation was selected given its superior performance relative to other equations (19). The absolute 72-h change in CrCl (ΔCrCl72h) was calculated by using preoperative and postoperative CrCl: ΔCrCl72 h = (CrCl72 h − CrClpreop). The percentage 72-h change in CrCl (%ΔCrCl72h) was subsequently determined: %ΔCrCl72 h = (ΔCrCl72 h/CrClpreop) × 100%.
Patients who needed postoperative dialysis received either intermittent hemodialysis or continuous venovenous hemodialysis. Decisions on implementing dialysis were made by a consulting nephrologist.
Analyses were performed with SAS 8.02. All tests of significance were two tailed, with P values less than 0.05 considered statistically significant.
The distributions of ΔCr72h and %ΔCrCl72h were assessed by using frequency histograms and normal probability plots. We assessed the distributions of the percentage change in Cr concentration, %ΔCr (%ΔCr = ΔCr72 h/Crpreop × 100%), and the absolute change in CrCl (ΔCrCl72h) to ensure that differences in their distributions were not due to %ΔCrCl72h being a percentage. We also determined whether any mathematical trans-formations (logarithmic, etc.) converted ΔCr72h or %ΔCrCl72h to symmetric, bell-shaped distributions.
ROC curves (18) were plotted for the relationship of ΔCr72h and %ΔCrCl72h with postoperative dialysis, in-hospital mortality, and prolonged ICU length of stay. Prolonged ICU stay was defined as more than or equal to 5 days (95th percentile for the study sample). The methods of Hanley and McNeil (18) were used to calculate areas under each ROC curve with 95% CI. The ROC curve area gives a global assessment of the performance of ΔCr72h or ΔCrCl72h in discriminating between individuals who do or do not have an outcome. Measures that perform no better than chance have ROC curve areas of 0.50; measures with perfect discrimination have areas of 1.0. The measures’ ROC curve areas were compared by using the methods of Hanley and McNeil (20).
We developed several alternative definitions of clinically significant changes in perioperative renal function by dichotomizing ΔCr72h and %ΔCrCl72h along convenient cut-points, including a 72-h increase in Cr concentration more than 44.2 μmol/L (0.5 mg/dL). We used the ROC curves to identify the cut-points that resulted in the optimal balance of sensitivity and specificity for predicting postoperative dialysis.
We also measured the correlation of ΔCr72h and %ΔCrCl72h with dialysis, mortality, and prolonged ICU stay by using univariate logistic regression modeling. The statistical significance of the relationship of ΔCr72h or %ΔCrCl72h (as continuous variables) with these outcomes was assessed by using the Wald statistic (21). Odds ratios with 95% CI were calculated. The discrimination and calibration of the regression models were determined with the c statistic and the Hosmer-Lemeshow goodness-of-fit statistic, respectively (21). Discrimination refers to a model’s ability to assign higher probabilities to individuals who sustain outcomes as opposed to those who do not (22). Calibration describes the degree to which models’ predicted probabilities compare against actual outcomes (22).
The effect of perioperative changes on renal function may be affected by the presence of preoperative renal dysfunction. We therefore determined the predictive performance of ΔCr72h and %ΔCrCl72h in two strata defined by CrClpreop. Only two strata were used, given the limited statistical power of a 2000-patient sample. We initially plotted an ROC curve comparing CrClpreop and postoperative dialysis; this curve was subsequently used to identify the CrClpreop value with the optimal balance of sensitivity and specificity for predicting dialysis. The sample was then divided into two strata (normal and reduced preoperative renal function) on the basis of this CrClpreop value. ROC curves were subsequently plotted comparing ΔCr72h and %ΔCrCl72h with dialysis within each stratum.
The study sample (Table 1) had an in-hospital mortality rate of 1.1% (n = 21). The median ICU length of stay was 1 day (25th percentile, 1 day; 75th percentile, 2 days). The incidence of postoperative dialysis and prolonged ICU stay was 1.0% (n = 20) and 6.3% (n = 126), respectively.
