Guidelines for Neuroprognostication in Comatose Adult Survivors of Cardiac Arrest

The scope of these Grading of Recommendations Assessment, Development and Evaluation (GRADE) guidelines is the prognostication of neurological outcome in adult survivors of nontraumatic OHCA and IHCA who remain comatose following ROSC. The purpose of these guidelines is to provide evidence-based recommendations on the reliability of predictors of neurological outcome in comatose survivors of cardiac arrest to aid clinicians in formulating a prognosis. The target audience consists of clinicians responsible for such counseling.

These guidelines provide recommendations on the reliability of select demographic and clinical variables as well as prediction models when counseling families and surrogates of comatose survivors of cardiac arrest. We categorized these predictors as reliable, moderately reliable, or not reliable. We based this categorization on a GRADE-based assessment of certainty in the body of evidence, as well as effect size (quantification of predictor accuracy) across published studies, primarily the false positive rate (FPR), as shown in Table 1. Reliable predictors, for the purposes of these guidelines, may be used to formulate a prognosis when the appropriate clinical context is present in the absence of potential confounders. These are predictors with clear actionable thresholds or clinical/radiographic definitions and a low rate of error in prediction of poor outcomes and with at least moderate certainty in the body of evidence using GRADE criteria. When the prognosis is formulated on the basis of one or more reliable predictors, the clinician may describe the outcome as “very likely” during counseling. Given the inherent limitations in neuroprognostication research, the clinician must nevertheless acknowledge the presence of uncertainty, albeit low, in the prognosis. Moderately reliable individual predictors may be used for prognostication only when additional reliable or moderately reliable predictors are present, in addition to the appropriate clinical context as specified above. These are also predictors with clear, actionable thresholds or clinical/radiographic definitions and a low rate of error in prediction of poor outcomes but with lower certainty in the body of evidence using GRADE criteria, often a result of smaller studies that result in imprecision. When the prognosis is formulated on the basis of multiple moderately reliable predictors, the clinician may describe the outcome as “likely” during counseling but must acknowledge “substantial” uncertainty in the prognosis. Moderately reliable clinical prediction models that generate predicted probabilities of outcomes, in contrast, may be used for prognostication during counseling in the absence of other reliable or moderately reliable predictors. However, it is recommended that the clinician describe the predicted probability of the outcome as “an objective estimate only, subject to considerable uncertainty.” Although the panelists recognize that predictors that do not meet the criteria to be described as reliable or moderately reliable are often used by clinicians in formulating their subjective impressions of prognosis, they have nevertheless been deemed not reliable for the purposes of these guidelines and cannot be formally recommended for prognostication on their own. However, variables deemed not reliable may be a component of reliable or moderately reliable prediction models.

Candidate predictors were selected on the basis of clinical relevance and the presence of an appropriate body of literature. Candidate predictors and prediction models were considered “clinically relevant” if, in the subjective opinion of the content experts and guideline chairs, the predictor or components of the prediction models were (a) accessible to clinicians, although universal availability was not required, and (b) likely to be considered by clinicians while formulating a neurological prognosis for comatose survivors of cardiac arrest. Predictors addressed in prior cardiac arrest neuroprognostication guidelines were thought particularly likely to be considered by clinicians and therefore prioritized. An appropriate body of literature was considered present for any clinical variable that fulfilled two criteria: (1) evaluated in at least two published studies that included a minimum of 100 study participants and (2) established as an independent predictor in a multivariate analysis. An appropriate body of literature was considered present for clinical prediction models with at least one external validation study of at least 100 patients in addition to the initial report on development of the model (also with a minimum of 100 patients).

Based on these criteria, the following candidate predictors were selected:

Clinical variables:

Clinical prediction models:

The Population, Intervention, Comparator, Outcome, Timing, Setting (PICOTS) question was then framed for the specific candidate predictors as follows: “When counseling surrogates of comatose adult survivors of cardiac arrest, should [predictor, with time of assessment if appropriate] be considered a reliable predictor of [outcome, with time frame of assessment]?”.

