Using best subset regression to identify clinical characteristics and biomarkers associated with sepsis-associated acute kidney injury

Abstract

Sepsis-associated acute kidney injury (AKI) is a complex clinical disorder associated with inflammation, endothelial dysfunction, and dysregulated coagulation. With standard regression methods, collinearity among biomarkers may lead to the exclusion of important biological pathways in a single final model. Best subset regression is an analytic technique that identifies statistically equivalent models, allowing for more robust evaluation of correlated variables. Our objective was to identify common clinical characteristics and biomarkers associated with sepsis-associated AKI. We enrolled 453 septic adults within 24 h of intensive care unit admission. Using best subset regression, we evaluated for associations using a range of models consisting of 1−38 predictors (composed of clinical risk factors and plasma and urine biomarkers) with AKI as the outcome [defined as a serum creatinine (SCr) increase of ≥0.3 mg/dL within 48 h or ≥1.5× baseline SCr within 7 days]. Two hundred ninety-seven patients had AKI. Five-variable models were found to be of optimal complexity, as the best subset of five- and six-variable models were statistically equivalent. Within the subset of five-variable models, 46 permutations of predictors were noted to be statistically equivalent. The most common predictors in this subset included diabetes, baseline SCr, angiopoetin-2, IL-8, soluble tumor necrosis factor receptor-1, and urine neutrophil gelatinase-associated lipocalin. The models had a c-statistic of ∼0.70 (95% confidence interval: 0.65–0.75). In conclusion, using best subset regression, we identified common clinical characteristics and biomarkers associated with sepsis-associated AKI. These variables may be especially relevant in the pathogenesis of sepsis-associated AKI.