Can AI Help Prevent Acute Kidney Injury? What Hospitalists Should Know About ESTOP-AKI

Can AI Help Prevent Acute Kidney Injury? What Hospitalists Should Know About ESTOP-AKI

Acute kidney injury (AKI) is one of the most common and important complications encountered in hospital medicine.

The challenge is timing.

By the time serum creatinine rises enough to meet conventional criteria for AKI, kidney injury may already be established.

That creates an intriguing question:

What if we could identify patients at high risk for AKI before creatinine rises?

Machine learning may help clinicians do exactly that.

But predicting AKI and preventing AKI are two very different things.

The ESTOP-AKI randomized clinical trial provides an important lesson for hospitalists interested in artificial intelligence, predictive analytics, and clinical decision support.

The study suggests that a machine-learning model can identify patients at elevated risk for AKI—but simply identifying those patients and triggering an early nephrology consultation does not necessarily improve outcomes.

What Is ESTOP-AKI?

ESTOP-AKI stands for:

Electronic Signal to Prevent Acute Kidney Injury.

ESTOP-AKI is a machine-learning risk model developed to identify hospitalized patients at increased risk of developing more severe AKI.

Rather than waiting for serum creatinine to rise, the model attempts to recognize risk earlier in the patient’s hospitalization.

This represents an important shift in thinking.

Traditional AKI recognition asks:

“Does this patient have AKI?”

Predictive analytics asks:

“Is this patient likely to develop AKI?”

That difference potentially creates an opportunity for prevention.

But only if identifying risk leads to an intervention that actually changes the patient’s outcome.

Why Predict AKI Before It Happens?

AKI is associated with substantial morbidity and healthcare utilization.

For hospitalists, kidney injury can complicate the management of:

• Sepsis
• Heart failure
• Shock
• Infection
• Surgery
• Volume depletion
• Medication toxicity
• Critical illness

AKI can also affect decisions involving:

• IV fluids
• Diuretics
• Antibiotics
• Contrast
• Antihypertensive medications
• Medication dosing
• Electrolytes
• Nutrition
• Disposition

Earlier identification of high-risk patients therefore has obvious appeal.

The hope is that clinicians could intervene before significant kidney injury develops.

The Critical Difference Between Prediction and Prevention

This is the central lesson from ESTOP-AKI.

A predictive model may successfully identify risk.

But that does not automatically mean the model improves patient outcomes.

For AI to improve care, several things must happen:

Prediction → Clinical recognition → Appropriate intervention → Implementation → Better outcome

A failure anywhere in that chain can reduce the clinical value of the prediction.

ESTOP-AKI provides a useful real-world example.

How Did the ESTOP-AKI Trial Work?

The randomized clinical trial was conducted at the University of Chicago.

Researchers enrolled hospitalized adults who:

• Did not already have serum creatinine–based AKI
• Had an elevated ESTOP-AKI risk score
• Met the study’s other eligibility criteria

A total of 180 patients were randomized.

Patients were assigned to:

Early nephrology consultation

or

Usual care.

The early consultation was triggered by the machine-learning risk score.

What Did the Early Nephrology Consultation Include?

The intervention was more than simply notifying the primary team that a patient was high risk.

An attending nephrologist performed a structured assessment.

Recommendations could address:

• Volume status
• Kidney perfusion
• Medication selection
• Medication dosing
• Fluids
• Diuretics
• Vasopressors
• Electrolytes
• Nutrition
• Additional diagnostic testing

The concept was logical:

Use AI to identify high-risk patients → involve kidney specialists earlier → intervene before AKI develops.

The question was whether this strategy actually worked.

What Was the Primary Outcome?

The primary outcome was the peak change in serum creatinine during the seven days after enrollment.

The investigators wanted to determine whether early nephrology consultation triggered by the ESTOP-AKI model reduced subsequent kidney injury.

It did not.

There was no statistically significant difference in the primary outcome between the early-consultation and usual-care groups.

Did ESTOP-AKI Prevent Acute Kidney Injury?

No significant reduction in AKI was demonstrated.

Stage 1 or higher AKI developed in:

42% of the early nephrology consultation group

versus

36% of the usual-care group.

The difference was not statistically significant.

Stage 2 or higher AKI occurred in:

19%

versus

13%,

again without a statistically significant difference.

Therefore, this randomized trial did not demonstrate that the ESTOP-AKI-triggered early nephrology intervention prevented AKI.

Did It Reduce Dialysis or Mortality?

The intervention also did not demonstrate significant improvements across the other major clinical outcomes studied.

The investigators evaluated outcomes including:

• Kidney replacement therapy
• Inpatient mortality
• Readmission
• 90-day mortality

The trial did not demonstrate significant improvement in these outcomes with the early nephrology consultation strategy.

What About 90-Day Outcomes?

At 90 days, readmission occurred in approximately:

34% of the early consultation group

and

44% of the usual-care group.

Despite the numerical difference, it was not statistically significant.

Ninety-day mortality was approximately:

15%

versus

19%,

which was also not statistically significant.

These results should therefore not be interpreted as evidence that the intervention reduced readmissions or mortality.