The outcomes rated “critical” using the GRADE 1–9 scale were functional outcome (average rating 8.33) assessed at or beyond 3 months from ROSC, mortality (average rating 7.67) assessed at or beyond discharge, and cognitive outcome (average rating 7.33) assessed at or beyond 3 months from ROSC. However, no studies that included cognitive outcomes met other full-text screening criteria for the systematic review. Following the systematic review, the mortality outcome (particularly when assessed at discharge, the most common time point in the literature) was recognized to be inseparable from WLST, which accounts for up to 80% of deaths in this clinical setting [5, 6, 8, 9]. Although the panel did provide recommendations for predictors of mortality, available in Supplementary Appendix 2, the body of evidence for the prediction of all-cause mortality was thought to be compromised by an unacceptably high risk of bias from the self-fulfilling prophecy and not reflect the probability of death when life support measures are used indefinitely and in their entirety. The body of evidence for predictors of progression to death by neurological criteria or death in the absence of WLST was considered insufficient to serve as the basis for recommendations. Therefore, the primary focus of recommendations in these guidelines will be the prediction of long-term functional outcome.

Neuroprognostication in these guidelines is primarily focused on the prediction of poor outcomes, reflecting the overwhelming majority of research in comatose survivors of cardiac arrest. More recent publications have examined the ability to predict good outcomes in this population [10]. Functional outcome assessment of comatose survivors of cardiac arrest in the published literature has overwhelmingly been performed with the Cerebral Performance Category (CPC) scale (Supplementary Appendix 3) [11, 12]. The CPC is typically dichotomized at the ability to perform activities of daily living or work in a sheltered environment (CPC 1–2 vs. 3–5). Following this convention, for the purposes of this systematic review, a poor functional outcome was defined as severe disability, a minimally conscious state or a vegetative/unresponsive-wakeful state. Importantly, this definition of poor outcome is focused on recovery of functional ability and not on recovery of responsiveness. Although recovery of responsiveness in patients with long-term disorders of consciousness may be seen many months or years following hospital discharge, most of these patients have severe, persistent disability [13, 14]. Additionally, because this definition of poor outcome is focused on recovery of functional ability, a distinction between a minimally conscious and chronic vegetative/unresponsive-wakeful state is not relevant to these guidelines. Severe disability was defined as the equivalent of CPC > 2: the inability to perform activities of daily living or work in a sheltered environment. Several studies used the modified Rankin scale (Supplementary Appendix 3), which was developed for the assessment of outcome in cerebrovascular disease [15, 16] and was considered an appropriate alternative to the CPC. Other functional outcome scales that incorporated the ability to perform activities of daily living were also considered acceptable. The assessment of functional outcome at 3 months or later is consistent with recommendations of the AHA consensus statement on primary outcomes for resuscitation science studies [17], as well as the AHA standards for neuroprognostication research following cardiac arrest [8]. There is evidence that a significant proportion of patients demonstrate an improvement in functional outcomes following discharge [18, 19]. In one study, 50% of patients progressed from poor to good functional outcome between 1 and 3 months [20]. In addition, the ability to work in a sheltered environment, travel by public transportation, or prepare food, all of which are inherent to the CPC score, cannot be adequately assessed in the inpatient setting. Although a longer duration from time of injury to outcome assessment is ideal to capture the entirety of functional recovery, this may result in loss to follow-up. Significant loss to follow-up in observational studies may result in a selection bias based on the patients most likely to respond or return to the index hospital for further medical care.

An in-depth description of systematic review methodology for these guidelines is in Supplementary Appendix 1. The librarian search string used for this systematic review is in Supplementary Appendix 4 and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram in Fig. 1. Full-text screening was performed with the following exclusion criteria: sample size less than 100, studies focused on a highly selected subgroup (such as traumatic cardiac arrest), studies of predictors not established as independent with multivariate analysis, studies focused on a genetic polymorphism as a predictor, and studies of clinical prediction models that did not report model discrimination. Studies of laboratory biomarkers were included only if the biomarker was considered clinically relevant and had been evaluated in more than one published study that met other criteria.

Studies with no restrictions on WLST and likely incorporation of predictors under investigation into clinical neuroprognostication during the course of the study were considered to have a high risk of bias from the self-fulfilling prophecy. Studies that included a systematic restriction of WLST for at least 72 h and those that blinded clinicians to the predictor under investigation were considered to have a moderate risk of bias in this domain. Studies from countries with restrictions or cultural limitations on withdrawal of life support, including from East Asia, were judged to have a lower risk of bias from the self-fulfilling prophecy [21]. However, withholding escalation of therapy, including cardiopulmonary resuscitation in the event of recurrent cardiac arrest, is relatively common in these settings [21]. Studies from these settings with a mortality outcome that did not include specific restrictions on withholding escalation of care were therefore judged to have a moderate risk of bias from the self-fulfilling prophecy.

A summary of individual studies of predictors is in Supplementary Appendix 5. The GRADE evidence profile (EP) and summary of findings (SoF) table for predictors of functional outcome is in Table 2, and the EP/SoF table for predictors of mortality is in Supplementary Appendix 2.