The distribution of ΔCr72h (Fig. 1) was skewed to the right (median, 3 μmol/L; 25th percentile, −5 μmol/L; 75th percentile, 14 μmol/L). The variable was not normally distributed on a normal probability plot. The transformation to %ΔCr (median, 3%; 25th percentile, −5%; 75th percentile, 15%) was also skewed to the right and not normally distributed. Transformation of ΔCr72h to the difference between the reciprocals of preoperative and postoperative Cr concentrations (1/Crpreop − 1/Cr72hmax) resulted in a more symmetric, bell-shaped distribution.
ΔCr72h was correlated with postoperative dialysis (ROC area, 0.98; 95% CI, 0.94–1.00), mortality (ROC area, 0.83; 95% CI, 0.72–0.94), and prolonged ICU stay (ROC area, 0.74; 95% CI, 0.69–0.79). The sensitivity, specificity, and prevalence of several cut-points in ΔCr72h are presented in Table 2. The cut-point that resulted in the optimal balance of sensitivity and specificity was an increase in Cr concentration more than 50 μmol/L, which had a prevalence of 5.5%. ΔCr72h was also associated with clinical outcomes in univariate logistic regression models (Table 3). All models had good discrimination (Table 3). The models were also well calibrated, aside from the model predicting prolonged ICU stay (Table 3).
The %ΔCrCl72h used here was a mathematical transformation of ΔCr72h. The distribution of %ΔCrCl72h (Fig. 1) was approximately normal (mean, −5.1%; sd, 16.4%). The distribution of the absolute change in CrCl, ΔCrCl72h, was more symmetric and bell shaped than that of ΔCr72h. However, normal probability plots showed that the transformation of ΔCrCl72h to %ΔCrCl72h resulted in a closer approximation to the normal distribution.
%ΔCrCl72h was correlated with postoperative dialysis (ROC area, 0.97; 95% CI, 0.91–1.00), mortality (ROC area, 0.82; 95% CI, 0.71–0.93), and prolonged ICU stay (ROC area, 0.74; 95% CI, 0.69–0.79). The ROC curve areas were not significantly different from ΔCr72h with respect to death (P = 0.89), dialysis (P = 0.49), or prolonged ICU stay (P = 0.85). The sensitivity, specificity, and prevalence of several cut-points in %ΔCrCl72h are presented in Table 2. The cut-point that resulted in the optimal balance of sensitivity and specificity was a decrease in CrCl more than 25%, which had a prevalence of 10.2%.
%ΔCrCl72h was also associated with clinical outcomes in univariate logistic regression models (Table 3). All models had good discrimination (Table 3). The models were well calibrated, except the model predicting prolonged ICU stay (Table 3).
The area under the ROC curve relating CrClpreop and postoperative dialysis was 0.77 (95% CI, 0.65–0.90). The optimal definition of preoperative renal dysfunction was a CrCl <60 mL/min, which had a prevalence of 27% (n = 538). The relative risk for dialysis among individuals with preoperative renal dysfunction was 5.0 (95% CI, 2.0–12.6;P = 0.0001). The ROC curves relating ΔCr72h and %ΔCrCl72h with postoperative dialysis in the entire cohort and two subgroups are presented in Figures 2 and 3.
For individuals with normal preoperative renal function, the ROC curve areas relating dialysis with ΔCr72h and %ΔCrCl72h were 0.99 (95% CI, 0.95–1.00) and 0.99 (95% CI, 0.94–1.00), respectively. An increase in Cr concentration more than 50 μmol/L, which had a prevalence of 4% (n = 53), was associated with a sensitivity and specificity for dialysis of 97% and 100%, respectively. By comparison, a decrease in CrCl larger than 25%, which had a prevalence of 9% (n = 127), had a sensitivity of 92% and specificity of 100%.
