How well can we predict hospital acquired acute kidney injury? (AKI)

Summary

• AKI is an acute and usually reversible decline in renal function. AKI can be life threatening requiring immediate fluid and electrolyte attention.

• Investigators have long searched for an ability to predict early AKI

• A strong reliable predictive model to be able to risk stratify patients at risk remains elusive.

Background

AKI is typically defined as an acute increase in serum creatinine (SCr) or decrease in urine volume. The incidence of AKI is reported to be increasing to 1 out of 5 hospitalized adults worldwide. AKI is associated with a variety of etiological factors and results in significant morbidity, mortality, prolonged length of stay and increased hospitalization costs. Renal injury occurs over time with increases in SCr occurring as a late stage event after significant kidney function has been lost.

Investigators have long searched for the ability to predict early AKI in order to possibly enable effective early treatment to prevent AKI. This search has produced limited beneficial results, including a recent publication using an artificial intelligence deep learning model only able to predict with 56% accuracy1 .

Hodgson et al2 provided a systematic review of prognostic prediction models in general hospital patients (pts). Their aim was to review models from which conclusions can be drawn and decisions made regarding AKI predictors.

Predictive Performance

11 studies consisting of 474,478 pts were reviewed. The studies were difficult to evaluate due to wide variability in study design, definitions, markers evaluated, timeframes used and external validation techniques. The overall incidence of hospital acquired AKI was 7% and ranged from <1% to 28% based on the patient group being studied.

Studies tested between 11-48 predictors, with the final predictors including 3-29 parameters. Performance was assessed via discrimination (AUROC) which separates and ranks patients who experienced AKI vs those who did not. The performance of all models was modest with a median AUROC was 0.745 (range of 0.71-0.80).

Derived Value

• 8 studies

• 0.75 Mean AUROC

Internally Validated Value

• 7 studies

• 0.73 Mean AUROC

Externally Validated Value

• 5 studies

• 0.68 Mean AUROC

Conclusion

• Early Identification and management of AKI patients presents a significant challenge.

• A strong predictive model to risk stratify pts and aid in early/prompt recognition remains elusive.

References: 1. Tomasev N, Glorot X, Rae JW, et al. A clinically applicable approach to continuous prediction of future acute kidney injury. Nature 2019:572;116-119. 2. Hodgson LE, Sarnowski A, Roderick PJ, et al. systematic review of prognostic prediction models for acute kidney injury (AKI) in general hospital populations. BMJ Open 2017;7:e016591. Doi:10.1136/bmjopen-2017-016581

---