Not All Sepsis-Associated Acute Kidney Injury Is the Same

Artificial intelligence can be broadly defined as the development and implementation of computer systems that are capable of performing tasks that normally require human intelligence. Artificial intelligence is omnipresent in our modern world, with smartphones providing facial recognition and applications (apps) using machine learning algorithms to see through the infinite options to predict exact items for a future purchase on Amazon or find the perfect Netflix show to binge watch while decompressing from a weekend call. Artificial intelligence can also be found throughout medicine; it is in the electronic medical record alerting clinicians to the patient’s risk for a diagnosis/complication or guiding medication dosing (1). The last several years have seen increased investigation around artificial intelligence in the fields of AKI and critical care. It is with this background that Nadkarni et al. (2) used artificial intelligence and advanced machine learning techniques to sift through laboratory results, vital signs, and clinical data to identify subphenotypes of critically ill patients with sepsis-associated AKI.

Although standardized consensus definitions have informed the epidemiology and clinical care of patients with AKI, these definitions remain imperfect in their sole reliance on serum creatinine and urine output (3). All AKI is classified by changes in these two limited biomarkers, with no current diagnostic distinction between sepsis-associated AKI, nephrotoxin-associated AKI, and cardiac surgery–associated AKI. For decades, non-nephrologists have thought of AKI in terms of changes in serum creatinine and urine output, not necessarily seeing the changes in other laboratory results: rising phosphate levels, decreasing serum bicarbonate, and elevations of lactate. Harnessing the power of these other laboratory results as well as vital signs, medication exposures, and other diagnostics has allowed our group and others to develop machine learning algorithms to predict the development of AKI over 1 day prior to changes in serum creatinine or urine output (4,5). As an example, these risk scores have utilized rising white blood cell counts, hypotension, tachycardia, positive blood cultures, and increasing lactates as potential surrogates for sepsis and sepsis-associated AKI.

As we move away from sole reliance on serum creatinine and urine output as the only biomarkers for AKI, we recognize that not all AKI behaves in the same fashion. Within the main subheadings of AKI (e.g., sepsis, cardiac surgery, nephrotoxin), there are disease clusters that likely represent patient subsets with distinct pathophysiologic mechanisms. In those with sepsis-associated AKI, some patients may have a more inflammatory injury pattern, whereas others may be more ischemic from hypotension, and others may have nephrotoxin injury (e.g., from aminoglycosides or elevated vancomycin levels) in the setting of septic shock. Our current classifications and biomarkers do not allow for distinguishing between these clinical entities. Understanding and identifying which of these injury patterns predominates in a unique patient will be paramount to determining who is at risk for specific clinical outcomes and which targeted novel therapeutics are likely to be most effective in which patients.

This process of determining the disease clusters, referred to as subphenotypes, has already begun in the field of sepsis-associated AKI. Using latent class analysis with a discovery (n=794) cohort and a validation (n=425) cohort from the Vasopressin in Septic Shock Trial, Bhatraju et al. (6) identified two distinct AKI subphenotypes. Each subphenotype displayed different clinical outcomes (7-day kidney nonrecovery and 28-day mortality) and a unique response to vasopressin therapy. The subphenotypes discriminated clinical outcome better than Kidney Disease Improving Global Outcomes staging, and this performance was further enhanced when the phenotype classification was combined with endothelial (angiopoietin 1 [Ang-1] and Ang-2) and inflammatory (IL-8) biomarkers. Further, this same group looked at genetic variation and single-nucleotide polymorphisms within the Ang-1, Ang-2, and TNF receptor 1A genes in 421 subjects of European ancestry. They demonstrated that a single-nucleotide polymorphism within the Ang-2 gene led to decreased plasma Ang-2 concentrations and likely plays a causal role with regard to subphenotype categorization (7). It remains to be seen if phenotype-specific interventions targeting Ang-2 production or other aspects of these sepsis-associated AKI phenotypes can improve or alter care.

