Sandra L. Kane-Gill, PharmD, MS, is a professor of pharmacy and therapeutics and associate dean of faculty affairs at the University of Pittsburgh School of Pharmacy. 

Sandra Kane-Gill, PharmD, MS

More than half of alerts generated by a deep learning AKI prediction system were not actionable in the MEnD-AKI trial. Only 12% of high priority pharmacist recommendations reached physicians by phone, the intended primary route.¹

American Society of Nephrology (ASN) Kidney Week 2026

 in Denver, Colorado, by Sandra L. Kane-Gill, PharmD, MS, the data from the cluster randomized trial suggest that a well validated model may not translate into an effective bedside alert without attention to how alerts are generated and delivered. The pharmacist intervention did not significantly reduce major adverse kidney events at 30 days (MAKE30) compared with a passive alert, adding to 

What MEnD-AKI Revealed About AI Alerts, With Sandra Kane-Gill, PharmD, MS

The trial used a deep learning model, a type of artificial intelligence that analyzes patterns in electronic health record data, to predict progression to stage 2 or higher acute kidney injury (AKI) within 48 hours. In prior validation, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.82.² Alerts from the model prompted pharmacists to send medication recommendations to physicians, with the highest priority Level B alerts intended to reach physicians by phone.¹ ²

“Interestingly enough, our deep learning model performed well. It was robust. It was externally validated. It had good discrimination, but that doesn’t necessarily translate into a good alert,” said Kane-Gill, professor of pharmacy and therapeutics and associate dean of faculty affairs at the University of Pittsburgh School of Pharmacy, in an interview with 

. “So understanding how the model is put in place to generate the alert is pretty important.”

Earlier studies of AKI decision support had focused on detection and passive alerts, which showed only modest effects on outcomes. MEnD-AKI tested whether pairing prediction with an active pharmacist intervention could close that gap in hospitalized adults outside the intensive care unit.¹

“What we really learned here is how to put an AI model in place at the bedside, making sure there’s a human in the loop. It went through the pharmacist, and the physician decided if they wanted to follow those recommendations. We learned a lot about how to bring this type of model to the bedside,” Kane-Gill said.

MEnD-AKI Trial Design and Primary Outcome

Investigators randomized 38 hospital service clusters across 8 hospitals to the pharmacist intervention or to usual care with a passive electronic medical record alert. The intention to treat analysis included 5939 patients. MAKE30 occurred in 50.8% of intervention patients and 42.9% of usual care patients (adjusted OR, 1.05; 95% CI, 0.82–1.35), and no secondary outcome differed significantly.¹

Of the alerts in the intention to treat population, 52.3% were not actionable.¹ Kane-Gill pointed to the model’s reliance on structured health record data as one factor. 

 was often documented in clinical notes rather than the structured medical history, so some alerts suggested AKI in patients who had CKD, she said.

Most Level B Recommendations Went by Email, Not Phone

Although phone was the intended primary route, only 12% of Level B recommendations were delivered by phone rather than email.¹ Investigators cited this reliance on asynchronous email, along with non-actionable alerts and baseline comorbidity imbalances between groups, as key limitations.

They concluded that real-time delivery and improved alert actionability are needed before the approach can be considered effective.¹ Kane-Gill said future models may benefit from incorporating unstructured data, such as clinical notes.

Editor’s Note: Disclosures for Kane-Gill include the National Institute of Diabetes and Digestive and Kidney Diseases.

Kane-Gill S, Bihorac A, Kellum JA, et al. Multihospital Electronic Decision Support for Drug-Associated AKI (MEnD-AKI): A Prospective Cluster-Randomized Clinical Trial. Abstract presented at: American Society of Nephrology Kidney Week 2026; October 21–25, 2026; Denver, CO. doi:10.1681/ASN.2026fbjd7rne

Stottlemyer BA, Kellum JA, Bihorac A, et al. Multi-hospital electronic decision support for drug-associated acute kidney injury (MEnD-AKI): study protocol for a randomized clinical trial. 

