Invited Commentary
Nephrology
AI-Enabled Acute Kidney Injury Prediction and the Challenge of Prevention
Charat Thongprayoon, Francesco Pesce, Wisit Cheungpasitporn
JAMA Netw Open 2026;9;(7):e2622565. doi:10.1001/jamanetworkopen.2026.22565
Acute kidney injury (AKI) remains one of the most common, serious, and consequential complications of hospitalization. It is associated with longer hospital stays, higher costs, greater short-term and long-term mortality, and increased risk of chronic kidney disease. Yet, despite decades of investigation, preventing AKI in hospitalized patients remains difficult. A central challenge is timing: by the time serum creatinine rises, the injury may already be established. This reality has driven enthusiasm for predictive analytics, biomarkers, and electronic decision support systems that can identify patients before overt AKI develops. The key question, however, is no longer whether we can predict AKI. It is whether prediction can be converted into prevention.
In this issue of JAMA Network Open, Churpek et al1 report a randomized clinical trial evaluating whether a structured early nephrology consultation, triggered by a real-time artificial intelligence (AI) machine-learning AKI risk score, could reduce subsequent kidney injury among hospitalized patients without AKI at enrollment. The investigators used the ESTOP-AKI (electronic signal to prevent AKI) model2 to identify patients at increased risk of stage 2 AKI and randomized them to early nephrology consultation or usual care. The intervention was thoughtful and clinically grounded: nephrologists assessed volume status, kidney perfusion, medication selection and dosing, electrolytes, nutritional needs, and additional diagnostic testing. Despite this proactive approach, early nephrology consultation did not reduce the peak change in serum creatinine over 7 days, incident AKI, severe AKI, kidney replacement therapy, mortality, readmission, or 90-day outcomes.
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. 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. 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. Those actions require human and system-level implementation.
Indeed, one of the most revealing findings was that consult recommendations in the early consultation arm were often not followed. This observation may explain, at least in part, why the intervention did not improve outcomes. However, adherence should be interpreted carefully. Recommendations in the usual care arm were made after the primary team requested nephrology consultation, likely when kidney dysfunction, electrolyte abnormalities, or volume issues were already clinically apparent. In contrast, recommendations in the early consultation arm were triggered by predicted risk, before serum creatinine–defined AKI had developed. The same recommendation may be perceived differently when it is made in response to an overt clinical problem rather than a predicted future event.
This distinction is central to the implementation challenge. Preventive nephrology consultation asks clinicians to act before injury is visible. Primary teams may view such recommendations as less urgent, particularly when patients have competing acute problems, unclear volume status, active infections, hemodynamic instability, or complex medication trade-offs. Some recommendations may also overlap with tasks that primary teams believe they already manage, such as electrolyte repletion, diet modification, or medication adjustment. In this context, a consult note, even when delivered by an expert, may be insufficient as a stand-alone intervention to reliably change downstream care.
The findings align with prior AKI alert and recommendation trials. Electronic alerts alone have often failed to improve AKI outcomes.3 More intensive kidney action team approaches have also shown limited clinical benefit when recommendations are not consistently implemented.4 Conversely, trials of biomarker-guided or protocolized KDIGO (Kidney Disease: Improving Global Outcomes)–based care bundles in selected high-risk surgical populations have suggested that AKI prevention may be possible when risk identification is paired with concrete, structured, and adhered-to actions.5,6 The lesson is not that alerts, risk scores, or consultations are futile. Rather, they must be embedded within workflows that make the desired action easy, timely, and accountable.
Another important question is whether peak change in serum creatinine over 7 days is the optimal primary end point for prevention trials. It is pragmatic and biologically relevant, but serum creatinine is delayed, influenced by volume status, affected by muscle mass, and imperfectly linked to patient-centered outcomes. Larger trials may be needed to assess outcomes such as persistent AKI, major adverse kidney events, kidney replacement therapy, and postdischarge kidney function. At the same time, implementation outcomes should be elevated from secondary observations to central trial end points: Was the risk score delivered? Was it trusted? Was the recommendation acknowledged? Was the action completed? How quickly? Which components were most strongly associated with benefit?
The next generation of AI-enabled AKI prevention trials should move 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. Not all recommendations are equal. Future interventions should identify and prioritize the highest-yield actions, such as stopping or dose adjusting nephrotoxins, addressing hypotension or venous congestion, optimizing fluid and diuretic strategies, avoiding unnecessary contrast exposure, and evaluating obstruction when appropriate. These interventions should be assigned to specific team members, tracked in real time, and followed to completion. Future systems may also benefit from phenotype-guided deployment strategies, as patients with hemodynamic, nephrotoxic, congestive, or inflammatory AKI risk profiles may differ substantially in modifiability and responsiveness to intervention.
Churpek et al1 should be commended for conducting a randomized clinical trial of an AI-enabled AKI prevention strategy rather than stopping at model development or retrospective validation. Their study demonstrates that machine-learning risk scores can identify a high-risk population before creatinine-defined AKI, but it also shows that identification alone is insufficient. The central lesson is clear: prediction is not prevention. To improve AKI outcomes, the field must now focus less on whether algorithms can forecast kidney injury and more on how health systems can reliably act on those forecasts.