Why Didn’t the Intervention Work?

This is where the study becomes especially interesting for hospitalists.

The negative trial does not necessarily mean:

“AI doesn’t work.”

It also does not necessarily mean:

“Early nephrology consultation is useless.”

Instead, the trial highlights a larger challenge in clinical AI:

Accurate risk prediction is only the first step.

The intervention triggered by that prediction must also be effective and implemented.

Recommendation Adherence May Be Important

One notable finding involved implementation of nephrology recommendations.

The early-consultation arm generated substantially more recommendations.

However, several recommendations involving:

• Medication dosing or discontinuation
• Fluids
• Diuretics
• Vasopressors

were less likely to be fully followed in the early-consultation group than comparable recommendations in the usual-care group.

That observation raises an important possibility:

What happens after an AI alert may matter as much as the alert itself.

An algorithm can identify risk.

A consultant can make recommendations.

But the recommendations still need to become actual clinical care.

The Prediction-to-Action Gap

Hospitalists should become familiar with what might be called the:

prediction-to-action gap.

Imagine that an AI model tells you:

This patient has a high probability of developing AKI.

That information is only clinically useful if there is something effective to do with it.

The next questions become:

• What intervention should be triggered?
• Who receives the alert?
• How quickly must someone respond?
• Which medications should be reviewed?
• Should fluids change?
• Should diuretics change?
• Should nephrotoxins be avoided?
• Does the patient need additional monitoring?
• Who owns the recommendation?
• How do we know the intervention occurred?

These are workflow questions as much as artificial-intelligence questions.

AI Risk Scores Are Not Treatment

This distinction is essential.

A risk score is not a therapy.

Predictive algorithms can potentially identify:

• AKI risk
• Sepsis risk
• Clinical deterioration
• Readmission risk
• Mortality risk

But prediction alone does not treat the patient.

Clinical benefit depends on whether the prediction leads to an effective intervention.

This is one reason healthcare AI should be evaluated based on patient outcomes, not merely model performance.

AUC Is Not the Same as Clinical Benefit

AI studies often report measures of predictive performance such as:

• Area under the receiver operating characteristic curve
• Sensitivity
• Specificity
• Positive predictive value
• Negative predictive value

These metrics can tell us how well a model predicts an outcome.

They do not necessarily tell us whether deploying that model improves patient care.

A model could be highly accurate yet produce little clinical benefit if:

• There is no effective intervention
• Alerts arrive too late
• Alerts are ignored
• Clinicians experience alert fatigue
• Recommended actions are not implemented
• The workflow is poorly designed

ESTOP-AKI demonstrates why randomized implementation studies are so important.

Why This Trial Matters Beyond AKI

The implications extend far beyond nephrology.

Hospitals are increasingly interested in predictive tools for:

• Sepsis
• Clinical deterioration
• ICU transfer
• Readmission
• Falls
• Delirium
• Length of stay
• Discharge readiness
• Patient flow

The temptation is to assume:

Better prediction = better care.

But that equation is incomplete.

A more accurate framework is:

Prediction + effective intervention + reliable implementation = potential clinical benefit.

Each component needs evidence.

What Should Hospitalists Do With an AKI Risk Alert?

An AI alert should not replace clinical judgment.

Instead, a high-risk signal can potentially prompt clinicians to reassess the patient.

Questions may include:

Is the Patient Hemodynamically Stable?

Review:

• Blood pressure
• Perfusion
• Volume status
• Recent fluid balance
• Urine output

Are There Potential Nephrotoxins?

Review medications and exposures that may contribute to kidney injury.

Are Medications Appropriately Dosed?

Renally cleared medications may require reassessment as kidney function changes.

Does the Patient Need Fluids?

Some patients may benefit from fluid administration.

Others—particularly patients with heart failure or volume overload—may be harmed by unnecessary fluids.

The risk score cannot answer this question by itself.

Does the Patient Need Diuresis?

Again, clinical context matters.

A patient with cardiorenal physiology may require a very different strategy from a patient with true volume depletion.

For additional hospital medicine review, see Heart Failure in Hospital Medicine.

AKI in Sepsis

Sepsis represents an important clinical setting for AKI.

Patients may experience kidney injury related to:

• Hemodynamic instability
• Inflammation
• Microcirculatory dysfunction
• Medication exposure
• Critical illness
• Other mechanisms

Management requires treating the underlying infection while carefully evaluating hemodynamics, medications, fluid strategy, and organ function.

For additional review, read 2026 Surviving Sepsis Campaign Guidelines: What Every Hospitalist Needs to Know.

AKI in Heart Failure

Kidney function can also become challenging in patients hospitalized with acute heart failure.

Rising creatinine may occur in the setting of:

• Congestion
• Diuresis
• Hemodynamic changes
• Medication changes
• Underlying chronic kidney disease

Hospitalists should avoid interpreting every increase in creatinine identically.

Clinical context matters.

For a broader review, see Heart Failure in Hospital Medicine.

AKI and Critical Care

Critically ill patients often have multiple simultaneous AKI risk factors.

