Predicting Nephrotoxic Acute Kidney Injury in Hospitalized Adults
A GRU-based recurrent neural network layered onto NINJA criteria halved false nephrotoxic AKI alerts in adult inpatients while preserving sensitivity.
Key Findings
- Alert burden reduction. False alerts per AKI case dropped from 2.5 to 0.7 while positive predictive value climbed to 60%.
- Meaningful risk signals. Hemoglobin nadir, blood pressure minima, and leukocytosis emerged as high-impact predictors.
- Targeted interventions. Combining nephrotoxin exposure with physiologic data enables proactive stewardship and nephrology consults.
Introduction
Adult hospitals struggle to adapt the pediatric nephrotoxic injury negated by just-in-time action (NINJA) alert because it triggers too many false positives. Incorporating granular clinical features promised to focus attention on patients truly at risk of AKI within 48 hours of high nephrotoxin exposure.
Methods
This single-centre retrospective study analysed 14,480 adults (18,180 admissions) at the University of Iowa Hospitals (2017–2022). High-nephrotoxin exposures were defined per NINJA rules. Eighty-five demographics, vital signs, laboratory results, and medication features fed a GRU recurrent neural network trained on 85% of events and validated on 15%. A secondary feed-forward network provided feature attribution.
Results
In the test cohort 29% of exposures progressed to AKI. The GRU achieved a precision of 0.60, raising the positive predictive value relative to the rule-based alert and cutting false alerts per AKI case from 2.5 to 0.7. Lowest haemoglobin, hypotension, leukocytosis, and exposure to acyclovir, piperacillin-tazobactam, calcineurin inhibitors, or ACEi/ARB were the dominant predictors.
Discussion
By contextualising nephrotoxin exposure with clinical state, the model preserved sensitivity while delivering a manageable alert workload. Generalisability will require multicentre validation and thoughtful integration into stewardship rounds.
Clinical Implications
Targeted, accurate alerts free pharmacists and physicians to focus on patients most likely to benefit from medication adjustments, potentially preventing progression to severe AKI and dialysis.
Conclusion
Machine learning meaningfully enhances nephrotoxin surveillance in adults by embedding physiologic context into NINJA criteria.
Future Directions
Expand to multi-hospital datasets, integrate drug dosing and serum levels, and trial prospective deployment to confirm patient outcome benefits.