AI-Triggered Nephrology Consult Fails to Prevent Acute Kidney Injury in Trial

Early nephrology consultations triggered by real-time machine-learning risk scores failed to prevent increases in peak serum creatinine among hospitalized patients at risk for severe acute kidney injury (AKI), a randomized trial showed.

Among 180 patients, the adjusted mean difference in 7-day serum creatinine was comparable between those who had an early nephrology consultation and those who received usual care (0.04 mg/dL vs -0.03 mg/dL, P=0.30), reported Jay L. Koyner, MD, of the University of Chicago, and colleagues.

Additionally, there was no significant difference between the two arms in the development of Kidney Disease: Improving Global Outcomes (KDIGO) stage 1 or higher AKI (42% vs 36%, P=0.47) or stage 2 or higher AKI (19% vs 13%, P=0.28), they wrote in JAMA Network Open.

The intervention also did not show a significant effect on secondary endpoints such as AKI severity, hospital length of stay, inpatient mortality, or 90-day outcomes.

While the intervention group received substantially more recommendations versus the usual care group, clinician adherence to these recommendations was low. “It is possible that this low uptake of early consultative recommendations contributed to the trial outcome,” Koyner and team pointed out.

During the study period, there were 121 early nephrology consultations containing 270 recommendations compared with 19 usual care consultations and 36 recommendations. The most common recommendations in the intervention and usual care groups included:

• Changing patients’ diet: 28.9% vs 13.9%
• Stopping medications: 20.7% vs 16.7%
• Changing medication doses: 11.9% vs 11.1%

Recommendations for medication dosage and discontinuation, diuretics or fluids, and vasopressors were more likely to be completely followed in the usual care arm (68%) compared with the intervention arm (41%).

Evidence-based recommendations may simply be ineffective before KDIGO-defined AKI fully manifests, Koyner and colleagues suggested. “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,” they wrote.

Because so many hospitalized patients have general AKI risks, primary teams may struggle to distinguish critical recommendations from minor ones, they added.

In an accompanying commentary, Wisit Cheungpasitporn, MD, of the Mayo Clinic in Rochester, Minnesota, and co-authors called the negative trial findings “important,” cautioning that they do not reflect a failure of machine learning or proactive kidney care.

“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,” they wrote. “A model may identify a patient at risk, but it does not discontinue nephrotoxins, adjust medication doses, optimize hemodynamics, administer fluids or diuretics, evaluate obstruction, or ensure follow-up.”

Because those actions require human and system-level execution, the commentators argued that the field must pivot from asking whether algorithms can forecast kidney injury to how health systems can reliably act on those forecasts.

The single-center trial ran at the University of Chicago from March 2019 to August 2024, randomizing 180 hospitalized patients without baseline AKI who triggered an Electronic Signal to Prevent AKI (ESTOP-AKI) score greater than 0.01. Median patient age was 62.5 years, 56.7% were men, 52.8% were white, and 41.1% were Black.

Triggered by ESTOP-AKI score, the automated nephrology consultation included an in-person assessment covering volume status, kidney perfusion, drug dosing, electrolytes, nutrition, and diagnostic testing. Because these findings were framed as traditional, non-mandatory consultation notes rather than direct orders, primary teams were not obligated to follow them. The usual care group only received nephrology consultations upon a direct, manual request from the primary team.

The researchers noted that conducting the trial at a single academic center limited the generalizability of the findings to other settings.

To bridge the gap, Cheungpasitporn’s team suggested that future artificial intelligence-enabled AKI trials shift from “risk-triggered advice to risk-triggered action.”

“This may require closed-loop systems in which high-risk alerts activate prioritized care bundles, pharmacist-led medication review, nurse-supported volume and urine output assessment, embedded order sets, nephrotoxin stewardship, and rapid reassessment,” they wrote.