In this edition of CJASN, Nadkarni et al. (2) expand on subphenotyping to further our understanding of sepsis-associated AKI. Using data from the Medical Information Mart for Intensive Care III database, they performed a retrospective observational study identifying 4001 patients with sepsis-associated AKI within the first 48 hours of intensive care unit (ICU) admission. Then, using vital signs and laboratory measurements from the first 48 hours along with comorbidities (188 variables/features), they used deep learning techniques to identify three unique clinical subphenotypes of sepsis-associated AKI. They demonstrated significant differences in the presence of preadmission comorbidities (e.g., baseline CKD, liver disease, and congestive heart failure all identified via diagnostic coding) across these three groups as well as differences in laboratory values (e.g., liver function tests, lactate, and white blood cell count) and clinical outcomes (need for KRT or 28-day mortality). They demonstrated that the highest-risk group (subphenotype 3) had more severe disease as measured by higher sepsis scores and acute-phase reactants (lactate, white blood cell count, liver function tests, and lactate dehydrogenase). They then performed manual chart review for 30 patients from each subphenotype in order to determine the predominant cause of AKI. There were significantly different numbers of patients adjudicated to have acute tubular necrosis and prerenal azotemia across these three groups, with more patients with acute tubular necrosis in subphenotype 3.

The authors should be commended for defining AKI on the basis of both the serum creatinine and urine output criteria, something that was not done with the AKI risk prediction algorithms (4,5). They clearly demonstrate subphenotypes of sepsis-associated AKI, each with unique clinical trajectory and outcomes. However, there remain several limitations to this work. First and foremost, as the authors acknowledge, this study suffers from all of the inherent biases of retrospective single-center data. Second, they define sepsis via the use of diagnostic coding rather than using a consensus definition (8). Third, although they had their reasons, the authors limited the patients with sepsis-associated AKI to only those with cases that occurred within the first 48 hours of ICU admission. This is potentially problematic, especially when one considers that many patients develop AKI after this time point and that sepsis and AKI have a bidirectional relationship (9). Although most are familiar with AKI coming after the onset of sepsis, data from the Program to Improve Care in Acute Renal Disease showed that sepsis developing a median of 5 days after AKI was predictive of worse clinical outcomes. Furthermore, the patients in their cohort had sepsis and AKI, but this did not guarantee that they had sepsis-associated AKI, as their AKI could have been from nephrotoxins or ischemia rather than from sepsis itself. Finally, owing to the limitations of the dataset, they defined baseline serum creatinine (necessary to subsequently determine the incidence of AKI) as the lowest value within 7 days of ICU admission. Although this represents a limitation, it is also important to note that there is no standard method for determining baseline in the absence of prehospital data. Importantly, the authors acknowledge many of these limitations as well as state that the next step should be external validation of their data in other cohorts.

This last point is of the utmost importance, as the majority of artificial intelligence tools that have been developed (in AKI and elsewhere) started with the use of retrospective data. These tools, whether used to develop subphenotypes or stratify AKI risk, need to be prospectively and independently validated, just like any other medical intervention. The validation process should include external multicenter validation and then potentially a prospective randomized examination of that tool’s ability to improve patient outcomes. Recently, various forms of artificial intelligence have been prospectively evaluated (10,11). Wilson and colleagues (10) prospectively implemented an AKI risk prediction tool and used it as a trigger to collect and measure urinary biomarkers in patients at high risk for AKI. Separately, Goldstein et al. (11) implemented a nephrotoxin screening tool and demonstrated its ability to produce a sustained reduction in nephrotoxin exposure in pediatric patients at high risk for AKI.

Prospective investigation of these new sepsis-associated AKI subphenotyping tools to determine their clinical utility is essential. It is easy to imagine a scenario where the sicker patients of subphenotype 3 are identified within the first day of ICU admission and randomized to receive a novel therapeutic that specifically targets sepsis’ microvascular dysfunction or the immunologic dysregulation of sepsis-associated AKI. However, if targeted treatment of the subphenotypes described by Nadkarni et al. (2) does not affect or improve the clinical care or outcomes of patients with sepsis-associated AKI in a meaningful manner, they should not be used. This is no different than artificial intelligence tools outside of medicine and is why when you are done reading this editorial, you may view streaming content on Netflix or Hulu rather than on failed platforms, such as Aereo or Seeso.

Disclosures

J. Koyner reports consultancy agreements with Astute Medical, Baxter, and Sphingotec; research funding from Astute Medical, Bioporto, National Institutes of Health, Nxstage Medical, and Satellite Dialysis–Coplon Grant; honoraria from the American Society of Nephrology and the Society of Critical Care Medicine; is a scientific part of the editorial board for American Journal of Nephrology, Kidney360, and the Scientific Advisory Board for the National Kidney Foundation and the National Kidney Foundation of Illinois; and is part of the speakers bureau of NxStage Medical. The remaining author has nothing to disclose.

Funding

None.