A novel machine learning-based risk assessment tool successfully identified hospitalized patients at risk for acute kidney injury (AKI) and triggered early nephrology consultation, but the intervention did not significantly improve kidney outcomes. Adherence to the resulting recommendations was also lower than with usual care, underscoring the challenges of translating early risk prediction into effective prevention.

“To our knowledge, our [randomized controlled trial] is the first to use a machine learning risk score to identify patients at high risk and then trigger an AKI-focused intervention,” the authors wrote in the study, published in JAMA Network Open.

“However, using this score to trigger an [early nephrology consultation], we found no statistical differences in the peak rise in serum creatinine or the subsequent development of KDIGO [Kidney Disease: Improving Global Outcomes] AKI in those receiving an [early nephrology consultation].”

Commenting on the findings, nephrologist Wisit Cheungpasitporn, MD, professor of medicine at the Mayo Clinic in Rochester, Minnesota, said the central message was that “prediction is not prevention.”

“Artificial intelligence may help identify the right patient at the right time, but improved outcomes will require risk-triggered action rather than risk-triggered advice alone,” he told Medscape Medical News.

AKI in critically ill patients is associated with substantial morbidity, chronic kidney disease, and mortality. Although efforts to prevent AKI have been underway for decades, progress has been limited.

Risk scores have been developed to identify vulnerable patients before serum creatine-based AKI occurs. However, most clinical trials of preventive interventions have focused on patients who already developed AKI, the authors noted.

Assessing a Deep Learning Strategy

To evaluate whether a real-time deep learning score-based strategy could trigger early nephrology consultation to prevent AKI, first author Matthew M. Churpek, MD, and colleagues conducted a randomized trial of 180 patients hospitalized at the University of Chicago, Illinois, between March 2019 and August 2024.

Eligible patients did not have serum creatinine-based AKI at enrollment but had a machine learning-based electronic signal to prevent AKI (ESTOP-AKI) of at least 0.01, indicating a high risk for stage II AKI.

Patients were randomly assigned 1:1 to receive either a structured early nephrology consultation from an attending nephrologist (n = 89; 49.4%) or usual care (n = 91; 50.6%).

The early consultation included an in-person assessment and recommendations to the patient’s care team regarding volume status, kidney perfusion, medication selection and dosing, electrolyte management, nutritional needs, and additional testing.

In the usual care group, nephrology consultation was only provided when requested by the primary team.

The patients had a median age of 62.5 years, 43.3% were women, and 52.8% were White. Baseline characteristics, including medical history, were similar between the study groups.

Comparable Outcomes

For the primary endpoint, there was no significant between-group difference in peak serum creatinine change during the 7 days after enrollment. This remained true after adjustment for a higher ESTOP-AKI score, age, sex, patient location, and baseline serum creatinine (0.04 mg/dL vs -0.03 mg/dL with usual care; P = .30).

Overall, 70 patients (38.9%) developed AKI of any stage over the 7-day follow-up. The incidence of stage I or higher AKI did not significantly differ between the early consultation and usual care groups (42% vs 36%; P = .47), nor did the incidence of stage II or higher AKI (19% vs 13%; P = .28).

At further 90-day follow-up, there were also no significant differences between the early consultation and usual care groups in hospital readmission rates (34.1% vs 44.4%; P = .21) or mortality (14.8% vs 18.7%; P = .62).

The intervention group received 121 early nephrology consultations that resulted in 270 recommendations compared with only 19 consultations and 36 recommendations in the usual care group.

Despite the greater number of recommendations, adherence was lower in the early consultation group. Recommendations involving medication dosage or discontinuation, diuretics or fluids, and vasopressor use were completely followed in 41% of cases in the early consultation group and 68% in the usual care group.

The authors noted that the low compliance rates were similar to those observed in previous studies evaluating implementation of strategies among patients with established AKI.

“While [previous] trials differ from ours in that we excluded patients with established AKI, our rate of adherence to recommendations was similarly under 50% (ie, 48%),” they wrote.

Lack of Perceived Urgency May Hinder Adherence

One reason for the low adherence may have been that patients had not yet developed overt AKI or the marked creatinine elevations that typically suggest urgency, said lead author Jay L. Koyner, MD, of the Section of Nephrology at the University of Chicago.

“It is fairly common to only call nephrology for a consult when creatinine levels have already doubled or tripled,” he told Medscape Medical News.

“So the idea of receiving recommendations when the creatinine has increased by 0.1-0.2 mg/dL rather than 1.0-2.0 mg/dL is foreign to many physicians and may explain, in part, why consult recommendations were not completely followed.”

Even when recommendations are implemented, evidence remains lacking regarding their effectiveness before AKI develops.

“It is not clear if stopping exposure to a nephrotoxin or dose reducing a medication before there is evidence of serum creatinine-based AKI improves outcomes,” the authors wrote.

“I suspect that different interventions carry a different level of significance in different patients,” Koyner added.

For instance, “maintaining mean arterial pressures may be more important in cardiac patients compared to stopping nephrotoxic medication, which may be more important in other patients, such as septic patients.”

Not Necessarily a Failure

In an accompanying editorial, Cheungpasitporn and colleagues argued that the results should not necessarily refute the benefits of early counseling.

“The negative result is important, but it should not be interpreted as a failure of machine learning or as evidence that proactive kidney care has no value,” they wrote. “Rather, the trial highlights a recurring lesson in clinical decision support: risk identification is only the first step in a much longer chain of clinical translation.”

Cheungpasitporn said the high number of recommendations may itself have contributed to the relatively low compliance.

“When clinicians receive numerous recommendations for a patient without visible kidney injury, the most important actions may lose urgency amid competing priorities,” he said.

Patient selection and broader study outcomes could also better reflect improvements, he added.

“Future studies may benefit from selecting patients whose risk is both high and potentially modifiable, rather than treating all predicted risk as equivalent,” Cheungpasitporn said. “Outcomes beyond short-term peak creatinine may also be more informative, including persistent AKI, major adverse kidney events, kidney replacement therapy, and post-discharge kidney function,” he said.

Prevention Still Depends on Individualized Care

Cheungpasitporn said the most effective AKI prevention strategies remain individualized and focused on reversible risks. These include:

• Avoiding or dose-adjusting nephrotoxic medications
• Promptly addressing hypotension and impaired kidney perfusion
• Assessing both volume depletion and venous congestion
• Using fluids and diuretics judiciously
• Avoiding unnecessary contrast exposure without delaying clinically indicated diagnostic imaging
• Monitoring urine output and kidney function
• Evaluating obstruction when appropriate

“Pharmacist-supported medication review and structured kidney care bundles may be particularly useful in selected high-risk populations when the recommended actions are promptly and consistently implemented,” he added.

Koyner reported receiving grants from the National Institute of Diabetes and Digestive and Kidney Diseases, the National Institutes of Health, and SphingoTec, as well as consulting and/or research fees from bioMérieux, BioPorto, Fresenius Medical, Alexion, Vantive, SeaStar Medical, Novartis, AstraZeneca, and Guard Therapeutics outside the submitted work. Cheungpasitporn reported no relevant disclosures.