Antimicrobial Exposure and the Risk of Delirium in Critically Ill Patients

Jessica J. Grahl; Joanna L. Stollings; Shayan Rakhit; Anna K. Person; Li Wang; Jennifer L. Thompson; Pratik P. Pandharipande; E. Wesley Ely; Mayur B. Patel


Crit Care. 2018;22(337) 

In This Article


Study Design and Population

The Vanderbilt University Institutional Review Board approved the study protocol. Described in detail elsewhere,[8] the parent BRAIN-ICU Study included adults who received treatment in the medical or surgical ICU for respiratory failure or shock (or both) between March 2007 and May 2010. This parent cohort excluded patients if they met one or more of the following criteria: recent substantial critical illness requiring ICU admission, conditions that would make assessments for delirium unreliable (for example, deafness and blindness), conditions that would prohibit long-term follow-up owing to active substance abuse, psychotic disorder, homelessness, or residence 200 miles or more from the enrolling center, life expectancy less than 24 h, lack of informed consent, known or suspected severe neurologic disease such as from an anoxic or traumatic brain injury, and severe dementia. As previously reported, significant pre-existing cognitive impairment was excluded by using the combination of the Short Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE)[20] score of at least 3.3 and the Clinical Dementia Rating score of more than 2.0.[8,21,22] Pregnant patients, prisoners, and patients younger than 18 years of age were also excluded from the parent cohort.

For this current investigation, additional inclusion criteria consisted of BRAIN-ICU patients restricted to enrollment at a single hospital, Vanderbilt University Medical Center, with available medication administration records from time of enrollment to ICU discharge, and additional exclusion criteria consisted of those without medication records and those who died or withdrew within the first 48 h.

Risk Factors for Delirium

The medication-administration record was used to collect information from the time of enrollment and throughout the ICU stay. Antimicrobial exposure was categorized as a previous-day exposure in the first model and total days of delirium during antimicrobial exposure in the second model. Antimicrobials were divided into four non-mutually exclusive groups: beta-lactams (subclasses: penicillins, first- through third-generation cephalosporins, cefepime, and carbapenems), macrolides, fluoroquinolones, and other while in the ICU. Antimicrobial exposure was assessed on a daily basis throughout ICU stay, and antimicrobial-free days were included in the statistical analysis in order to account for differences in total antimicrobial exposure during the study period. To avoid model overfitting, we created an "other" antimicrobial class, which included vancomycin, antifungals, antivirals, metronidazole, aminoglycosides, linezolid, sulfamethoxazole-trimethoprim, antiretrovirals, clindamycin, rifampin, doxycycline, tigecycline, dapsone, pentamidine, isoniazid, and ethambutol.

All other covariates were chosen a priori on the basis of clinical judgment and previous research, owing to their expected associations with antimicrobial exposure (independent variable) and with delirium (dependent variable) and thus their potential to be confounders. Baseline covariates included age, mechanical ventilation status, presence of severe sepsis, ICU type, Charlson Comorbidity Index, and Short IQCODE.[8,20,23] Previous-day covariates consisted of delirium, total daily dose of analgesics and sedatives (benzodiazepine, propofol, opiate, and dexmedetomidine), and antipsychotic use (typical antipsychotic [for example, haloperidol] and atypical antipsychotic [for example, quetiapine and olanzapine]). Analgesic and sedative drug doses were transformed with the use of their cube root to reduce the influence of extreme outliers. Same-day covariates consisted of modified Sequential Organ Failure Assessment (SOFA) score.[24] As previously done, the modified SOFA score was used and excluded the neurological components since coma was accounted for in all statistical models to determine primary outcome.[8,23] This daily modified SOFA score was incorporated into statistical analyses in order to account for severity of illness, including degree of renal dysfunction, which is a risk factor for developing central nervous system toxicity and delirium in patients.


Trained research personnel evaluated patients for delirium and level of consciousness daily until hospital discharge. Delirium was assessed with the utilization of the Confusion Assessment Method for the ICU (CAM-ICU) twice daily in the ICU and once daily on the ward.[25,26] As previously described and in conjunction with structured evaluations for agitation and sedation,[27] mental status on each day was classified as normal, delirious, or comatose.[8,21–23,28] Coma was defined as a Richmond Agitation Sedation Scale score of −4 or −5. The day of discharge was assigned a mental status of normal. All records prior to death or discharge were included for the longitudinal analysis.

Statistical Analysis

To test the associations between antimicrobial use and delirium, we used a multivariable adjusted regression model. To account for dependency between repeated measures[29] and longitudinal change in any mental status to delirium (that is, delirium versus normal/coma) in our multivariable adjusted model, we applied a logistic regression with cluster sandwich covariance estimator. Our second main model employed a proportional odds logistic regression model to explore the outcome of total days of delirium for each patient included in the cohort study as a function of total antimicrobial exposure.

Continuous variables were summarized as median with interquartile range (IQR). Categorical variables were summarized as a number with percentage (n, %). Multivariable analyses were reported with odds ratios (ORs) and 95% confidence intervals (CIs). Data and associated electronic materials were stored in a password-protected Research Electronic Data Capture (REDCap) database.[30] All analysis was performed by using statistical software R version 3.3.0 (R Development Core Team Vienna, Austria;