Predictor accuracy is often described using measures such as the odds ratio (OR), which measures the relative probability of the outcome when the predictor is present compared to the probability of the outcome in the absence of the predictor. In the context of counseling families of comatose survivors of cardiac arrest on the likelihood of a poor outcome, the single most important measure of accuracy of individual clinical variables may be the FPR, the proportion of patients with a good outcome in whom the predictor was observed. FPR = [false positives/(false positives + true negatives)] = [1 – specificity]. The sensitivity of the predictor, or the proportion of patients who suffer a poor outcome in whom the predictor is observed, is also useful. Predictors with lower sensitivity, such as the bilateral absence of both the pupillary light response and corneal reflex, typically have lower prevalence in comatose cardiac arrest survivors, whereas the rate of poor outcome may be high. Clinical prediction model performance is evaluated using measures of model discrimination and calibration, as described in Supplementary Appendix 1 [22].

A summary of all recommendations for the prediction of long-term functional outcome is in Table 3. Recommendations for the prediction of mortality are in Supplementary Appendix 2.

In accordance with recommendations of the GRADE network, these statements were considered by the panel to be actionable, supported by indirect evidence where appropriate, and essential to guide the practice of neuroprognostication [25]. The good clinical practice reflected in these statements lacked a meaningful body of direct supporting evidence (typically because of insufficient clinical equipoise) but was considered by the panel to be unequivocally beneficial.

We recommend deferral of assessment of the neurological prognosis of comatose survivors of cardiac arrest for at least 72 h following ROSC in patients not treated with therapeutic hypothermia (goal temperature < 36.5 °C) and at least 72 h following rewarming in patients treated with hypothermia. However, persistence of coma beyond this period in the ICU must not be equated with a poor neurological prognosis (strong recommendation, evidence cannot be graded).

The rationale for this recommendation is that the majority of patients who awaken from coma following cardiac arrest will do so within the recommended time frame, rendering neuroprognostication unnecessary [3, 5‐7, 20, 26‐32]. In the Targeted Temperature Management (TTM) clinical trial, 45% of patients in the hypothermia (33 °C) arm and 52% of patients in the normothermia arm (36 °C) woke from coma within 72 h of rewarming [4]. About 64–88% of patients not treated with hypothermia who awaken will do so within 72 h from ROSC, whereas 51–90% of patients treated with hypothermia who awaken will do so within 72 h from rewarming [4, 5, 20, 27‐29, 31‐39]. A further 10–15% of patients will suffer death by neurological criteria, and a similar percentage will die of cardiopulmonary instability in the first few days despite the use of life support measures, again rendering neuroprognostication unnecessary [4, 5, 40‐42]. Note that some tests required for neuroprognostication, such as SSEP and imaging, can be performed beyond 48 (rather than 72) hours. Most, but not all, patients who awaken in the ICU will have a good long-term neurological recovery. There is an insufficient body of evidence to identify predictors of poor outcome in patients who do awaken.

Therapeutic hypothermia to a goal temperature of 33 °C is frequently used as a neuroprotective strategy following cardiac arrest, despite two large clinical trials, TTM and TTM-2, demonstrating equivalent outcomes with the use of therapeutic normothermia (36–37.5 °C) [43, 44]. The use of therapeutic hypothermia most likely delays awakening [5, 31, 33, 36], possibly through prolongation of the impact of sedation, although this is not a universal finding [37]. It is therefore reasonable to allow a longer period (72 h from rewarming rather than from ROSC) following the use of therapeutic hypothermia to maximize the number of patients in whom neuroprognostication is rendered unnecessary.

Although neuroprognostication is relevant only to patients who remain comatose (do not awaken) following cardiac arrest, awakening in the ICU must not be equated with neurological prognosis: the persistence of coma beyond 72 h or even seven days does not automatically imply a poor long-term outcome. A substantial number of comatose survivors will awaken beyond this period; 10–22% of all cardiac arrest survivors in the ICU not treated with hypothermia will awaken > 72 h from ROSC, and 10–19% of patients treated with hypothermia will awaken > 72 h from rewarming [5, 27‐29, 35, 39]. Of patients who awaken beyond this period, 67–88% will have a good long-term functional outcome [20, 31, 32, 39]. When the waiting period in the ICU is extended to 7 days, 83–100% of patients who eventually awaken will have done so within this time period [33, 35, 38]. In a study from Taiwan, where WLST is restricted, 7% of all cardiac arrest survivors woke beyond 7 days from ROSC, of whom 74% had an excellent functional outcome (CPC = 1) at 6 months [38]. Multiple instances of awakening beyond 2 weeks [31, 32, 36, 38] or even 4 weeks [13, 14, 36, 39] have been described. In the TTM clinical trial, seven patients awoke between days 15 and 22, of whom three (43%) had a good long-term functional outcome [31]. Neuroprognostication in patients with persistent coma beyond 72 h from ROSC/rewarming must therefore be based on objective predictors with low FPRs, as described below, rather than the persistence of coma alone.

