Prediction of Acute Kidney Injury (AKI) After Surgery: A Pragmatic Three-Arm Cluster-Randomized Trial

Summary

This investigator-initiated, pragmatic trial evaluates whether displaying a machine learning (ML)- derived perioperative AKI risk score-alone or paired with an interruptive Best/Our Practice Advisory (BPA/OPA)-improves kidney-protective care and reduces kidney injury after non-obstetric surgery at UCSF. Approximately 75-100 attending anesthesiologists (clusters) are randomized 1:1:1 to: (a) Control (risk score hidden), (b) Score Only (visible preoperative AKI risk probability with passive KDIGO bundle recommendation), or (c) Score + BPA (visible risk plus interruptive KDIGO prompt for high-risk patients). CRNAs/residents follow their attending' s assignment. Adult inpatients (age ≥18) with expected overnight stay and eGFR ≥15 mL/min/1.73 m² are included; obstetrics, chronic dialysis, and kidney transplant patients are excluded. The underlying preoperative model was prospectively validated at UCSF and outperforms anesthesiologist risk estimation reported in the literature. The model was reviewed and approved by the AI Oversight Committee at UCSF. Primary endpoint is the continuous change in serum creatinine (mg/dL) from baseline to POD 1-2. Secondary outcomes include KDIGO-defined AKI, adherence to bundle elements (hemodynamics, balanced fluids, nephrotoxin avoidance, glycemic control), intraoperative hypotension time, fluid volumes, nephrotoxin exposure, perioperative hyperglycemia, length of stay, unplanned ICU transfer, readmission, dialysis, and in-hospital mortality. Data are obtained from the EHR; analysts are blinded. No direct subject interaction is planned; the investigators will request a waiver of patient consent. The study aims to demonstrate that ML-enabled, workflow-embedded decision support can safely and feasibly improve guideline concordant care and decrease early postoperative kidney injury.

Eligibility

Age range

18+ years

Sex

All

Healthy volunteers

Yes