These may include:

• Shock
• Sepsis
• Mechanical ventilation
• Vasopressors
• Surgery
• Medications
• Multiorgan failure

An AI prediction in this setting should therefore be interpreted alongside the overall clinical picture.

For practical critical-care education, read ICU Pearls for Hospitalists.

Avoid Alert Fatigue

Predictive tools can create another problem:

too many alerts.

If clinicians receive frequent warnings that do not clearly change management, alerts may eventually be ignored.

Hospitals implementing predictive analytics should therefore evaluate:

• How frequently alerts fire
• Positive predictive value
• Who receives alerts
• When alerts appear
• Whether clinicians can act on them
• Whether alerts duplicate existing information
• Whether alerts improve outcomes

The goal is not to generate the largest number of predictions.

The goal is to identify actionable risk at the right time.

Should Every High-Risk Patient Receive a Nephrology Consultation?

ESTOP-AKI does not support a blanket conclusion that every patient identified as high risk by a machine-learning model should automatically receive early nephrology consultation.

The trial tested a particular strategy:

machine-learning risk identification + structured early nephrology consultation.

That strategy did not improve the primary outcome or reduce incident AKI in the trial.

This does not mean nephrology consultation is inappropriate when clinically indicated.

Hospitalists should continue to request specialist input based on the patient’s condition, diagnostic uncertainty, severity of kidney disease, electrolyte abnormalities, need for kidney replacement therapy, and other clinical considerations.

What ESTOP-AKI Does Tell Us

The trial provides several important lessons.

Lesson 1: AI Can Identify Risk

Machine-learning tools can potentially recognize patients at elevated risk before traditional clinical thresholds are reached.

Lesson 2: Prediction Is Not Prevention

Knowing who is at risk does not automatically change the outcome.

Lesson 3: The Intervention Matters

A predictive model is only as clinically useful as the intervention it triggers.

Lesson 4: Implementation Matters

Recommendations that are not carried out cannot improve care.

Lesson 5: Randomized Trials Matter

Predictive performance alone is not enough.

Healthcare AI should ultimately be tested for its effect on clinically meaningful outcomes.

Questions Hospital Leaders Should Ask Before Deploying Predictive AI

Before implementing an AI clinical prediction tool, hospitals should ask:

What exactly does the model predict?

How accurate is it in our patient population?

Has it been externally validated?

When does the prediction occur?

What action does the prediction trigger?

Is that intervention evidence based?

Who is responsible for responding?

How quickly must they respond?

How will adherence be measured?

Could the tool create alert fatigue?

Could it create unintended harm?

Has implementation been shown to improve patient outcomes?

These questions are often more important than simply asking how sophisticated the algorithm is.

What This Means for Hospitalists

Hospitalists are likely to encounter increasing numbers of predictive tools integrated into electronic health records.

These may involve:

• AKI
• Sepsis
• Deterioration
• Readmission
• Discharge
• Patient flow

Clinicians should approach these tools with informed skepticism rather than either automatic enthusiasm or automatic rejection.

Ask:

What does this tool predict?

How reliable is that prediction?

What am I supposed to do differently because of it?

Is that action supported by evidence?

Has the entire intervention actually improved patient outcomes?

Those questions help distinguish a useful clinical tool from an impressive prediction model.

Frequently Asked Questions

What is ESTOP-AKI?

ESTOP-AKI is a machine-learning risk model designed to identify hospitalized patients at increased risk of developing stage 2 acute kidney injury.

Was ESTOP-AKI tested in a randomized trial?

Yes. A randomized clinical trial evaluated whether structured early nephrology consultation triggered by an elevated ESTOP-AKI score improved outcomes in hospitalized patients who did not yet have serum creatinine–based AKI.

How many patients were included?

The trial randomized 180 hospitalized patients.

Did early nephrology consultation prevent AKI?

No significant reduction in AKI was demonstrated.

Did it reduce mortality?

No statistically significant mortality benefit was demonstrated.

Does that mean the AI model failed?

Not exactly.

The model identified a high-risk population. The trial tested whether a specific intervention triggered by that model—early structured nephrology consultation—improved outcomes.

That intervention did not significantly improve the measured outcomes.

Does this mean AI cannot prevent AKI?

No.

The trial evaluated one particular predictive model and intervention strategy. It does not establish that every future AI-based AKI prevention strategy will fail.

Should hospitalists ignore AI AKI alerts?

No. Predictive alerts should be interpreted in clinical context.

The important question is whether an alert provides actionable information and whether the action it triggers is supported by evidence.

Bottom Line

ESTOP-AKI illustrates one of the most important lessons in clinical artificial intelligence:

Predicting disease is not the same as preventing disease.

The machine-learning model identified hospitalized patients at increased risk for AKI.

But using that risk prediction to trigger early structured nephrology consultation did not significantly reduce:

• Peak creatinine change
• Incident AKI
• Severe AKI
• Kidney replacement therapy
• Mortality
• Readmission

The next generation of clinical AI research therefore needs to move beyond asking:

“Can the algorithm predict the outcome?”

and increasingly ask:

“Does using the algorithm change what clinicians do—and does that improve outcomes that matter to patients?”

For hospitalists, that is the distinction that matters most.