Editorial March 8, 2022
Predictive Models for Acute Kidney Injury Following Cardiac Surgery: The Importance of Accurate and Actionable Prediction
Marlies Ostermann, Nuttha Lumlertgul, Francis Perry Wilson
JAMA. 2022;327(10):927-929. doi:10.1001/jama.2022.1823
Acute kidney injury (AKI) complicates recovery from cardiac surgery in up to 40% of patients; impairs the function of the heart, lungs, brain, and gut; and is associated with an increased risk of death during hospitalization.1 AKI that requires kidney replacement therapy after cardiac surgery is associated with an increased 28-day mortality ranging from 15% to 85%, depending on acute and chronic comorbidities.2
The current international Kidney Disease: Improving Global Outcomes consensus definition of AKI is based on a serum creatinine increase of 0.3 mg/dL or more in 48 hours or less, a decrease in urine output, or both.3 However, not every increase in serum creatinine necessarily represents kidney injury. Similarly, AKI can be present despite a smaller or no increase in serum creatinine, such as in the setting of liver disease, fluid overload, or muscle wasting.4 Lassnigg et al analyzed data from 4118 patients who underwent cardiac and thoracic aortic surgery and showed that small changes of serum creatinine between the preoperative baseline value and the maximal serum creatinine value within 48 hours after the surgical procedure were associated with an increase in mortality.5 Furthermore, patients with a decrease of serum creatinine of more than 0.3 mg/dL also had progressively increasing 30-day mortality of up to 8%. Thirty-day mortality was lowest (2.1%) among patients in whom serum creatinine decreased by 0.3 mg/dL or less and increased to 6% among patients in whom serum creatinine remained unchanged or increased up to 0.5 mg/dL. Patients with a serum creatinine increase greater than 0.5 mg/dL had a 30-day mortality rate of 32.5%.5
Patients with AKI who survive are also more likely to develop chronic kidney disease and dialysis-dependent kidney failure in the following years than those without postoperative AKI. Xu et al analyzed the data from 3245 patients with normal preoperative kidney function (estimated glomerular filtration rate ≥60 mL/min/1.73 m2) who underwent cardiac surgery between April 2009 and December 2012.6 Overall, 1295 patients (39.9%) developed postoperative AKI. The 1- and 2-year survival rates of patients with AKI were significantly lower than in those who did not have AKI (1-year: 85.9% vs 98.1%; 2-year: 82.3% vs 93.7%) (P < .001), even after initial recovery of kidney function. The accumulated prevalence of progressive chronic kidney disease was also significantly higher in patients with AKI than in those without AKI in the 2 years after surgery (6.8% vs 0.2%; P < .001), including among patients who had initially recovered kidney function at hospital discharge.
An accurate assessment of an individual patient’s risk provides an opportunity to directly influence clinical decision-making and potentially prevent AKI through adjustment of hemodynamic management, fluid therapy, and avoidance of nephrotoxic drugs. Further, accurate risk models can support the decision-making process before surgery and be used as research tools to identify high-risk patients for clinical trials.
Several predictive models of cardiac surgery–associated AKI exist.7-10 A 2012 systematic review of the risk prediction models available at the time concluded that most models used differing AKI criteria, were based on small cohorts, or lacked external validation.9 The models for AKI requiring dialysis after cardiac surgery were the most robust and externally validated. In that review, the Cleveland Clinic Score by Thaker et al10 was considered the most widely studied model with high discrimination in most of the tested populations.
More scores have been evaluated since then. Kristovic et al applied risk scores of 5 prediction models to 1056 adult patients undergoing cardiac surgery from 2012 to 2014.7 Again, the Cleveland Clinic Score10 showed the best performance in the prediction of dialysis-dependent AKI (area under the receiver-operating characteristic curve [AUC], 0.837 [95% CI, 0.810-0.862]) and in the prediction of stage 1 AKI and higher (AUC, 0.731 [95% CI, 0.639-0.761]) and stage 2 AKI and higher (AUC, 0.811 [95% CI, 0.783-0.838]).
