Acute Kidney Injury, Mortality, Length of Stay, and Costs in Hospitalized Patients

Study Setting

The study was conducted at Brigham and Women’s Hospital, a 720-bed urban academic medical center in Boston, MA. Data were obtained for a study to examine the effects of a computer-order entry-based decision tool on drug prescribing for hospitalized patients with impaired kidney function (3). As part of the data library collected for evaluation of the appropriateness of drug prescription, serial SCr determinations were collected on a consecutive series of hospitalized patients on the medical, surgical (including subspecialty surgical services), neurology, and obstetrics and gynecology services between September 1997 and April 1998.

There were 19,982 admissions in which at least one SCr was obtained. For 9210 (46%) admissions, SCr was determined two or more times. Five (0.05%) admissions were excluded because vital status was unknown at discharge. The number of SCr determinations ranged from two to 92.

Description of Data

Routinely available demographic factors, specific drug prescriptions, Center for Medicare and Medicaid Services diagnosis-related group (DRG) weights (a proxy for disease severity [4]), LOS, hospital charges, itemized costs, and vital status at hospital discharge were also collected. For the purpose of these analyses, we additionally obtained primary International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes for each hospitalization. We categorized the ICD-9-CM codes according to conventional methods. For risk adjustment, we grouped diagnoses into “diseases of the circulatory system,” “diseases of the respiratory system,” “diseases of the gastrointestinal system,” “infectious diseases,” “neoplasms,” and all others, with the last group as the referent category. We considered people whose gender-stratified admission SCr was above the 90th percentile as “chronic kidney disease (CKD),” except when the admission ICD-9-CM code was 580, 584, or 788, potentially indicating community-acquired AKI. The CKD designation included men with SCr ≥ 1.6 mg/dl and women with SCr ≥ 1.4 mg/dl. We estimated baseline creatinine clearance using the Cockcroft-Gault equation, incorporating SCr, age, gender, and body weight (5).

To help determine the presence and severity of AKI, we calculated the difference between peak and baseline SCr concentrations. For individuals with only two SCr determinations, the second, when higher than the first, was considered the peak. We assumed that individuals with no further monitoring or change in SCr had not experienced AKI. For individuals with more than two SCr determinations, we considered the minimum of the first three as baseline and the maximum of the second through nth as peak. We considered nominal changes in SCr, percentage changes, percentage changes to a minimum peak, and other previously used algorithms that incorporated the exponential relation between SCr and GFR. Finally, we broadly classified individuals by a combination of ICD-9-CM diagnosis codes and prescribed medications as potentially associated with reduced renal perfusion (“prerenal”) versus intrinsic renal injury.

Statistical Analyses

Continuous variables are described as mean ± SD or median with interquartile range and compared with t test or the Wilcoxon rank sum test, when appropriate. Categorical variables are described as proportions and compared with the χ² test. We used logistic regression to estimate the odds of death with AKI, adjusting for associations with age, gender, DRG weight, ICD-9-CM group, and CKD. Because death was relatively common among patients with large increases in SCr, risk ratios were estimated using the method of Zhang and Yu (6). Model discrimination was determined using the area under the receiver operating characteristic curve (7). Model fit was assessed with the Hosmer-Lemeshow goodness-of-fit test (8). We used linear regression analysis to evaluate the associations of AKI with hospital LOS and costs. We log-transformed LOS and costs to accommodate data that were expectedly right-skewed and used Mallow’s Cp to assess model fit. Residual diagnostics indicated few outliers. We tested multiplicative interaction terms of AKI with other model covariates to evaluate for effect modification. To determine whether the risk associated with AKI differed by presumed cause (prerenal versus other), we used the Breslow-Day test for homogeneity of odds ratios (OR) and added an AKI × cause interaction term to multivariable models. Given that we were exploring multiple definitions of AKI and three outcome variables, we considered two-tailed P < 0.01 to be statistically significant. Statistical analyses were conducted using SAS 8.2 (SAS Institute, Cary, NC).

Mortality

Table 2 shows the unadjusted, age- and gender-adjusted, and multivariable-adjusted OR associated with AKI when AKI was determined by changes in SCr. Older patients and patients with CKD, higher disease severity, and admission diagnoses in the infectious disease, respiratory, gastrointestinal, and malignancy ICD-9-CM categories had an increased risk for death.

