The Attributable Mortality of Delirium in Critically Ill Patients
The Attributable Mortality of Delirium in Critically Ill Patients
We prospectively evaluated consecutive adults admitted for at least 24 hours to the 32 mixed bed intensive care unit of the University Medical Centre Utrecht, the Netherlands, between January 2011 and June 2013. We excluded patients with acute or premorbid neurological disease at baseline, those in whom assessments of delirium could not be performed owing to a language barrier, and those transferred from or to another intensive care unit. The local ethical review board gave approval for an opt-out consent method (institutional review board number 10-056/12-421) whereby participants and family members were notified of the study by a brochure that was provided at admission to the intensive care unit with an attached opt-out card.
A research team dedicated to this study used a validated flowchart to classify the mental status of patients daily until discharge from intensive care. All relevant information was available to the study team, including the 12 hourly confusion assessment method for the intensive care unit conducted by nurses. We categorised patients as comatose, sedated, awake and delirious, or awake and non-delirious. Firstly, we assessed the level of consciousness using the Richmond agitation-sedation scale. Patients with maximum scores of -5 or -4 during the entire 24 hour observation period could not be assessed for delirium and were classified as either comatose or sedated. A sedated state was defined as propofol continuously administered at a rate of >1 mg/kg/h and/or midazolam at a dose of >50 mg/d or equivalent either at the time of assessment or at any time in the 48 hours before assessment. All other patients with scores of -5 or -4 on the Richmond agitation-sedation scale were classified as comatose. We assessed the remaining patients for delirium using the confusion assessment method for the intensive care unit as well as inspection of medical notes and nursing charts by the research team. These patients were classified as delirious when they tested positive on the confusion assessment method in the intensive care unit and/or when there was a description of fluctuation in the level of consciousness, agitation, disorientation, or hallucinations. Furthermore, because haloperidol and quetiapine were exclusively used for the treatment of delirium during the study period, we also classified patients as delirious on the day of initiation of either of these drugs. In case of doubt, a neurologist (AS) was consulted, who cast the decisive vote for classification of mental status. This procedure had a sensitivity of 0.75 (95% confidence interval 0.47 to 0.92), specificity of 0.85 (95% confidence interval 0.68 to 0.94), and an excellent inter-rater agreement (Fleiss’ K 0.94) (unpublished data). To enable our primary analysis we dichotomised the mental status by reclassifying sedated patients as non-delirious and comatose patients (without sedation) as delirious (see Supplementary Figure 1). The clinical team responsible for the patients was unaware of the results of the delirium assessments made by the study team.
(Enlarge Image)
Supplementary Figure 1.
Reclassification of sedation and coma days in the various sensitivity analyses
In all multivariable models we adjusted for covariables that we chose a priori based on their expected associations with delirium and mortality after careful consideration of the literature. These covariables included age, sex, history of dementia, history of alcohol misuse, Charlson comorbidity index, acute physiology and chronic health evaluation IV score, admission type, readmission status, and presence of sepsis on admission to the intensive care unit. These are all time fixed variables, representing the risk of delirium at baseline. However, because the risk of delirium onset is likely to vary over the course of admission to an intensive care unit depending on the evolution of disease severity, we also incorporated time dependent variables in our primary analysis (Fig 1). These included daily measurements of the sequential organ failure assessment score, sepsis status, core temperature, mechanical ventilation status, use of sedative and analgesic drugs, and plasma sodium, urea, acidosis, and haematocrit levels. Several physiological and laboratory variables (temperature, sodium, urea, and haemoglobin) were transformed to account for their non-linear relation with delirium or mortality, using the cut-offs from the acute physiology and chronic health evaluation IV model. Observers dedicated to this study collected data prospectively as part of a large cohort study and regularly checked for data integrity.
(Enlarge Image)
Figure 1.
Evolution of disease severity before onset of delirium in two hypothetical patients admitted to the intensive care unit. Both patients have similar severity of disease, but the condition of patient A worsens, whereas that of patient B improves. As delirium preferentially develops in more severely ill patients, confounding occurs when disease severity after baseline is not adjusted for in the analysis. Logistic regression and survival analysis adjusts for baseline variables at t=0 only. A marginal structural model adjusts for changes in disease severity until the onset of delirium (area to left of arrow), but not thereafter (area to right of arrow)
No data were missing for baseline variables, daily mental status classifications, or the outcome. However, for daily observations of several laboratory and physiological variables, 3.1% of data were missing overall (range 0-6.9% for individual variables). Because of the availability of longitudinal data before each observation day for each patient, we performed a trend imputation for missing covariables. Mortality during admission to an intensive care unit was the primary outcome of interest in all analyses. Although long term sequelae of delirium are known to exist, in the present study we deliberately focused on short term outcomes because previous authors have claimed up to a threefold increased mortality rate in the intensive care unit even after correction for confounders. Furthermore, adjustment for time dependent covariables using marginal structural models requires daily information about severity of disease and therapeutic interventions for the duration of observation. Practically, this precludes the use of outcomes that lie beyond discharge from the intensive care unit.