The ROC curve areas relating postoperative dialysis with ΔCr72h and %ΔCrCl72h for individuals with preoperative renal dysfunction were 0.96 (95% CI, 0.89–1.00) and 0.95 (95% CI, 0.86–1.00), respectively. An increase in Cr concentration more than 50 μmol/L, which had a prevalence of 11% (n = 57), was associated with a sensitivity of 91% and a specificity of 92%. A decrease in CrCl more than 25%, which had a prevalence of 14% (n = 76), had a sensitivity and specificity of 88% and 92%, respectively.
We evaluated the performance of two simple surrogate measures that can be readily used in clinical trials: the 72-hour changes in Cr concentration and calculated CrCl. Both were highly correlated with postoperative dialysis, death, and prolonged ICU stay. ROC analysis did not demonstrate any significant differences between the measures with regard to predicting clinical outcomes. Their performance was similar regardless of whether preoperative renal function was normal or abnormal. Preexisting renal dysfunction, as defined by CrCl less than 60 mL/min, itself increased the risk of postoperative dialysis fivefold; these findings mirror those of other investigators (1).
Calculated CrCl is a mathematical transformation of Cr concentration. Nonetheless, this transformation resulted in a normally distributed and clinically meaningful variable that retained high correlation with patient-relevant outcomes. The other mathematical transformation that resulted in a symmetric, bell-shaped distribution was the difference between the reciprocals of preoperative and postoperative Cr concentrations. This alternative transformation is difficult to interpret. We suspect that clinicians will more readily interpret 95% CI for a proportionate change in calculated CrCl than for a difference in reciprocals.
Both Cr concentrations and calculated CrCl may be used as continuous outcome variables. In this situation, CrCl has an important advantage over Cr concentration. Parametric statistical tests (e.g., Student’s t-tests and analysis of variance) all assume that the outcome variable conforms to an approximately normal distribution. Whereas CrCl has a symmetric, bell-shaped distribution, serum Cr has a skewed distribution. The symmetric distribution of CrCl did not simply reflect its use as a percentage. The percentage change in Cr concentration (%ΔCr72h) did not have a normal distribution. When performing statistical analyses on Cr concentration, one must therefore either apply nonparametric tests or transform the data such that they follow a normal distribution. The sd and 95% CI from analyses of transformed data are often difficult to interpret (23). In contrast, one can immediately apply convenient parametric statistical tests to CrCl and derive results that are readily understandable to clinicians. This difference in distributions will therefore simplify statistical analyses during preliminary sample size calculations and final data analyses (3).
Both Cr and CrCl can also be dichotomized to define clinically significant changes in perioperative renal function. We analyzed the most common definition of perioperative renal dysfunction, a 0.5 mg/dL (44 μmol/L) increase in serum Cr, and determined its predictive performance for dialysis (sensitivity, 95.0%; specificity, 61.9%), mortality (sensitivity, 61.9%; specificity, 94.1%), and prolonged ICU stay (sensitivity, 37.3%; specificity, 95.6%). ROC analysis suggested that a clinically significant change in renal function should be defined as either an increase in serum Cr concentration more than 50 μmol/L or a decrease in CrCl more than 25%. The latter definition is more conveniently used during clinical trials because of its higher prevalence (10.2% vs 5.5%). A clinical trial that defines clinically significant renal dysfunction as a decrease in CrCl more than 25% would therefore need only half as many patients to develop statistical power equivalent to that in a trial that defines clinically significant dysfunction as an increase in serum Cr more than 50 μmol/L.
There is limited prior information on surrogate measures of perioperative renal function. Charlson et al. (24) used ROC analysis to evaluate the performance of changes in serum Cr concentrations. However, they related changes in Cr to postoperative measured CrCl, not clinical end-points (e.g., death, dialysis, or prolonged ICU stay).
