The impact of delirium on withdrawal of life-sustaining treatment after intracerebral hemorrhage

Abstract

Objective

To determine the impact of delirium on withdrawal of life-sustaining treatment (WLST) after intracerebral hemorrhage (ICH) in the context of established predictors of poor outcome, using data from an institutional ICH registry.

Methods

We performed a single-center cohort study on consecutive patients with ICH admitted over 12 months. ICH features were prospectively adjudicated, and WLST and corresponding hospital day were recorded retrospectively. Patients were categorized using DSM-5 criteria as never delirious, ever delirious (either on admission or later during hospitalization), or persistently comatose. We determined the impact of delirium on WLST using Cox regression models adjusted for demographics and ICH predictors (including Glasgow Coma Scale score), then used logistic regression with receiver operating characteristic curve analysis to compare the accuracy of ICH score–based models with and without delirium category in predicting WLST.

Results

Of 311 patients (mean age 70.6 ± 15.6, median ICH score 1 [interquartile range 1–2]), 50% had delirium. WLST occurred in 26%, and median time to WLST was 1 day (0–6). WLST was more frequent in patients who developed delirium (adjusted hazard ratio 8.9 [95% confidence interval (CI) 2.1–37.6]), with high rates of WLST in both early (occurring ≤24 hours from admission) and later delirium groups. An ICH score-based model was strongly predictive of WLST (area under the curve [AUC] 0.902 [95% CI 0.863–0.941]), and the addition of delirium category further improved the model's accuracy (AUC 0.936 [95% CI 0.909–0.962], p = 0.004).

Conclusion

Delirium is associated with WLST after ICH regardless of when it occurs. Further study on the impact of delirium on clinician and surrogate decision-making is warranted.

Established clinical predictors of poor outcome after intracerebral hemorrhage (ICH)^1–3 may lead to a self-fulfilling prophecy, as they often factor into clinician decision-making and may serve as primary determinants of withdrawal of life-sustaining treatment (WLST) in many cases.⁴ Although the phenomenon of self-fulfilling prophecy in this patient population has been well-described,^5–8 there have been few studies examining the relative contribution of individual patient-level factors in WLST. Further, the association between WLST and composite prognostic models is unclear. Given the prominence of impaired consciousness in many prognostic scores, and the high prevalence of delirium after ICH,^9–12 we aimed to determine the impact of delirium on WLST after ICH in the context of other established predictors. We also tested the hypothesis that delirium is a robust predictor of WLST regardless of whether it is present at the time of hospital admission or subsequently develops over the course of a patient's hospitalization.

Methods

Study population

We performed a retrospective cohort study using prospectively collected data from the Brown ICH registry at Rhode Island Hospital. We included consecutive patients who were determined to have a spontaneous ICH by 2 attending neurocritical care or vascular neurologists over a 12-month period from February 16, 2018 (the registry's start date) to February 15, 2019. Patients diagnosed with ICH due to trauma, hemorrhagic conversion of a known ischemic stroke, or a known intracranial malignancy at the time of admission were excluded from the registry.

At our institution, the criteria for patients with ICH to be admitted to our Neurocritical Care Unit include the following: hematomas larger than 20 mL, presence of intraventricular hemorrhage (IVH), infratentorial location, impaired arousal, and anticoagulation use. All other patients with ICH are cared for in our intermediate care stroke unit.

Standard protocol approvals, registrations, and patient consents

The use of data for this study was approved by our hospital's Institutional Review Board, and the requirement for informed consent was waived.

Data collection

We prospectively collected all data related to standard clinical stroke care in a REDCap^13,14 database (Vanderbilt University, Nashville, TN) as part of an ongoing institutional quality improvement project. These data included patient demographics, comorbidities, neuroimaging, and other diagnostic testing. Two attending neurologists with board certification in neurocritical care or vascular neurology prospectively adjudicated ICH-related clinical predictors until consensus was achieved. These clinical predictors included hematoma location and size (measured via the ABC/2 method¹⁵ with modifications for irregularly shaped hematomas¹⁶), ICH score, and etiology. We diagnosed possible or probable cerebral amyloid angiopathy according to modified Boston criteria¹⁷ using clinical history and neuroimaging (with MRI available in 65% of patients).