To obtain first estimates of the association between delirium and mortality in our cohort, and to be able to compare our results with previous literature, we performed a multivariable logistic regression analysis, adjusting for a priori selected baseline confounders. To comply with existing literature, we assumed that patients who develop delirium are at increased risk for the duration of their stay in the intensive care unit, even if delirium develops only several days after admission. The resulting bias can be overcome by using a Cox proportional hazards analysis and with inclusion of delirium as a time dependent variable. In this type of analysis, however, informative censoring of the survival time should additionally be taken into account by considering discharge as a competing risk for mortality, because patients who are discharged from the intensive care unit alive are in a different health state from patients who remain admitted beyond that time point. A competing risks analysis provides two measures of association: the cause specific hazard ratio, which in this case estimates the direct effects of delirium on outcome (both intensive care unit discharge and death), and the subdistribution hazard ratio, which describes the instantaneous risk of dying from delirium given that the patient has not died from delirium. The subdistribution hazard ratio is therefore a summary measure of all separate cause specific hazards and can be used to calculate the cumulative incidence of the outcome of interest (that is, death in this study). Although the methods can adjust for baseline confounders, the time varying nature of delirium onset and informative censoring caused by discharge from the intensive care unit, a limitation of these methods is that neither can adjust for other, potentially important sources of confounding. Firstly, the severity of disease on the day of admission to the intensive care unit may not be representative of the health state at the time of delirium onset, which typically occurs later on during stay in the intensive care unit (Fig 1). As delirium preferentially develops in patients who are more severely ill, bias occurs when such changes in disease severity are not adjusted for during the analysis. Secondly, bias might occur when a time dependent covariable is not only a risk factor for death but also predicts subsequent delirium, and when delirium status at a previous time point predicts the risk factor. For instance, severely agitated patients with delirium may eventually be treated with sedatives, whereas sedative use itself is a known risk factor for delirium. A marginal structural model analysis deals with these limitations by adjusting for the changes in disease severity before delirium onset, while preventing bias. It enables assessment of what the mortality in the intensive care unit would have been in a hypothetical population in which all patients remained delirium-free, and is therefore called a counterfactual analysis.
To accomplish such a counterfactual analysis, we performed two steps. Firstly, we modelled the daily probability of acquiring delirium in the intensive care unit, using a multivariable logistic regression analysis that included both baseline and daily patient characteristics. Based on these estimated daily probabilities, we calculated stabilised patient specific weights (so called inversed probability weights) that represent the cumulative risk of acquiring delirium for each patient. Because adjustment for time varying variables measured after the start of delirium may result in bias, we used lagged values from the preceding day to predict delirium on each day. We adjusted for lagged values of the sequential organ failure assessment scores two days before to acknowledge that the scores measured within 24 hours before the onset of delirium may have been influenced by an insidious onset of delirium. Secondly, we performed an inverse probability weighted Cox regression analysis with competing endpoints (death and discharge alive) and estimated both the daily hazard and cumulative risk of death. To aid in the interpretation of the results, we computed the population attributable fraction, which indicates the percentage of patients who have died from delirium.
We performed several post hoc sensitivity analyses (see Supplementary Figure 1). Firstly, instead of categorising sedated patients as non-delirious, we reclassified these patients based on the first available valid assessment for delirium after the cessation of sedation, using backward imputation. Secondly, we applied a more rigorous definition of delirium by considering patients as being delirious only when they had been classified as delirious on at least two consecutive days. Thirdly, to assess possible effect modification by the underlying condition, we performed subgroup analyses in patients with sepsis only, and stratified by acute physiology and chronic health evaluation IV score.
All analyses were performed using SAS 9.2 (Cary, NC) and R 2.14 (www.r-project.org). We used the R-package “IPW” for the marginal structural model analysis. P values less than 0.05 were considered to be statistically significant. We used robust estimators (Huber sandwich) to calculate confidence intervals for the hazard ratios resulting from the marginal structural model analyses, and we used bootstrapping to estimate the confidence intervals for the attributable mortality.