In this issue of JAMA, Demirjian et al from the Cleveland Clinic report the performance of new prediction models in adult patients undergoing coronary artery bypass graft, valve, or aorta surgery.11 The authors derived the models in a cohort of 58 526 patients from the main academic hospital and performed external validation in a cohort of 4734 patients from 3 affiliated community hospitals. The models were based on routinely available results of postoperative blood samples obtained for clinical reasons, including changes in serum creatinine from preoperative values, serum potassium, bicarbonate, sodium, albumin, and blood urea nitrogen, adjusted for the time between conclusion of the surgical procedure to the first postoperative blood draw. The median (IQR) time from the conclusion of cardiac surgery and first-available blood results was 10 (7-12) hours in the derivation cohort and 6 (2-11) hours in the validation cohort. The diagnosis of AKI was based on a modified Kidney Disease: Improving Global Outcomes definition for AKI diagnosis using serum creatinine criteria only with the most recent preoperative result as baseline value. The authors stratified the model performance by time from surgery to blood draw (≤8 h vs >8 h) and results were similar.
In the derivation cohort, the metabolic panel–based models showed very good discrimination for moderate to severe AKI within 72 hours and 14 days after the surgical procedure, with an AUC of 0.876 within 72 hours and 0.854 within 14 days, and for AKI requiring dialysis within 72 hours and 14 days after the surgical procedure, with an AUC of 0.916 within 72 hours and 0.900 within 14 days. In the validation cohort, the models for moderate to severe AKI within 72 hours following surgery had an AUC of 0.860 and within 14 days had an AUC of 0.842; models for AKI requiring dialysis within 72 hours following surgery had an AUC of 0.879 and within 14 days had an AUC of 0.873.
The strengths of this study include the size of the patient cohort, a robust statistical analysis, the conduct of an external validation, and the restriction of variables in the models to quantitative data that are routinely available and relatively inexpensive. Factors that are usually influenced by clinical decision-making, such as the choice of “on-pump” vs “off-pump” bypass surgery and the use of potentially nephrotoxic drugs or an intra-aortic balloon pump, were not included in the models, nor were procedural factors that may be difficult to measure, such as aortic cross-clamp time.
In addition to its predictive utility, the study by Demirjian et al serves as a reminder that even relatively minor changes in serum creatinine matter and are associated with worse outcomes, as previously shown by Lassnigg et al.5,11 Importantly, the rate of serum creatinine change between preoperative and postoperative results outperformed absolute changes. In the derivation cohort, perioperative median (IQR) serum creatinine increased by 0.20 (0-0.40) mg/dL in patients with stage 2/3 AKI within 72 hours, 0.15 (0-0.39) mg/dL in patients with stage 2/3 AKI within 14 days, 0.13 (−0.07 to 0.39) mg/dL in patients requiring dialysis within 72 hours, and 0.1 (−0.10 to 0.35) mg/dL in patients with AKI requiring dialysis within 14 days.
Strategies aimed at optimizing hemodynamics and fluid status and avoiding hyperglycemia and nephrotoxic exposures have been shown to be effective at preventing moderate or severe AKI in high-risk patients after cardiac surgery.12 Thus, it is widely acknowledged that these patients should be identified as early as possible. A major limitation in clinical practice is the lack of reliable tools to complete this task in a timely and practical manner. Although novel kidney biomarkers are effective and included in some guidelines, these injury biomarkers are not universally available.13
Demirjian and colleagues offer a promising alternative.11 The depth and breadth of data in the electronic health record, combined with ever more powerful statistical and machine learning tools, has led to a proliferation of accurate risk models for a variety of diseases, including AKI.14 Future studies are needed to evaluate whether the provision of these risk models can improve patient outcomes. For that to occur, the models must not only be accurate, but also actionable, and must offer information beyond that available based on the perception or intuition of a treating clinician. Ideally, techniques are needed that determine and monitor glomerular filtration rate continuously to answer many questions that serum creatinine cannot so that patients with AKI could be identified in real time and different AKI subtypes could be determined. Further, the search for specific drugs to prevent AKI needs to continue to support personalized action and care.
It has been argued that hemodynamic optimization and avoidance of nephrotoxic drugs should be routine practice in every patient undergoing cardiac surgery, avoiding the expense of additional prediction scores and novel diagnostic tests. However, observational data from 237 patients at 12 hospitals confirmed that in routine clinical practice, all recommended AKI prevention strategies were applied in less than 10% of patients.12,15 An accurate predictive model may prompt clinicians to improve on these best practices. However, even a tool with perfect predictive ability will only be useful if it leads to changes in clinical decisions that lead to changes in outcome. A prediction score without appropriate clinical action is a potentially futile exercise.