Delirium diagnosis

Delirium was diagnosed by an attending neurologist with additional training in delirium according to DSM-5 criteria¹⁸: disturbances in attention and awareness (often accompanied by disturbances in other cognitive domains, such as psychomotor slowing or agitation, disorientation, disorganized thinking, impaired executive function, or perceptual disturbance) that develop over a short period of time and tend to fluctuate, represent a change in function, and are due to an underlying medical condition or toxic/withdrawal syndrome. In many cases, we determined diagnoses prospectively, either as part of a nested research study¹² or in the course of standard clinical care. In all other cases, we determined diagnoses retrospectively via detailed chart review using the same DSM-5 criteria, with established chart-based methods that have been previously validated.¹⁹ Common words and phrases used to aid in retrospective delirium diagnosis included the following when used to describe a patient: delirium, delirious, mental status change, altered mental status, fluctuating mental status, waxing and waning mental status, encephalopathy, encephalopathic, confusion, agitation, agitated, and sundowning. However, we also required that individual features conformed to DSM-based criteria, and we gave preference to cases with a clear underlying medical cause or documented risk factor. We further supplemented this method with Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) scores²⁰ documented as part of clinical care. However, the CAM-ICU was not used as a primary source of delirium diagnosis given its purpose as a screening rather than diagnostic instrument, and given that we have previously identified concerns about its accuracy in this patient population.¹²

We established inattention, the defining feature of delirium, as described previously,¹² using validated verbal-based methods such as backwards tasks when possible. However, we also considered a broad scope of manifestations of attention including patients' distractibility, ocular fixation and tracking of external stimuli, and ability to shift focus between multiple external stimuli, especially in the context of severe stroke-related deficits such as aphasia. While we considered these components in prospective assessments, such details were also frequently reported in clinical documentation by physicians, nurses, and physical and occupational therapists. In some cases, serial assessments were necessary to determine whether a patient's inattention was out of proportion to his or her expected ICH-related deficits, a distinction aided by evidence of fluctuations of attention over time.

In addition to determining whether patients were “never delirious” or “ever delirious,” we also determined whether patients had delirium that was present within 24 hours of hospital arrival (“early delirium”), or if they had delirium that occurred at some subsequent point during their hospitalization (“later delirium”). Patients who had no evidence of purposeful response to any stimulus (including command-following, vocalization, attention to external stimuli via ocular, head, or body movement, or localization) for the duration of their hospitalization were categorized as being persistently comatose.

Measurements and outcomes

We measured hospitalization time for each patient in calendar days, with the day of hospital admission counted as hospital day 1. We retrospectively abstracted code status documentation for each patient, including the hospital day on which a patient's code status was changed. We defined WLST as a change in code status to “comfort measures only” with a corresponding change in subsequent management (including terminal extubation, deferring life-prolonging measures, or transition to hospice care). Our primary outcomes were the hospital day on which WLST occurred and the binary classification of patients who had WLST as opposed to those who did not.

Statistical analysis

We used standard descriptive statistics to report patient characteristics, describing data that had a normal distribution with means and SDs, and non-normal data with medians and interquartile ranges (IQRs). We analyzed differences between continuous variables using t tests or the Mann-Whitney U test, as appropriate, and differences between categorical variables using the χ² test.

To account for differences in exposure time that might arise in cases of later as opposed to early delirium, we used survival statistics and adjusted Cox regression models reporting hazard ratios (HRs) and 95% confidence intervals (CIs) to compare WLST between patients with and without delirium. Exposure time included calendar days from hospital admission until WLST and patients who did not have WLST were censored on day of hospital discharge or death. In our base model, we included patient demographics (age, sex, and race classified as white vs nonwhite given previously described disparities in WLST^21,22) and established ICH predictors (hematoma size, infratentorial vs supratentorial location, presence of IVH, and initial Glasgow Coma Scale [GCS] score) as covariates, with age and hematoma size represented by continuous variables. In a second model, we included delirium category (never delirious, ever delirious, or persistently comatose) as an additional covariate. We also performed a post hoc sensitivity analysis using this model, in which we excluded patients with hematoma expansion or cerebral edema due to the likelihood that this may have contributed to WLST. In a third model, we also considered delirium that occurred later rather than early, which we included as an interaction variable due to concerns that the proportional hazards assumption might otherwise be violated.

Finally, to allow estimation of marginal risk differences and formal comparison of model discrimination, we performed logistic regression with postestimation procedures as follows: first, we determined predicted probabilities of WLST and marginal effects associated with delirium status when ICH-related predictors were held at mean values for the sample, then with a more clinically relevant hypothetical scenario with a 10 mL supratentorial ICH; then, we performed receiver operating characteristic (ROC) curve analysis with comparisons of area under the ROC curve (AUC) using the DeLong test.²³ We compared the predictive accuracy of a model comprising the individual components of the ICH score (including age and hematoma size as continuous variables) with a similar model that also included a covariate for delirium category, then performed further exploratory analyses comparing the 2 models in groups of patients stratified by GCS score (≥13 or <13). We also determined the predictive accuracy of the original ordinal ICH score to allow comparison with other models.

