AKI Prediction Models: Room for Improvement

The power to predict an outcome with reliable certainty helps both providers and patients make better therapeutic choices. An area that is ripe for prediction modeling is acute kidney injury (AKI). I tend to group AKI prediction models into one of three categories: dynamic, static, and hybrid.

An actively researched dynamic model is the furosemide stress test. As eloquently described by Chawla and colleagues,^[1] administering a moderate dose of furosemide (1.0-1.5 mg/kg body weight) to patients with mild acute kidney injury (AKIN stage 1) can identify those who will progress to severe AKI (AKIN stage 3). Patients whose urine output falls below the 200 mL threshold in the subsequent 2 hours are predicted to worsen to AKIN stage 3. While the furosemide stress test needs further validation, a number of leading-edge nephrologists from around the world recommend its use, including those in Bolivia^[2,3] (see poster).

(Enlarge Image)

Fig. 1.

Courtesy of Rolando Claure-Del Granado, MD
Download Poster

The BioMaRK model operates on the other end of the AKI spectrum. It can predict the probability of dialysis freedom in AKI patients currently receiving renal replacement therapy (AKI-D). Serial measurements of urinary biomarkers can elucidate individuals who will become free of dialysis within the next 60 days. The BioMaRK model offers good discrimination when the urinary markers are followed over time (see infographic) but unfortunately has limited utility because of the lack of labs that can measure these markers.

(Enlarge Image)

Fig. 2.

Download Infographic

This bring us to the "Kaiser model," which offers a static tool to predict the likelihood of dialysis freedom in AKI-D patients within 90 days.^[4] The model has two characteristics that make it attractive: the use of readily measured predictor variables (eg, platelet count, eGFR, hemoglobin) and the need for measurements at a single point in time. Unfortunately, the discriminatory power of the model is not as robust as BioMaRK, with the best probability of dialysis independence near 50% (a coin toss). The validation study is also limited by the small number of patients included (just over 2200) despite the large Kaiser database.

The Takeaway

Prediction modeling remains a hot area of research, and given the limitations of existing models, the quest continues to devise a mathematical formula that can reliably predict AKI progression and/or recovery. From a technical standpoint, the best models are those whose predictor variables are easy to measure and require only one measurement at any point in time. From a clinical utility standpoint, however, the best models are those that meaningfully help both providers and patients make better-informed decisions. The work needed to develop such a model will require the combined efforts of modeling researchers, clinicians, and patients.

What models do you use to predict outcomes in acute kidney injury? Use the comments section to share your experiences with the nephrology community.

Follow Tejas P. Desai, MD, on Twitter: @nephondemand

Follow Rolando Claure-Del Granado, MD, on Twitter: @RClaure_nefro

Follow Medscape on Facebook, Twitter, Instagram, and YouTube