Methods
Study Population
We prospectively evaluated consecutive adults admitted for at least 24 hours to the 32 mixed bed intensive care unit of the University Medical Centre Utrecht, the Netherlands, between January 2011 and June 2013. We excluded patients with acute or premorbid neurological disease at baseline, those in whom assessments of delirium could not be performed owing to a language barrier, and those transferred from or to another intensive care unit. The local ethical review board gave approval for an opt-out consent method (institutional review board number 10-056/12-421) whereby participants and family members were notified of the study by a brochure that was provided at admission to the intensive care unit with an attached opt-out card.
Delirium
A research team dedicated to this study used a validated flowchart to classify the mental status of patients daily until discharge from intensive care. All relevant information was available to the study team, including the 12 hourly confusion assessment method for the intensive care unit conducted by nurses. We categorised patients as comatose, sedated, awake and delirious, or awake and non-delirious. Firstly, we assessed the level of consciousness using the Richmond agitation-sedation scale. Patients with maximum scores of -5 or -4 during the entire 24 hour observation period could not be assessed for delirium and were classified as either comatose or sedated. A sedated state was defined as propofol continuously administered at a rate of >1 mg/kg/h and/or midazolam at a dose of >50 mg/d or equivalent either at the time of assessment or at any time in the 48 hours before assessment. All other patients with scores of -5 or -4 on the Richmond agitation-sedation scale were classified as comatose. We assessed the remaining patients for delirium using the confusion assessment method for the intensive care unit as well as inspection of medical notes and nursing charts by the research team. These patients were classified as delirious when they tested positive on the confusion assessment method in the intensive care unit and/or when there was a description of fluctuation in the level of consciousness, agitation, disorientation, or hallucinations. Furthermore, because haloperidol and quetiapine were exclusively used for the treatment of delirium during the study period, we also classified patients as delirious on the day of initiation of either of these drugs. In case of doubt, a neurologist (AS) was consulted, who cast the decisive vote for classification of mental status. This procedure had a sensitivity of 0.75 (95% confidence interval 0.47 to 0.92), specificity of 0.85 (95% confidence interval 0.68 to 0.94), and an excellent inter-rater agreement (Fleiss’ K 0.94) (unpublished data). To enable our primary analysis we dichotomised the mental status by reclassifying sedated patients as non-delirious and comatose patients (without sedation) as delirious (see Supplementary Figure 1). The clinical team responsible for the patients was unaware of the results of the delirium assessments made by the study team.
(Enlarge Image)
Supplementary Figure 1.
Reclassification of sedation and coma days in the various sensitivity analyses
Covariables and Outcome
In all multivariable models we adjusted for covariables that we chose a priori based on their expected associations with delirium and mortality after careful consideration of the literature. These covariables included age, sex, history of dementia, history of alcohol misuse, Charlson comorbidity index, acute physiology and chronic health evaluation IV score, admission type, readmission status, and presence of sepsis on admission to the intensive care unit. These are all time fixed variables, representing the risk of delirium at baseline. However, because the risk of delirium onset is likely to vary over the course of admission to an intensive care unit depending on the evolution of disease severity, we also incorporated time dependent variables in our primary analysis (Fig 1). These included daily measurements of the sequential organ failure assessment score, sepsis status, core temperature, mechanical ventilation status, use of sedative and analgesic drugs, and plasma sodium, urea, acidosis, and haematocrit levels. Several physiological and laboratory variables (temperature, sodium, urea, and haemoglobin) were transformed to account for their non-linear relation with delirium or mortality, using the cut-offs from the acute physiology and chronic health evaluation IV model. Observers dedicated to this study collected data prospectively as part of a large cohort study and regularly checked for data integrity.
(Enlarge Image)
Figure 1.
Evolution of disease severity before onset of delirium in two hypothetical patients admitted to the intensive care unit. Both patients have similar severity of disease, but the condition of patient A worsens, whereas that of patient B improves. As delirium preferentially develops in more severely ill patients, confounding occurs when disease severity after baseline is not adjusted for in the analysis. Logistic regression and survival analysis adjusts for baseline variables at t=0 only. A marginal structural model adjusts for changes in disease severity until the onset of delirium (area to left of arrow), but not thereafter (area to right of arrow)
No data were missing for baseline variables, daily mental status classifications, or the outcome. However, for daily observations of several laboratory and physiological variables, 3.1% of data were missing overall (range 0-6.9% for individual variables). Because of the availability of longitudinal data before each observation day for each patient, we performed a trend imputation for missing covariables. Mortality during admission to an intensive care unit was the primary outcome of interest in all analyses. Although long term sequelae of delirium are known to exist, in the present study we deliberately focused on short term outcomes because previous authors have claimed up to a threefold increased mortality rate in the intensive care unit even after correction for confounders. Furthermore, adjustment for time dependent covariables using marginal structural models requires daily information about severity of disease and therapeutic interventions for the duration of observation. Practically, this precludes the use of outcomes that lie beyond discharge from the intensive care unit.