. 2025;157:108055. doi:10.1016/j.cct.2025.108055

Videos

In the MEnD-AKI trial, pharmacists reviewed deep learning predictions of AKI risk across 38 hospital service clusters at 8 hospitals.

A pharmacist intervention prompted by a deep learning clinical decision support system (CDSS) did not significantly reduce major adverse kidney events at 30 days (MAKE30) among hospitalized patients outside the intensive care unit with or at risk of drug-associated .¹

 in Denver, Colorado, by Sandra L. Kane-Gill, PharmD, MS, the data from the MEnD-AKI cluster randomized trial show that pairing predictive alerts with pharmacist review did not change kidney outcomes compared with a passive alert. Investigators concluded that real-time delivery and more actionable alerts are needed before the approach can be considered effective.¹

AI Alerts Did Not Reduce Kidney Events, With Sandra Kane-Gill, PharmD, MS

Medications are a leading cause of AKI, with almost 30% of cases attributed to a drug.² CDSS tools have improved AKI detection, but passive electronic alerts alone have shown only modest effects on outcomes.¹

“As we know, nephrotoxins are a common cause of acute kidney injury, but they’re also a modifiable risk factor for acute kidney injury,” said Kane-Gill, professor of pharmacy and therapeutics and associate dean of faculty affairs at the University of Pittsburgh School of Pharmacy, in an interview with 

. “About two thirds of drugs are excreted by the kidney, and so when a patient has reduced kidney function, that drug accumulates, and there’s a potential for adverse effects.”

The trial tested whether adding a predictive deep learning model and an active pharmacist intervention to existing alerts could improve the prevention and management of drug-associated AKI in hospitalized adults outside the ICU.¹ Previous work in this population had largely focused on passive alerts.²

“We chose that model because prior work indicated that having a dedicated clinician to respond to medication management for patients with kidney dysfunction was a preferred method, and hopefully an effective method,” Kane-Gill said. “In addition, those attending physicians involved received a lot of alerts already, and they didn’t want additional alerts sent their way.”

How MEnD-AKI Paired Deep Learning Alerts With Pharmacists

Investigators randomized 38 hospital service clusters across 8 hospitals to a pharmacist intervention prompted by the machine learning CDSS or to an electronic medical record passive alert (usual care).¹ The CDSS used a deep learning model, a type of artificial intelligence that analyzes patterns in electronic health record data, to continuously predict each patient’s risk of progressing to stage 2 or higher AKI within 48 hours.²

Tiered Level A and Level B alerts led pharmacists to send recommendations to physicians via telemedicine. The intervention was delivered at the physician level, while outcomes were assessed at the patient level using de-identified data.¹

The primary endpoint was MAKE30, a composite of death, new kidney replacement therapy, or final serum creatinine ≥150% of reference.² Secondary endpoints included AKI progression, AKI intensity, and nephrotoxin burden.¹

No Reduction in Major Adverse Kidney Events at 30 Days

Of 698 enrolled physicians, 480 (40.0% female) cared for a patient with a qualifying alert. The intention to treat analysis included 5939 patients: 3841 in the intervention group and 2098 in usual care.¹

MAKE30 occurred in 50.8% of intervention patients and 42.9% of usual care patients (adjusted OR, 1.05; 95% CI, 0.82–1.35). No secondary outcome differed significantly between groups.¹

The groups were not balanced at baseline. Intervention patients were older and had more 

, congestive heart failure, diabetes, hypercholesterolemia, and hypertension. Usual care patients had more anemia, malignancy, and trauma (all P <.001).¹

Investigators cited reliance on asynchronous email over real-time phone delivery, a high proportion of non-actionable alerts, and the baseline comorbidity imbalances as key limitations.¹

Sandra Kane-Gill, PharmD, MS

Articles