Cardiac surgery-associated acute kidney injury (CSA-AKI) has received considerable attention in recent years.1,2 Despite its multifactorial origin, some characteristics of CSA-AKI make it an interesting topic for research: high incidence, a known time of onset, a detailed epidemiology, and several accurate and validated prediction models based on well-defined risk factors.3,4 For all these reasons, cardiac surgery patients are frequently included in studies on biomarkers of acute kidney injury.

The last decade has seen the introduction of new biomarkers,5 some of which have been widely tested,6–8 However, despite more than 10 years of intense research, the new biomarkers of acute kidney injury are not yet ready for clinical use. Clinicians treating this syndrome are still in need of a therapeutic tool to help their patients. The study of the predictive capacity of these biomarkers in different settings, including cardiac surgery (CS), has provided a considerable body of new evidence. Unfortunately, none of these biomarkers has reached the clinical phase,9 and diagnosis and discrimination are still based on serum creatinine (sCr) concentration and urine output.

The estimated glomerular filtration rate (eGFR) based on sCr levels has already been a focus of research in CSA-AKI.10 Preoperative eGFR has been shown to be superior to sCr in predicting the need for renal replacement therapy (RRT) in the postoperative period,11 and is independently associated with other negative outcomes, such as prolonged hospitalisation or in-hospital mortality.12 In this study, we hypothesised that postoperative eGFR may be useful as an early predictor of AKI after CS. To test this hypothesis, we studied the predictive capacity of several common parameters used to measure postoperative renal function in a cohort of cardiac surgery patients. Specifically, we evaluated the use of eGFR calculated using the MDRD-4 (4-variable Modification of Diet in Renal Disease) formula, serum urea and their respective changes over time as biomarkers to predict AKI one or two days in advance. Biomarkers were defined according to the criteria put forward by the Biomarkers Definition Working Group.13

A total of 4167 major cardiac surgeries were performed over the 12-year study period. A total of 205 interventions were excluded (85 due to the type of surgery, 48 due to preoperative RRT, 38 due to early death, 19 due to previous kidney transplant, 15 were the second intervention during the same admission, five due to missing data, and three involved an LVAD). Some patients met more than one exclusion criterion. Of the remaining 3962 cases, a preliminary exploratory sample of 1980 cases was randomly selected. Table 1 shows the main demographic and clinical characteristics of this sample. The information relating to the validation sample can be seen in Table 1B of the additional material.

The frequency and severity of AKI in both samples are shown in Table 2.

Only the results of the best biomarker (eGFR MDRD-4) are shown. The results for the rest of the parameters (urea, Δurea, ΔMDRD-4 and ΔsCr) can be seen in Appendix B of the additional material.

Fig. 1 shows the time-dependent ROC curves corresponding to the 3 categories of AKI, also taking into account the time from surgery to AKI. The curves in Fig. 1A are derived exclusively from data of cases with severe AKI (AKIN stage III) and controls, those in Fig. 1B from cases with moderate AKI (AKIN stage II) and controls, and finally those in Fig. 1C from mild cases and controls. These curves show the estimated cumulative sensitivity and dynamic specificity of eGFR MDRD-4. Estimates have been calculated for prediction one day (continuous line) and two days (dotted line) in advance. As shown, sensitivity and specificity depend not only on the time between measurement and diagnosis of the event, but also on the time from CS to the measurement. One-day predictions are always better (continuous line), particularly when MDRD-4 is measured two days after surgery.

Fig. 1.

Time-dependent ROC curves (cumulative/dynamic) in the exploratory sample. Cases of AKI versus control: (A) Severe AKI (stage III). (B) Moderate AKI (stage II). (C) Mild AKI (stage I). The unbroken line corresponds to the 1-day prediction and the dotted line corresponds to the 2-day prediction.

(0.53MB).