Of note, awakening from coma following cardiac arrest is variably defined in the literature as follows: a Glasgow Coma Scale (GCS) score > 8 [33, 34, 39]; a Richmond Agitation Sedation Score >  − 2 [28, 29]; orientation to either person, situation, or place [20, 38]; and (most commonly) the ability to follow commands as documented by a GCS motor score of 6 [4, 5, 27, 31‐33, 35‐39]. “Early” awakening has also been variably defined: within 72 h from ROSC [5, 27, 33, 34, 37], 48 h from rewarming [20, 35], 72 h from rewarming [4, 20, 31, 39], 48 h from cessation of sedation [28, 29, 32], 5 days from ROSC [36], 7 days from ROSC [33, 38], and 7 days from rewarming [35]. Regardless of definition, patients who awaken early are more likely to have a good functional outcome at 3–6 months than patients who awaken late: 82–97% vs. 67–93% [20, 31, 32, 38, 39].

We recommend that assessment of the neurological prognosis of comatose survivors of cardiac arrest be performed in the absence of sedation or other potential confounders (strong recommendation, evidence cannot be graded).

Sedation will delay awakening following cardiac arrest, and waiting an appropriate length of time for sedation to clear may permit awakening and render neuroprognostication unnecessary [28, 32]. Sedation may also confound EEG, and the use of opioids may result in pupillary constriction, confounding the subjective evaluation of pupillary reactivity [45]. The duration of sedative effect is dependent on medication half-life, duration of infusion, hepatic and renal function, drug interactions, patient age, temperature, and comorbidities, among others. The involvement of a clinical pharmacist may help clarify the expected duration of sedative effect in complex critically ill patients. Other factors that may confound the neurological examination and delay awakening following cardiac arrest include, but are not limited to, seizures [27, 31, 36], hypothermia [5, 31, 33, 36], sepsis, renal failure [29], delirium [32], and hepatic encephalopathy. A defined time period of waiting cannot be recommended to cover all potential confounders. Instead, this time period must be individualized with careful consideration of all relevant factors. Some predictors, such as SSEP and quantitative pupillometry, may be less subject to confounding.

We recommend that factors that impact overall prognosis, such as a poor baseline level of functioning, preexisting illness associated with limited life-expectancy, and multiorgan failure, be considered prior to, and distinct from, assessment of the neurological prognosis of comatose survivors of cardiac arrest (strong recommendation, evidence cannot be graded).

Neuroprognostication should not be conflated with an overall assessment of prognosis. Given the definition of a poor outcome in our systematic review, patients with severe dependence and disability at baseline are outside the scope of these guidelines. Similarly, assessment of long-term neurological outcome may not be relevant in critically ill patients with a high short-term risk of death from multiorgan failure. Assessment of long-term neurological prognosis may also not be relevant in patients with a poor prognosis for long-term survival from conditions such as advanced malignancy.

We recommend that assessment of the neurological prognosis of comatose survivors of cardiac arrest be multimodal, with consideration of the complete clinical scenario, and never based on a single variable (strong recommendation, evidence cannot be graded).

As previously described, an element of uncertainty is inherent in neuroprognostication, even with predictors considered “reliable.” Therefore, it is critical that a thorough assessment be performed with every patient and that the entirety of the clinical picture be considered. This uncertainty is related in part to the inability of studies to fully account for the self-fulfilling prophecy while evaluating the reliability of a predictor [46]. In addition, the possibility of technical, operator, or interpretation error while evaluating a predictor is always present [45, 47‐49], and the ability to account for confounders, such as sedation, is imperfect. The use of multiple modalities of assessment, such as clinical examination, imaging, and electrophysiological studies, will mitigate the risk of error from a single modality. Information from these modalities should be largely consistent. For example, the reliability of a subjective determination of nonreactive pupils should be called into question and operator error should be considered in a comatose cardiac arrest survivor with a withdrawal motor response, reactive EEG, and normal brain imaging at 72 h.