All hypothesis-testing was 2-sided, and we set the threshold for significance at α = 0.05. We performed our statistical analyses using Stata/MP 11 (College Station, TX).

Data availability

All relevant data are presented within the article and its supporting tables and figures. Additional information can be obtained upon reasonable request to the corresponding author.

Results

Baseline characteristics

We identified 311 patients with ICH during the 12-month study period, of whom 31% (n = 97) were assessed for delirium prospectively. Mean age was 70.6 years (SD 15.6), 50% were male, 83% were white, and 61% were transferred from another hospital. Median ICH score was 1 (IQR 1–2), and 50% (n = 157) of patients developed delirium at some point during their hospitalization (65% [n = 63] of patients with prospective assessments and 50% [n = 94] of noncomatose patients who were reviewed retrospectively). Delirium occurred early in 36% of patients (n = 111) and later in 15% (n = 46); 9% of patients (n = 27) remained persistently comatose without any clinical evidence of consciousness at any time during their hospitalization. Among those who were not persistently comatose, patients with delirium had higher ICH severity than those who did not have delirium (including larger hematoma volumes, more frequent hematoma expansion, and a higher frequency of IVH) and were more likely to have a history of dementia (table 1).

Table 1.

Baseline characteristics for noncomatose patients with intracerebral hemorrhage (ICH) comparing those who had delirium and those who did not

A total of 81 patients (26%) had a change in documented code status and subsequent care consistent with WLST, with a median time to WLST of 1 day (IQR 0–6). WLST occurred in 2% of patients who were never delirious (n = 2), 32% of those who had early delirium (n = 35), 41% of those who had later delirium (n = 19), and 93% of patients who were persistently comatose (n = 25) (2 additional patients died without WLST). Frequency of WLST was not significantly different between patients who had early vs later delirium (p = 0.31), and was also similar between patients with delirium who were assessed prospectively and those who were reviewed retrospectively (early delirium: 25% vs 37%, p = 0.19; later delirium: 38% vs 45%, p = 0.58).

Compared to patients who did not have WLST, patients with WLST were older and more likely to have ICH due to cerebral amyloid angiopathy, were more likely to be white, and had higher ICH severity (including higher ICH scores, larger hematoma volumes, and a higher frequency of IVH; table 2). These differences remained significant when considering only patients with delirium during their hospitalization, except for the distribution of patients by ICH etiology and the proportion of patients with IVH (the latter of which led to a somewhat less prominent difference in overall ICH score; table 2).

Table 2.

Baseline characteristics for patients with intracerebral hemorrhage (ICH) who had withdrawal of life-sustaining treatment (WLST) and those who did not, in the entire cohort and only patients with delirium during their hospitalization

Predictors of WLST

In our base Cox model, age, ICH volume, and especially initial GCS score were associated with WLST, while infratentorial location and IVH were not (table 3). WLST increased substantially with each ICH score point (figure 1A). In a second model that also included delirium category, the presence of delirium at any time during hospitalization had a marked association with WLST, while the effect of initial GCS score was substantially attenuated (table 3). We performed an additional post hoc sensitivity analysis using this second model, in which we excluded patients with hematoma expansion or cerebral edema due to the likelihood that this may have contributed to WLST, and found that the association between delirium and WLST remained significant (HR 4.7 [95% CI 1.03–21.6]). Finally, in an exploratory model that considered the interaction between delirium and early vs later delirium onset, we found no effect of delirium timing on subsequent WLST (p = 0.17; figure 1B).

Table 3.

Results of Cox regression analyses testing associations between relevant clinical factors and time to withdrawal of life-sustaining treatment after intracerebral hemorrhage (ICH)

Figure 1. Time to withdrawal of life-sustaining treatment.

Kaplan-Meier survival curves stratified by (A) intracerebral hemorrhage (ICH) score and (B) delirium category.

In our subsequent logistic regression and postestimation analyses, we found that the presence of delirium was associated with an adjusted marginal risk difference of 15% (95% CI 7%–23%) for having WLST when ICH-related clinical predictors were held at mean values for the sample, and 18% (95% CI 11%–26%) for a hypothetical 10 mL supratentorial ICH. When we considered delirium timing, we found that early delirium was associated with an adjusted marginal risk difference of 11% (95% CI 3%–18%) for having WLST when ICH-related clinical predictors were held at mean values for the sample, and 14% (95% CI 6%–21%) for a hypothetical 10 mL supratentorial ICH. For later delirium, these adjusted risk differences increased to 26% (95% CI 10%–42%) and 27% (95% CI 15%–40%), respectively.