There are several limitations to this study. First, postoperative CrCl was calculated by using preoperative weights. Given that the Cockcroft-Gault equation uses weight to estimate muscle mass, our calculations assumed that muscle mass did not significantly change over the first 72 postoperative hours. However, surgical stress increases protein catabolism. The urinary excretion of 3-methylhistidine, a marker of skeletal muscle protein catabolism, increases up to 40% in the early postoperative period after elective surgery (25). Nonetheless, it is unlikely that remeasuring weight on each postoperative day would have significantly improved the accuracy of our results. Immediate postoperative weight increases above preoperative values because of IV fluid administration. The Cockcroft-Gault equation would interpret this additional weight as increased muscle mass, when muscle mass is likely to have decreased. Using preoperative weights will therefore lead to less overestimation of postoperative GFR than using serial postoperative weights.
Second, these results demonstrate a close statistical association of these outcome measures with dialysis, death, and prolonged ICU stay. Although the relationship has biological plausibility, a cause-effect relationship must be validated prospectively. Future clinical trials of interventions that alter Cr or CrCl must also demonstrate similar changes in these clinical outcomes.
In conclusion, both Cr concentration and calculated CrCl are surrogate measures of perioperative renal function that are convenient, inexpensive, and highly correlated with patient-relevant clinical outcomes (mortality, dialysis, and prolonged hospitalization). Other surrogate outcomes now in use (e.g., measured CrCl and urinary protein markers) do not share all these characteristics. CrCl has the additional advantage of a symmetric, bell-shaped distribution amenable to parametric statistical tests. Either measure may be used to define a clinically significant change in perioperative renal function. The most appropriate definitions appear to be either an increase in serum Cr concentration >50 μmol/L or a decrease in CrCl more than 25%.
We propose that that Cr concentration and calculated CrCl be used as surrogate measures of renal function during initial clinical trials of novel renal-protective drugs. Demonstration of the benefit of such drugs on these surrogate measures is necessary before progressing to the large studies needed for evaluating effects on the outcomes themselves.
The authors acknowledge the superb data collection and management of Susan Collins at the University Health Network.
Appendix 1: Definitions of Perioperative Variables
- Timing of operation: semiurgent (cannot leave hospital without surgery), urgent (surgery required within 72 h of presentation), and emergent (surgery required within 12 h of presentation).
- Diabetes mellitus: preoperative diagnosis of diabetes mellitus treated with insulin, oral hypoglycemic drugs, or diet.
- Hypertension: preoperative systemic hypertension necessitating medical treatment.
- Peripheral vascular disease: known carotid, aortoiliac, or femoropopliteal disease operating room cases in which the patient had a previous carotid endarterectomy or peripheral vascular operation.
- Chronic obstructive pulmonary disease: severe chronic obstructive pulmonary disease requiring daily inhaled or oral medication to improve symptoms.
- Low cardiac output syndrome: need for intraaortic balloon pump support in the operating room or the intensive care unit or need for inotropic medication (dopamine >4 μg · kg−1 · min−1, dobutamine, milrinone, or epinephrine) for at least 30 min to maintain the systolic blood pressure >90 mm Hg and the cardiac output >2.2 L · min−1 · m−2 in the ICU.
- Perioperative myocardial infarction: a new Q wave on the postoperative electrocardiogram. A myocardial infarction was also diagnosed if the MB isoenzyme (CK-MB) of creatine kinase (CK) exceeded 50 U/L, the CK-MB/CK ratio exceeded 5%, and the postoperative electrocardiogram showed a new left bundle branch block, loss of R-wave progression, or changes in the ST segment or T wave.
1. Chertow GM, Lazarus JM, Christiansen CL, et al. Preoperative renal risk stratification. Circulation 1997; 95: 878–84.
2. Mangano CM, Diamondstone LS, Ramsay JG, et al. Renal dysfunction after myocardial revascularization: risk factors, adverse outcomes, and hospital resource utilization. Ann Intern Med 1998; 128: 194–203.
