Study design and participants
This study was based on a prospectively enrolled cohort of patients admitted to the mixed ICU of the University Medical Center Utrecht the Netherlands between August 2011–June 2013 and June 2015–March 2019 [
20,
21]. In between these two periods no data have been collected, due to lack of available research staff. Details on the patients in the first cohort have been published [
22]. For the current analysis, exclusion criteria were ICU admission < 24 h, readmissions, transfers from another ICU, or admission with a primary acute neurological or neurosurgical disorder confounding the delirium diagnosis [
23], or another condition that could hamper the assessment of delirium, such as mental retardation, anoxic brain injury after cardiopulmonary resuscitation, or inability to speak Dutch or English. ICU physicians adhered to a sedation protocol aiming for light sedation (target Richmond Agitation and Sedation Scale [RASS] score − 2 to 0), without using dexmedetomidine. Because of the non-interventional nature of this investigation, the ethics committee of the University Medical Center Utrecht waived the need for informed consent (protocol #19-768/c). This cohort study was reported using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement [
24].
Data collection and definitions
All data have been collected prospectively. Age, sex, ICU admission type, Acute Physiology and Chronic Health Evaluation (APACHE) IV [
25], and ICU and hospital length of stay were stored in the electronic patient data management system. On a daily level (one day was defined as a 24-h epoch from 10AM to 09.59AM), the following characteristics were collected: modified Sequential Organ Failure Assessment (SOFA) score (without the central nervous system component) [
26], presence of metabolic acidosis (defined as base excess < − 3 in arterial blood gas analysis), mechanical ventilation status, mortality, and administration of antipsychotics (haloperidol, quetiapine, olanzapine, clozapine, risperidone), sedatives (benzodiazepines, propofol, clonidine) and opioids.
Patients were prospectively classified per 24-h epoch as awake without delirium, delirious, unarousable, discharged or dead. The daily mental status was classified according to a delirium recognition algorithm based on three items [
22]: (1) the RASS and Confusion Assessment Method-Intensive Care Unit (CAM-ICU) scores, assessed by bedside nurses, (2) evaluation of medical charts for signs of delirium i.e. altered consciousness and > 1 symptoms of delirium, and (3) additional delirium assessments by trained researchers. If in a 24-h period at least one of these items was positive for delirium, patients were classified as delirious. In this current cohort study, the initiation of haloperidol or quetiapine was not used to classify patients as delirious. Sedation was assessed every three hours with the RASS [
27]. On a daily level, an unarousable state was defined as all RASS scores on that day being < − 3 or − 3 combined with sedative use—impeding delirium assessment [
28]. Patients were classified as awake without delirium if they had at least one RASS score > − 4, or > − 3 if in combination with sedative use, in a 24-h period, with all delirium assessments in that period being negative. With regards to delirium during ICU stay, patients were classified as delirious if they had at least one day with a positive delirium assessment according to the before mentioned algorithm.
Statistical analysis
Continuous data were summarized with medians and interquartile ranges (IQR) or means with standard deviations, depending on distribution. Categorical variables were shown as proportions with percentages. To describe differences between patients who were ever delirious and patients who were never delirious, continuous variables were compared using Mann–Whitney U tests or independent t tests, and categorical outcomes with Chi square tests. For differences in baseline medication exposures, we used logistic and linear regression analyses.
For the primary analysis the delirious observation days were used. We studied whether the administration of haloperidol, clonidine, or both haloperidol and clonidine on any given delirium day (day
X) was associated with a transition to an awake state without delirium on the following day (day
X + 1), compared to the probability of either remaining delirious or transitioning to unarousable state on day
X + 1, with mortality and discharge on day
X + 1 being competing events (Online Resource 1). The combination of delirium and unarousable state on the following day was chosen as a comparator, as we specifically aimed to assess delirium resolution as transition to an awake state without delirium and to minimize confounding of the treatment to the delirium assessment due to sedative side-effects of haloperidol [
29] and clonidine [
30]. For the primary analysis we used a first-order Markov multinomial logistic regression model. Model 1 involved administration of haloperidol only (yes/no), clonidine only (yes/no), or both haloperidol and clonidine (yes/no) on day
X. Additionally, to examine dose-dependency, we modeled two main predictors (Model 2): (1) dosage of intravenously administered haloperidol in milligrams on day
X, with a conversion of 0.6 per milligram enterally administered haloperidol based on bioavailability, and (2) dosage of intravenously administered clonidine in micrograms on day
X. To control for confounding, we included the following covariates to both Markov models: age, APACHE IV score, and admission type (acute surgery, elective surgery, medicine). Additionally, we included the following daily covariates: patients’ mental status on day X, modified SOFA score, metabolic acidosis (yes/no), use of mechanical ventilation (yes/no), and administration of any antipsychotic other than haloperidol (yes/no), benzodiazepines or propofol (yes/no), and opioids (yes/no). These covariates were associated with delirium development, or were suspected to confound the studied association [
31].
Additionally, we used propensity score matching for the primary analysis to match delirium days without medication exposure with delirium days on which haloperidol, clonidine, or both was administered. The match tolerance was set to 0.001. We adjusted for confounding by including the same covariates as in the Markov models.
Three sensitivity analyses were performed. First, as both haloperidol and clonidine may have mild sedative effects [
29,
30], they may influence the CAM-ICU assessment. Likewise, agitated delirious patients may have received more haloperidol and/or clonidine than patients with hypoactive delirium. Hence, our first sensitivity analysis stratified outcomes by delirium motor subtypes, comparing “hyperactive delirium days” (all RASS scores on that delirium day > 0), “hypoactive delirium days” (all RASS scores on that delirium day ≤ 0) and “mixed delirium days” (RASS scores of both > 0 and ≤ 0 on that delirium day). Second, we assessed whether the implementation of the 2013 Pain, Agitation and Delirium guideline [
32] and general changes in ICU delirium management over the years affected the resolution of delirium by assessing the effect of the inclusion period (2011–2013 versus 2015–2019). Third, as the clinical effects of haloperidol and clonidine may become apparent 24 h after administration, we examined whether there was an association of the administration of haloperidol and/or clonidine on delirium day
X with transition to an awake state without delirium on day
X + 2, using the same analysis as for
X + 1.
Secondary analyses were performed on patient level in delirious patients instead of the level of observation-days to assess whether the administration of haloperidol, clonidine or both haloperidol and clonidine versus none of these agents during ICU stay was associated with the secondary outcomes. Linear regression models were used to analyze delirium duration, number of delirium and ventilation days, and ICU and hospital length of stay. ICU mortality was analyzed with a logistic regression model. These models were adjusted for age, APACHE IV score, admission type, highest modified SOFA score during ICU stay, presence of metabolic acidosis during ICU stay (yes/no), ventilation during ICU stay (yes/no), delirium duration (in days), and administration of any antipsychotic other than haloperidol (yes/no), benzodiazepines or propofol (yes/no) and opioids (yes/no) during ICU stay. If needed, log transformation of the variables was performed.
Analyses were performed in SPSS, version 25, and Markov multinomial logistic regression analyses were performed in R, version 3.5.1. To evaluate the goodness of fit, likelihood ratio tests, F values and R2 values were used. Observations were excluded if delirium classification was missing. A two-sided p value less than 0.05 was considered statistically significant.