In our ROC curve analyses, we found that a composite model of individual predictors from the ICH score was strongly predictive of WLST (AUC 0.902, 95% CI 0.863–0.941; figure 2A). However, despite this excellent predictive ability, the addition of delirium category resulted in further improvement in the model's accuracy (AUC 0.936, 95% CI 0.909–0.962, p = 0.004). Meanwhile, the ordinal ICH score as measured by point level also had very good accuracy in predicting WLST (AUC 0.852, 95% CI 0.805–0.898).

Figure 2. Predicting withdrawal of life-sustaining treatment (WLST) using established predictors of intracerebral hemorrhage (ICH) severity and delirium status.

Receiver operating characteristic (ROC) curves depicting accuracy in predicting WLST in patients with ICH for 3 models: the ICH score by point level, a composite model comprising individual components of the ICH score, and a composite model with the individual components of the ICH score plus delirium category. ROC curves are shown for (A) the entire cohort, (B) patients with Glasgow Coma Scale (GCS) score ≥13 only, and (C) patients with GCS score <13 only.

In exploratory ROC curve analyses stratifying patients by GCS score, we found that the accuracy of both the composite model of individual ICH score predictors and the ordinal ICH score itself decreased substantially when considering only patients with GCS ≥13 (composite model: AUC 0.809, 95% CI 0.709–0.909; ordinal ICH score: AUC 0.715, 95% CI 0.607–0.823; figure 2B) or only those with GCS <13 (composite model: AUC 0.832, 95% CI 0.750–0.914; ordinal ICH score: AUC 0.718, 95% CI 0.621–0.816; figure 2C). However, the composite model that included delirium category maintained a high predictive accuracy for both groups of patients (GCS ≥13: AUC 0.902, 95% CI 0.852–0.953; GCS <13: 0.887, 95% CI 0.821–0.953).

Discussion

In this hypothesis-generating study, we found that established predictors of outcome after ICH have high accuracy in predicting WLST, and that among these, impairments in consciousness—namely, the continuum of delirium and coma—may be the most strongly predictive individual factors. Further, our findings suggest that delirium may foreshadow subsequent end-of-life decision-making whether it occurs early or later during hospitalization. However, whether this phenomenon is due to effects on clinician or surrogate decision-making remains unclear.

Though level of consciousness plays a prominent role in most predictive models of ICH outcome and has previously been associated with WLST via the admission GCS score,^4,7,21,22 such models assume that each underlying predictor is a fixed entity. However, just as hematoma expansion leads to worse outcomes than would be predicted by initial hematoma volumes,²⁴ subsequent worsening of mental status may have similar effects on decision-making leading to WLST. These effects may be abetted by the more apparent evidence of mental status changes for patients' surrogate decision-makers, whereas concepts of hematoma size, location, or IVH may be more challenging for nonclinicians to conceptualize.

Although the GCS is commonly used as a crude measurement of mental status in patients with ICH and those with other types of acute neurologic injury, it may not always accurately reflect true impairments of consciousness and is therefore not a surrogate for delirium. Some neurologic sequelae of stroke may lead to deceivingly lower scores in awake patients—for example, in cases of aphasia leading to decreased verbal or motor subscores,²⁵ or eyelid opening apraxia leading to decreased eye subscores. Further, delirium without decreased arousal may result in near-normal GCS scores, and our findings suggest that the development of delirium still accurately predicts WLST even in such cases. Therefore, while delirium likely acts as a marker of more severe ICH in many cases, these findings in patients with less severe ICH suggest that delirium may confer its own additional mediating impact on WLST.

There is debate regarding delirium diagnosis in patients with underlying neurologic disorders, with varying nonspecific terminology (including encephalopathy, confusion, and altered mental state) used to describe clinical features that meet criteria for delirium because they occur in the setting of acute stroke. In some cases, this may represent a semantic issue, since delirium is an umbrella diagnosis that does not imply a single cause or mechanism, and acute cerebrovascular disorders are not excluded from consideration as an underlying medical cause (with higher rates of even covert strokes identified in perioperative patients diagnosed with delirium²⁶). However, disentangling expected stroke symptoms from superimposed delirium features is not straightforward, and there is no single neuroanatomic localization that is pathognomonic of delirium (although some might increase the likelihood²⁷). Nevertheless, even in view of possible disagreements regarding standardized classification, our finding that acute cognitive disturbances after ICH are associated with WLST raises concerns for possible biases that had previously gone unrecognized, while demonstrating a need for clarification of delirium diagnosis in further prospective studies of neurologic patients.

