A pair of clinical models developed using machine learning techniques perform well in predicting the risk of death and adverse kidney outcomes in critically ill patients with incident acute kidney injury (AKI), reports a pre-proof paper in the American Journal of Kidney Diseases.
The multicenter, retrospective cohort study included a derivation set of 7354 patients at one U.S. university medical center who were diagnosed with AKI (based on serum creatinine Kidney Disease: Improving Global Outcomes [KDIGO] criteria) within 3 days after intensive care unit (ICU) admission. Data on 71 validated clinical variables during the first 3 ICU days were extracted from electronic medical records. Four machine learning algorithms—logistic regression, random forest, support vector machine, and extreme gradient boost—were used to train the models for prediction of in-hospital death and major adverse kidney events (MAKEs). The latter outcome was a composite of death, renal replacement therapy, and 50% or greater reduction in estimated glomerular filtration rate from baseline to 120 days after discharge.
The developed clinical models included 15 features for prediction of mortality and 14 for prediction of MAKEs. Predictive performance was evaluated using tenfold cross-validation in the derivation cohort, followed by external validation of 2333 patients from a different center.
The 15-variable clinical model outperformed the Sequential Organ Failure Assessment score for prediction of mortality in both the derivation cohort (area under the curve [AUC], 0.79 vs. 0.71) and the validation cohort (AUC, 0.71 vs. 0.74). In the validation cohort, among patients classified as being at >50% predicted risk of mortality, 41% actually died.
The 14-variable model also improved prediction of MAKEs compared with the maximum AKI KDIGO score (AUC, 0.78 vs. 0.66 in the derivation cohort and 0.66 vs. 0.73 in the validation cohort). Among patients at 50% or higher risk, 24.5% developed a MAKE.
AKI occurs in up to 50% of patients admitted to the ICU. Although clinical models are useful in predicting AKI risk, there are few tools for prediction of AKI recovery or outcomes.
The newly developed models perform well in predicting in-hospital mortality and MAKEs in a heterogeneous population of ICU patients with AKI. “[I]f further validated, [the models] could enable risk stratification for timely interventions that promote kidney recovery,” the researchers conclude. They have developed an online tool for predicting outcomes in critically ill adults with incident AKI within the first 3 days of an ICU stay, available at http://phenomics.uky.edu/taki/ [Neyra JA, et al. Prediction of mortality and major adverse kidney events in critically ill patients with acute kidney injury. Am J Kidney Dis, published online ahead of print July 13, 2022. doi: 10.1053/j.ajkd.2022.06.004; https://www.ajkd.org/article/S0272-6386(22)00774-0/fulltext].