The information provided by these curves can be easily interpreted from their corresponding AUCs (Table 3). The eGFR MDRD-4 biomarker showed a high predictive capacity for AKI in all categories (i.e., mild, moderate and severe) regardless of the day the measurement was taken, with AUC values higher than 0.7 in all cases. Similar results are obtained when sensitivity values are calculated with a specificity of 80% (Table 4).

The optimal cut-off points for eGFR MDRD-4, with a sensitivity of 80%, are shown in Table 5. Appendix B of the additional material (Table A.4) shows similar analyses using baseline changes in sCr (ΔsCr) as the biomarker. As expected, sCr is not a good predictor, with low AUC values in all cases. Although serum urea, Δurea and ΔMDRD-4 showed an acceptable predictive capacity in some cases, they were never superior to MDRD-4.

The results obtained from the exploratory sample were validated in the validation sample (n=1982). The 90% CIs obtained from the equivalence tests are shown in Fig. 2. The results from this second sample were similar to those obtained from the first.

Fig. 2.

90% CI for the difference in AUC (log transformation) between the exploratory and validation samples.

The dotted lines show the equivalence margin with Δ=20% (widest range) and with Δ=10% (narrowest range).

(0.27MB).

In this study, we used time-dependent ROC curves to evaluate the capacity of certain common, inexpensive kidney function laboratory biomarkers to predict onset of CSA-AKI one or two days in advance. The main finding is that eGFR calculated using the MDRD-4 formula is a reliable predictor of AKI after major CS. The limitations of sCr for the early diagnosis of AKI are well known, and have also been confirmed using this methodology. In the specific context of CS, the usual delay before sCr values become elevated after the intervention is amplified due to the haemodilution and positive fluid balance typical of cardiopulmonary bypass. It is also well known that, in the presence of AKI, eGFR decrease precedes an increase in sCr. Likewise, the determination of GFR based on equations such as MDRD does not reflect real filtration in acute situations. These formulas are unsuitable for patients with AKI.22 In our study, however, eGFR was used as a biomarker to predict a future event, not as a parameter to quantify real-time kidney function. Surprisingly, the estimation of GFR based on sCr weighted by certain demographic characteristics (age, race, gender) was superior to GFR obtained from sCr alone in predicting the onset of AKI.

Severe CSA-AKI, defined by the need for RRT, has a poor prognosis, with hospital mortality rates of around 50%. Various predictive indexes that include preoperative risk factors have been developed3,4,23 and validated16,24 to detect at-risk subjects and improve patient information. New serum and urine biomarkers of AKI have greatly improved the early detection of an episode of AKI in different clinical scenarios, starting with paediatric CS more than a decade ago.25 The usefulness of all these AKI biomarkers has been evaluated with conventional ROC curves. These studies estimate the sensitivity and specificity of the biomarker, the optimal cut-off points, and the time from the peak value of the biomarker to the diagnosis of AKI.

We took a different approach. Cox regression models are used to process time to event data when the aim is to study the association between certain biomarkers and future events. The main problem is that a biomarker can show a close association with the event, and nevertheless, not be useful for discrimination. Heagerty17 introduced the concept of the time-dependent ROC curve as an extension of standard ROC curves for the analysis of time to event data. To our knowledge, time-dependent ROC curves have not been used in the specific context of CSA-AKI with a real set of clinical data. Pepe et al.26 compared different methods of estimating time-dependent ROC curves, but used simulated data. More recently, Pitili et al.27 used time-dependent ROC curves to evaluate a combination of biomarkers to predict the onset of RRT in mechanically ventilated ICU patients. In our study, we used cumulative/dynamic time-dependent ROC curves, the most widely used approach.17

The information provided by time-dependent ROC curves should be interpreted in terms of proportions. For example, one day after surgery (see Table 4), using eGFR MDRD-4 as a biomarker and assuming a specificity of 80%, 89.4% of cases of severe AKI (AKIN stage III) can be predicted one day in advance, and 85.6% two days in advance. As expected, when the biomarker is measured on the day of surgery, the predictive capacity decreases. For example, on the day of surgery, only 67.4% of patients with severe AKI and 45.4% of patients with moderate AKI (AKIN stage II) can be predicted one day in advance. This same result has also been reported with some of the latest biomarkers,28 and is due to both the brief interval from the start of injury to measurement, and to the haemodilution that occurs during cardiopulmonary bypass (CPB), which is when the parameters are measured on admission to the ICU.