The reasons that delirium might occur after ICH are numerous, and this heterogeneity should be considered when interpreting its impact on decision-making and in overall prognostication. For example, the long-term implications of a patient developing delirium due to hematoma expansion or cerebral edema, which we considered in a post hoc sensitivity analysis, are likely different from those for patients whose delirium is due to potentially reversible causes such as infections, hydrocephalus, or nonconvulsive seizures and postictal states. However, the external manifestations of delirium, regardless of its cause, may be sufficient to bias clinicians and patient surrogates, a phenomenon that would be especially concerning if it deprives a patient of the opportunity for a potentially reversible cause to resolve. As such, though guidelines on end-of-life care in patients with stroke have recognized the importance of delirium and the effects it may have on family perceptions,²⁸ more awareness of its implications for prognostication is warranted, especially in the context of guidelines that also recommend avoidance of early care limitations after ICH.²⁹

Our findings reinforce concerns about self-fulfilling prophecies arising from the use of the ICH score and its components. Indeed, while physician judgement may have moderate accuracy in predicting outcomes after ICH,³⁰ there is a great deal of variability in physician prognosis and recommendations, and providing a prognostic score suggestive of poor outcome may increase the likelihood that physicians will recommend against aggressive care.³¹ However, we found that even such prognostic scores are less accurate in predicting WLST without accounting for mental status, and as such, delirium and other forms of impaired consciousness may be the driving factor behind the self-fulfilling prophecy of predictive models of ICH outcomes.

Our study has several limitations. First, it is a single-center study, which may reduce the generalizability of our findings based on patient demographics and clinical practice patterns. However, accounting for differences in approach and population, our results reasonably approximate the frequency of delirium, mortality, and WLST in prior studies, suggesting the results are likely generalizable. Second, delirium diagnoses were made retrospectively in most cases, which limited our ability to precisely determine the day of delirium onset for each patient. However, a sizeable subset of patients was prospectively assessed by an expert clinician, a potential strength in comparison to other contemporary studies of poststroke delirium, and rates of WLST were similar in patients whose delirium was diagnosed retrospectively and prospectively. Third, we did not have details on the underlying etiology for each case of delirium, including secondary causes such as fever or metabolic abnormalities, which may have proved especially helpful in patients with delirium that occurred later and that may not have been due to immediate complications of the ICH itself. Fourth, the proportion of patients in our cohort with a known history of dementia was not insubstantial, and the diagnosis of delirium superimposed on dementia presents its own challenges.³² While we used recognized strategies for the diagnosis of delirium superimposed on dementia, it is possible we may have misclassified some cases. Finally, we did not have details on the underlying reasoning behind WLST in each case, and therefore cannot explicitly characterize the conscious impact of delirium on clinician and surrogate decision-making. These limitations make a future prospective study using both qualitative and quantitative methods critically important to establishing the causal link between post-ICH delirium and decisions related to WLST.

In conclusion, we found that delirium predicts WLST after ICH regardless of when it occurs. Because delirium is a potentially reversible process, its impact on decisions related to WLST must be considered carefully, and further study of the effects it may have on clinician and surrogate decision-making is warranted.

Glossary

AUC

area under the receiver operating characteristic curve

CAM-ICU

Confusion Assessment Method for the Intensive Care Unit

CI

confidence interval

DSM-5

Diagnostic and Statistical Manual of Mental Disorders, 5th edition

GCS

Glasgow Coma Scale

HR

hazard ratio

ICH

intracerebral hemorrhage

IQR

interquartile range

IVH

intraventricular hemorrhage

ROC

receiver operating characteristic

WLST

withdrawal of life-sustaining treatment

Appendix. Authors

Study funding

No targeted funding reported.

Disclosure

M.E. Reznik received relevant support from the Rhode Island Foundation. S. Moody, K. Murray, S. Costa, B.C. Mac Grory, T.E. Madsen, A. Mahta, L.C. Wendell, B.B. Thompson, S.S. Rao, C. Stretz, K.N. Sheth, D.Y. Hwang, D.B. Zahuranec, M. Schrag, L.A. Daiello, W.F. Asaad, R.N. Jones, and K.L. Furie report no relevant disclosures. Go to Neurology.org/N for full disclosures.