Statistical Analysis
To obtain first estimates of the association between delirium and mortality in our cohort, and to be able to compare our results with previous literature, we performed a multivariable logistic regression analysis, adjusting for a priori selected baseline confounders. To comply with existing literature, we assumed that patients who develop delirium are at increased risk for the duration of their stay in the intensive care unit, even if delirium develops only several days after admission. The resulting bias can be overcome by using a Cox proportional hazards analysis and with inclusion of delirium as a time dependent variable. In this type of analysis, however, informative censoring of the survival time should additionally be taken into account by considering discharge as a competing risk for mortality, because patients who are discharged from the intensive care unit alive are in a different health state from patients who remain admitted beyond that time point. A competing risks analysis provides two measures of association: the cause specific hazard ratio, which in this case estimates the direct effects of delirium on outcome (both intensive care unit discharge and death), and the subdistribution hazard ratio, which describes the instantaneous risk of dying from delirium given that the patient has not died from delirium. The subdistribution hazard ratio is therefore a summary measure of all separate cause specific hazards and can be used to calculate the cumulative incidence of the outcome of interest (that is, death in this study). Although the methods can adjust for baseline confounders, the time varying nature of delirium onset and informative censoring caused by discharge from the intensive care unit, a limitation of these methods is that neither can adjust for other, potentially important sources of confounding. Firstly, the severity of disease on the day of admission to the intensive care unit may not be representative of the health state at the time of delirium onset, which typically occurs later on during stay in the intensive care unit (Fig 1). As delirium preferentially develops in patients who are more severely ill, bias occurs when such changes in disease severity are not adjusted for during the analysis. Secondly, bias might occur when a time dependent covariable is not only a risk factor for death but also predicts subsequent delirium, and when delirium status at a previous time point predicts the risk factor. For instance, severely agitated patients with delirium may eventually be treated with sedatives, whereas sedative use itself is a known risk factor for delirium. A marginal structural model analysis deals with these limitations by adjusting for the changes in disease severity before delirium onset, while preventing bias. It enables assessment of what the mortality in the intensive care unit would have been in a hypothetical population in which all patients remained delirium-free, and is therefore called a counterfactual analysis.
To accomplish such a counterfactual analysis, we performed two steps. Firstly, we modelled the daily probability of acquiring delirium in the intensive care unit, using a multivariable logistic regression analysis that included both baseline and daily patient characteristics. Based on these estimated daily probabilities, we calculated stabilised patient specific weights (so called inversed probability weights) that represent the cumulative risk of acquiring delirium for each patient. Because adjustment for time varying variables measured after the start of delirium may result in bias, we used lagged values from the preceding day to predict delirium on each day. We adjusted for lagged values of the sequential organ failure assessment scores two days before to acknowledge that the scores measured within 24 hours before the onset of delirium may have been influenced by an insidious onset of delirium. Secondly, we performed an inverse probability weighted Cox regression analysis with competing endpoints (death and discharge alive) and estimated both the daily hazard and cumulative risk of death. To aid in the interpretation of the results, we computed the population attributable fraction, which indicates the percentage of patients who have died from delirium.
We performed several post hoc sensitivity analyses (see Supplementary Figure 1). Firstly, instead of categorising sedated patients as non-delirious, we reclassified these patients based on the first available valid assessment for delirium after the cessation of sedation, using backward imputation. Secondly, we applied a more rigorous definition of delirium by considering patients as being delirious only when they had been classified as delirious on at least two consecutive days. Thirdly, to assess possible effect modification by the underlying condition, we performed subgroup analyses in patients with sepsis only, and stratified by acute physiology and chronic health evaluation IV score.
All analyses were performed using SAS 9.2 (Cary, NC) and R 2.14 (www.r-project.org). We used the R-package “IPW” for the marginal structural model analysis. P values less than 0.05 were considered to be statistically significant. We used robust estimators (Huber sandwich) to calculate confidence intervals for the hazard ratios resulting from the marginal structural model analyses, and we used bootstrapping to estimate the confidence intervals for the attributable mortality.
Source...