For the model that used a change in SCr of ≥0.5 mg/dl, the risks for death were higher for women than men (P = 0.004 for interaction) and for patients with de novo AKI compared with AKI superimposed on CKD (P = 0.001 for interaction). In models that used percentage change in SCr concentration (e.g., ≥50% increase in SCr), these interactions were not statistically significant. Discrimination and calibration were similar for models that used nominal and percentage changes in SCr (Table 2).

The OR associated with AKI was independent of the proxy for prerenal versus other causes (Breslow-Day χ², P = 0.56). The OR associated with AKI was significantly increased across all major ICD-9-CM diagnosis categories. Compared with respiratory, infectious, gastrointestinal, malignant, and other diseases, the adjusted OR associated with AKI was attenuated in patients who were admitted with cardiovascular disease (OR 4.3; 95% confidence interval [CI], 2.7 to 6.8), although the AKI × cardiovascular disease interaction was of marginal statistical significance (P = 0.02). There were no other significant interactions by ICD-9-CM code. LOS (another proxy for severity) was not significantly associated with mortality when added to these models. Forced inclusion of LOS augmented the OR estimates for AKI.

We performed additional analyses to compare different incremental changes in SCr. Figure 2 shows the unadjusted, age- and gender-adjusted, and multivariable-adjusted OR associated with increases in SCr of 0.3 to 0.4, 0.5 to 0.9, 1.0 to 1.9, and ≥2.0 mg/dl, relative to patients with little or no change in SCr. It is noteworthy that even very small increases in SCr (0.3 to 0.4 mg/dl) were significantly associated with mortality (multivariable OR 1.7; 95% CI, 1.2 to 2.6).

We evaluated the risk for AKI in subcohorts of patients with three, four, and five or more SCr determinations (cohorts with corresponding overall mortality rates of 3.7, 4.3, and 4.7%, respectively). Compared with an adjusted OR of 6.5 (95% CI, 5.0 to 8.5) in the full cohort, corresponding OR for AKI were 6.4 (95% CI, 4.9 to 8.5), 6.5 (95% CI, 4.9 to 8.7), and 6.5 (95% CI, 4.8 to 8.9) in subcohorts of patients with three or more, four or more, and five or more SCr determinations.

Finally, we conducted analyses that considered changes in SCr over the initial 7 ± 2 d of hospitalization. This method of AKI classification decreased the number of patients with AKI (n = 1098 [12%], n = 242 [3%], and n = 45 [0.5%] for increases in SCr of ≥0.5, ≥1.0, and ≥2.0 mg/dl, respectively). The corresponding multivariable OR for death associated with AKI were 4.3 (95% CI, 3.3 to 5.6), 5.9 (95% CI, 4.1 to 8.5), and 5.9 (95% CI, 2.8 to 12.4).

LOS

DRG weight was the strongest predictor of hospital LOS. Nevertheless, AKI was consistently associated with an independent increase in LOS. Using the ≥0.5 mg/dl increase in SCr criterion, we observed a 3.5-d increase in hospital LOS (P < 0.0001). In LOS models, there were no interactions among AKI, gender, and CKD. The models explained approximately 33% of the variance in LOS and were well fit. Larger increases in SCr were associated with longer relative increases in hospital LOS (e.g., 5.4 and 7.9 d with ≥1.0- and ≥2.0-mg/dl increases in SCr, respectively). In addition to AKI, DRG weight, CKD, and gastrointestinal and malignant diseases were associated with longer LOS. When we restricted the analysis to patients who survived, the relative increases in LOS were magnified (3.6, 5.8, and 9.0 d, for increases in SCr ≥0.5, ≥1.0, and ≥2.0 mg/dl, respectively). When we considered only those individuals with AKI during the 7 ± 2 d of hospitalization, the corresponding relative increases in LOS were 2.1, 2.8, and 4.6 d.

Costs

Table 3 shows unadjusted and multivariable-adjusted mean costs for hospitalization with and without AKI using corresponding changes in SCr. As with LOS, the dominant predictor of total hospital costs was DRG weight. Multivariable models explained >40% of the variation in total hospital costs. Figure 3 shows the unadjusted, age- and gender-adjusted, and multivariable-adjusted costs associated with incremental changes in SCr concentration. The cost results were robust, even with further adjustment for LOS. When LOS was included, the models explained >65% of the variation in costs.