Expert: ‘We Can Do Better’ to Assess Risk of Acute Kidney Injury

Historically, scoring a patient’s risk of acute kidney injury (AKI) is based on calculating serum creatinine levels, or looking at baseline kidney function; however, these measurements may not be the most effective means of identifying patients at risk of AKI. Machine learning can improve this risk scoring in the hospital setting, according to Jay Koyner, MD, during a session at the American Society of Nephrology (ASN) 2023 Kidney Week Annual Meeting. Koyner, an associate professor of medicine at the University of Chicago, suggested that a better risk scoring model could lead to better treatment outcomes with earlier intervention and reduced deleterious nephrotoxins.

“In patients with stage 2 AKI, we [still] don't do all the things we think we're doing,” Koyner said in the session. “Patients are allowed to get hyperkalemic… patients are allowed to be hypotensive, [and] that's just not right.”

Newer risk scoring mechanisms can include more data to stratify the risk or severity of AKI; further, assessing the risk score using dynamic data can only expound on current efforts to assess risk with static data (collected prior to hospitalization), Koyner said. Dynamic data includes information taken directly during operation, like diastolic blood pressure and pulse rate or basic metabolic panel tests collected while in the hospital.

Cleveland Clinic affirmed this hypothesis in a cohort study of 60,000 patients, which was validated in 5000 patients. During the study, collecting dynamic data improved the detection of severe AKI (stage 2/3 or patients who need dialysis). This data demonstrated that testing risk for AKI can go beyond creatinine and other biochemical biomarkers, according to Koyner. Researchers in Belgium were also able to successfully evaluate other disease markers not commonly used to assess AKI risk, including acute physiologic assessment and chronic health evaluation score.

Researchers in Belgium were also able to successfully evaluate other disease markers not commonly used to assess AKI risk, including acute physiologic assessment and chronic health evaluation score. Image Credit: © Crystal light - stock.adobe.com

“When you demonstrate any of the data that we're talking about, you need to actually demonstrate that it makes a clinical difference,” Koyner said.

Complex machine learning can have a hand in AKI risk analysis as well. It can be used, for instance, to alert practitioners about a risk score to inform subsequent biomarker testing. Colleagues of Koyner who collected data on alerts showed that it can be used to indicate AKI risk, according to Koyner. Likewise, it may be capable of doing phenotyping to help identify a source or medication driving AKI risk, or improve complicated risk score assessments, since it can analyze more variables than traditional systems, he said.

“[But] like most things, it's great in theory but a lot harder to do,” Koyner said.

While risk scoring models possess good sensitivity and specificity, they lack positive predictive value; Moreover, these machines cannot always account for the comorbidities and individuality of each person’s response to treatment, Koyner said.

Koyner also emphasized that these complex machine learning models need to be retrained over time. They will lose specificity gradually, which can affect risk scores and will necessitate recalibration.

“You need to make sure the model fires the way it should on the people it should,” Koyner said in the session.

In line with this concept, Koyner suggested that alerting physicians and patients about risk scores could reduce nephrotoxic care for patients with AKI. These alerts might even work to prevent AKI onset and improve mortality outcomes.

“To me, I like it when my patients stay alive,” Koyner said.

These alerts could enable physicians to change the course of their care and reduce nephrotoxic actions; for instance, taking a more thoughtful approach to offering dialysis (preventing over dialysis in patients who do not need it, or being alerted about patients who may need it most [surprisingly, these would be low risk patients]), NSAIDs, or proton pump inhibitors—the latter may have the worst impacts on mortality outcomes.

Likewise, education is key. In a 5-institution study that took place in England, providers who were educated about care for patients with AKI were more thoughtful about offering dialysis, higher rates of ultrasound (from 20% to 55%), and lower length of stay in the ICU, according to Koyner. “The idea is clear. We need to be thoughtful about the nephrotoxic care we’re providing,” Koyner said.

Reference

Koyner J, Not your Father’s Risk Score: Dynamic Risk Assessment and Alerts for AKI. Session. ASN Kidney Week. November 2 to 5, 2023.