We have also provided the optimal cut-off points (Table 5), which are extremely useful when applying the biomarker in daily practice. The eGFR MDRD-4 values measured on the day of surgery did not discriminate between cases and controls in any of the three AKI categories. However, if measured one day after surgery, a patient will be classified as AKIN stage III 24h in advance if their eGFR MDRD-4 is less than or equal to 51ml/min/1.73m2, as AKIN stage II when 51<MDRD-4≤75, as AKIN stage I when 75<MDRD-4≤80, and as no AKI when their eGFR MDRD-4 is >80. From our point of view, this information is particularly valuable, since many patients are discharged to the ward the day after surgery. Using these cut-off points, the value of eGFR MDRD-4 the day after surgery can help decide the best option for the patient. These options include: discharge from the ICU (versus remaining in the unit or transfer to an intermediate care unit), administration of nephrotoxic drugs (suspension, switch or restart), the need for strict glycaemic control, the maintenance of haemodynamic monitoring, and consultation with specialists from the Nephrology Department.29 Recently, Meersch et al.30 showed that the implementation of a “KDIGO measures bundle in cardiothoracic surgery” in patients undergoing CS with a high risk of AKI, identified by means of a new AKI biomarker measured four hours after CPB, reduced the frequency and severity of CSA-AKI.

This study has several limitations. First, its retrospective, single-site design called for prospective, external validation: the cut-off points obtained may not be transferable to other institutions. Secondly, defining AKI on the basis of the AKIN's sCr criterion is open to criticism: we were unable to diagnose AKI using the urine output criterion because the database did not include this variable – a common problem in most retrospective studies. In the specific context of CS, diagnosing AKI on the basis of the urine output criterion is controversial, and has recently been called into question.31 There is no consensus on the optimal definition and classification of AKI in CS, although some authors propose a modified RIFLE classification to avoid over-diagnosis of AKI when both AKIN criteria are used.32 This phenomenon mainly affects patients at AKIN stage I. However, in our sample, the predictive capacity of MDRD-4 improved with the severity of AKI (Table 3). Because prognosis is directly related to the severity of AKI, we believe that over-diagnosis is unlikely in patients with mild AKI, considering that using eGFR MDRD-4 as a biomarker is particularly useful in moderate and severe cases. A final limitation is the effect of the use of baseline values of eGFR MDRD-4. The performance of some new biomarkers (for example, urine NGAL) is known to be affected by the presence of chronic kidney disease,33 and this means that the test is not informative. We did not explore the role of baseline eGFR MDRD-4 values in predicting AKI due to the small number of cases with moderate or severe AKI. In patients with a preoperative eGFR under 60ml/min/1.73m2, the probability of presenting with an episode of AKI and the optimal cut-off points for MDRD-4 may differ from the figures given in Tables 4 and 5.

Determination of eGFR MDRD-4 one or two days after surgery is a good predictive tool for AKI in patients undergoing major CS. Until new, more accurate biomarkers are found, the cut-off points for this routine, low-cost biomarker can help guide decision-making in the ICU. The design of this study, with its sample size, primary endpoint defined as time to event, and the availability of serial measures for the different parameters analysed, is the ideal setting for using time-dependent ROC methodology, and this research will hopefully help to raise awareness and increase acceptance of this methodology.

This study was partially funded by research project no. MTM2016-75351-R of the Spanish Ministry of Economy, Industry and Competitiveness.

The authors declare that they have no conflicts of interest.