We recommend that in the absence of reliable (or multiple moderately reliable) predictors of outcome, surrogates of cardiac arrest survivors who remain comatose at the time of neuroprognostication should be counseled that the likelihood, extent, and time course of neurological recovery is uncertain. Surrogates should also be counseled that the timeline of any functional recovery that does occur may extend from several days to several months (strong recommendation, evidence cannot be graded).

Clinicians should directly acknowledge uncertainty when the prognosis is indeterminate. Many patients will suffer a poor outcome or recover only limited function. When functional recovery does occur the time course of recovery is highly variable. Patients may awaken beyond 72 h and achieve a good functional outcome prior to hospital discharge [5, 27, 29, 33, 35, 36]. Awakening and progressive improvement may also occur following discharge, especially in the first 3–6 months [18‐20].

We recommend an extended period of observation for signs of neurological recovery in comatose survivors of cardiac arrest with an indeterminate prognosis, if consistent with the goals of care as established through discussions with patient surrogates (strong recommendation, evidence cannot be graded).

The appropriate length of observation must be established through discussion with surrogates based on a best estimate of the patient’s willingness to undergo an extended duration of life-sustaining treatment. Although the majority of patients who have a good long-term functional recovery will have awakened in the first 2 weeks, a longer period of supportive care and observation, extending over three or more months, will provide a higher level of certainty in the final outcome [20, 31, 32, 38, 39]. The possibility of a good functional recovery in the setting of a chronic disorder of consciousness is never zero but does decline significantly beyond this period following cardiac arrest [14]. Although responsiveness to the outside environment may return in patients considered to be in a vegetative/unresponsive-wakeful state, this is infrequently associated with functional recovery [14]. Several other factors may impact the duration of observation. Supportive care over several months often requires invasive measures, such as tracheostomy and percutaneous gastrostomy, which may not be acceptable to all. An extended duration of supportive care may impose a financial and caregiver burden on surrogates, particularly in low- and middle-income countries, and large out-of-pocket health care expenditures may be associated with poverty in these settings [23, 24, 50]. A transition to hospice care in the setting of a persistent disorder of consciousness several months following cardiac arrest may involve withdrawal of artificial nutrition. This is medically, legally, and ethically feasible in many, but not all, states and countries [51‐54]. Surrogates should be counseled on local laws regarding withdrawal of artificial nutrition or other forms of life support.

The suggested approach to neuroprognostication in comatose survivors of cardiac arrest, incorporating the recommendations in these guidelines, is in Figs. 2 and 3.

Table 4 lists key considerations with the evidence-based use of predictors in this population.

Only a limited number of predictors had a sufficient body of evidence to support recommendations for use in clinical practice. Although these predictors met criteria for reliability, they were often insensitive. Therefore, a large number of patients will have an indeterminate prognosis on the basis of these guidelines, highlighting the importance of future high-quality neuroprognostication research. The AHA outlined standards for cardiac arrest neuroprognostication research in 2019 [8].

Based on the most common study limitations identified in our systematic review, future studies should consider the following general principles:

Individual clinical variables are unlikely to be reliable predictors in isolation. The only exceptions in our systematic review were the bilateral absence of the pupillary light response and the bilateral absence of N20 responses on SSEP testing. Clinical prediction models that incorporate multiple independent clinical, electrophysiological, and imaging predictors may therefore be optimal for neuroprognostication. The primary limitation of the most widely studied clinical prediction models in the setting of cardiac arrest is a lack of validation for the prediction of long-term outcomes. Several prediction models based on machine learning algorithms, including neural networks, have undergone preliminary evaluation and await larger multicenter validation studies for the prediction of long-term outcome [158‐160]. Several brain injury biomarkers other than NSE, such as neurofilament light chain, tau protein, S100 calcium-binding protein B, glial fibrillary acidic protein, and ubiquitin C-terminal hydrolase L1, are currently under investigation and may demonstrate value in multimodal neuroprognostication following cardiac arrest [161].

These guidelines and the preponderance of the neuroprognostication literature focus on the prediction of poor outcome. However, analyses of predictor accuracy for good long-term neurological outcome should be routinely incorporated into future neuroprognostication research. Discussions with patient and family representatives also highlighted the importance of lesser-studied outcomes following cardiac arrest. Several studies have established that cognitive dysfunction, anxiety, depression, posttraumatic stress disorder, and impairment of health-related quality of life are common in long-term survivors of cardiac arrest [162‐170]. However, in addition to examining predictors of impairments in these domains, future prospective studies should use standardized instruments and time points for evaluation and compare occurrence with that in an age- and sex-matched control population.

These guidelines were endorsed by the American Heart Association and the Society of Critical Care Medicine. The American Academy of Neurology affirms the value of these guidelines.