Study population
After obtaining institutional review board approval, data were obtained from the High-Density Intensive Care (HiDenIC) database, which includes clinical variables on all patients admitted to the University of Pittsburgh, a tertiary care academic medical center, from July 2000–September 2008. The HiDenIC database includes data on adult patients admitted to one of eight ICUs (medical, cardiac, transplant, surgical, neurological, and trauma). Exclusion criteria were applied including: (1) history of hemodialysis or renal transplant, (2) baseline creatinine > 3.5 mg/dl, (3) liver transplant during the index hospitalization, (4) insufficient information to determine AKI status, and (5) unknown age (Fig. 1). We defined the young adult population as those individuals 16–25 years of age. The remaining cohort was stratified into 10-year age increments including: 26–35 years, 36–45 years, and 46–55 years.
Clinical variables
The risk factors included the analysis are significant predictors of AKI in previous studies [19, 20]. The potential risk factors include sex, race, reference creatinine, estimated glomerular filtration rate (eGFR) derived from the reference creatinine [21], comorbid conditions defined by ICD-9 codes (cardiac disease, CKD, diabetes, fluid overload, history of hypertension, malignancies), admission type (medical or surgical), and moderate anemia (defined by The World Health Organization [22]). Fluid balance was calculated by subtracting total intake from output divided by the admission weight (kg) × 100 in the first 24 h of ICU admission [23]. We defined fluid overload as a fluid balance > 5%. Severity of illness was evaluated with the Acute Physiology and Chronic Healthy Evaluation (APACHE) III score [24]. In the first 24 h of ICU admission, the need for vasopressors, mechanical ventilation as well as concern for sepsis (the ordering of blood cultures and antibiotics within 24 h of each other) was also included. Additionally, we evaluated exposure to potentially nephrotoxic medications within the first 24 h of ICU admission, including angiotensin-converting enzyme inhibitors, angiotensin II receptor blockers, vancomycin, aminoglycosides, antibiotics other than vancomycin or aminoglycosides (including piperacillin/tazobactam, cephalosporins, quinolones, macrolides, sulfonamides, and carbapenems), calcineurin inhibitors, nonsteroidal anti-inflammatory drug (NSAID) medications, acyclovir, mannitol, and phenytoin.
Outcomes
We defined AKI according to the Kidney Disease Improving Global Outcomes (KDIGO) criteria [4]. Any patient meeting the criteria for KDIGO stage 1 or more based on either serum creatinine or urine output during their ICU stay was deemed to have AKI. We defined the reference creatinine as the baseline creatinine when available (lowest value between the most recent hospital creatinine value up to 1 year prior the index hospital admission and the creatinine recorded in the first 24 h of hospital admission) or the lowest value between the creatinine recorded in the first 24 h of hospital admission, first 24 h of ICU admission, and (for patients without a history of CKD) the creatinine derived from the Modification of Diet in Renal Disease (MDRD) equation for creatinine using an eGFR of 75 ml/min/1.73 m2 [25, 26]. The reference creatinine was used to determine creatinine changes for defining AKI. We evaluated for each age strata rates of AKI, need for renal replacement therapy (RRT), recovery from RRT, ICU length of stay, hospital length of stay, ICU mortality, hospital mortality, 90-day mortality, and 1-year mortality.
Statistical analysis
Categorical variables were summarized as number and percentage, and continuous variables were summarized as median with interquartile range. Given the large number of patients in the study, statistical differences alone are unlikely to be meaningful. Therefore, we set 10% as a clinically meaningful difference between age groups. Age per 5 years was included as a risk factor for each age group to account for differences within the age groups. To determine the susceptibilities and exposures associated with AKI, multivariable logistic regression was performed whereby: (1) the cohort was stratified by age group (each of 10 years, starting from age 16–25) and (2) with age group as a main effect and accounting for interactions between age group and all other risk factors. Age-stratified models were built using the following steps: (1) adding each risk factor to age as a continuous variable and using the Wald statistic to determine their significance, (2) the individual size of all variables in step 1 was tested with the Wald statistic as they were added to a multivariable logistic regression model, (3) variables with p ≥ 0.05 were taken out of the model and a reduced model was fit, and (4) lastly to compare nested models in steps 2 and 3 the likelihood ratio test was used to determine a final model. For the interaction models, in order to find a main effects model, age group was used as a main effect and steps 1 through 4 were repeated. With age retained in the models regardless of significance level, all possible interactions were added one at a time and their significance was determined with the Wald statistic. STATA’s “roctab” function was used to assess the area under the receiver operating characteristic curve (AUC) for each age-stratified model. In addition, the “rocreg” function that uses bootstrap (1000 replications) for inference was also used to assess nonparametric ROC estimation under the presence of covariates. Model selection for ICU mortality, hospital mortality, mortality at 90 days after ICU admission, and mortality 1 year after ICU admission across age groups was done using the stepwise selection methodology described above to identify the best model for mortality prediction. Goodness of fit was assessed using Hosmer–Lemeshow [27]. Statistical analyses were performed using STATA software (version SE 14.0, StataCorp LP) and SAS 9.4 with statistical significance set at p < 0.05.