3. Greenhalgh T. How to read a paper: statistics for the non-statistician. I. Different types of data need different statistical tests. BMJ 1997; 315: 364–6.
4. Baines AD. Strategies and criteria for developing new urinalysis tests. Kidney Int Suppl 1994; 47: S137–41.
5. Smith MS. Antifibrinolytic agents make α1
- and β2
-microglobulinuria poor markers of postcardiac surgery renal dysfunction. Anesthesiology 1999; 90: 928–9.
6. Levey AS, Perrone RD, Madias NE. Serum creatinine and renal function. Annu Rev Med 1988; 39: 465–90.
7. Lassnigg A, Donner E, Grubhofer G, et al. Lack of renoprotective effects of dopamine and furosemide during cardiac surgery. J Am Soc Nephrol 2000; 11: 97–104.
8. Story DA, Poustie S, Liu G, McNicol PL. Changes in plasma creatinine concentration after cardiac anesthesia with isoflurane, propofol, or sevoflurane: a randomized clinical trial. Anesthesiology 2001; 95: 842–8.
9. Solomon R, Werner C, Mann D, et al. Effects of saline, mannitol, and furosemide on acute decreases in renal function induced by radiocontrast agents. N Engl J Med 1994; 331: 1416–20.
10. Walser M. Assessing renal function from creatinine measurements in adults with chronic renal failure. Am J Kidney Dis 1998; 32: 23–31.
11. Cittanova ML, Zubicki A, Savu C, et al. The chronic inhibition of angiotensin-converting enzyme impairs postoperative renal function. Anesth Analg 2001; 93: 1111–5.
12. Cockcroft DW, Gault MH. Prediction of creatinine clearance from serum creatinine. Nephron 1976; 16: 31–41.
13. Robert S, Zarowitz BJ, Peterson EL, Dumler F. Predictability of creatinine clearance estimates in critically ill patients. Crit Care Med 1993; 21: 1487–95.
14. Bloor GK, Welsh KR, Goodall S. Comparison of predicted with measured creatinine clearance in cardiac surgical patients. J Cardiothorac Vasc Anesth 1996; 10: 899–902.
15. Hou SH, Bushinsky DA, Wish JB, et al. Hospital-acquired renal insufficiency: a prospective study. Am J Med 1983; 74: 243–8.
16. Rao V, Ivanov J, Weisel RD, et al. Predictors of low cardiac output syndrome after coronary artery bypass. J Thorac Cardiovasc Surg 1996; 112: 38–51.
17. Concato J, Feinstein AR, Holford TR. The risk of determining risk with multivariable models. Ann Intern Med 1993; 118: 201–10.
18. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982; 143: 29–36.
19. Spinler SA, Nawarskas JJ, Boyce EG, et al. Predictive performance of ten equations for estimating creatinine clearance in cardiac patients. Ann Pharmacother 1998; 32: 1275–83.
20. Hanley JA, McNeil BJ. A method of comparing the areas under receiver operating characteristic curves derived from the same cases. Radiology 1983; 148: 839–43.
21. Hosmer DW, Lemeshow S. Applied logistic regression. New York: Wiley, 1989.
22. Ash AS, Shwartz M. Evaluating the performance of risk-adjustment methods: dichotomous outcomes. In: Iezzoni LI, ed. Risk adjustment for measuring healthcare outcomes. Chicago: Health Administration Press, 1997: 427–69.
23. Bland JM, Altman DG. The use of transformation when comparing two means. BMJ 1996; 312: 1153.
24. Charlson ME, MacKenzie CR, Gold JP, Shires GT. Postoperative changes in serum creatinine: when do they occur and how much is important? Ann Surg 1989; 209: 328–33.
25. Gross E, Holbrook IB, Irving MH. The effect of elective surgery on 3-methylhistidine excretion. Br J Surg 1978; 65: 663–5.