FormalPara Take-home message

In patients who are comatose after resuscitation from cardiac arrest, the highly malignant electroencephalogram (EEG) patterns recommended in the 2021 guidelines from the European Resuscitation Council and the European Society of Intensive Care Medicine predict poor outcome with 93% specificity beyond 24 h after arrest. The combination with an unreactive EEG background significantly improves prognostic performance.

Introduction

Hypoxic–ischaemic brain injury is a leading cause of intensive care unit (ICU) admission after resuscitation from out-of-hospital cardiac arrest (CA) [1, 2]. In patients who are comatose after resuscitation, about two-thirds of deaths occurring after ICU admission are due to neurological causes and generally occur after withdrawal of life-sustaining therapy (WLST) following a predicted poor neurological outcome [3, 4]. Electroencephalography (EEG) is the most widely used and available method to assess prognosis after CA [5] and is included in the multimodal prognostication algorithm of the 2021 guidelines endorsed by the European Resuscitation Council (ERC) and the European Society of Intensive Care Medicine (ESICM) for post-resuscitation care [6]. These guidelines recommend using suppression or burst-suppression, with or without discharges (electronic supplemental material, ESM, Fig. E1), at 24 h or later after CA, combined with at least one other concordant predictor, to predict poor outcome. These EEG patterns are usually referred to as “highly malignant EEG patterns” (HMEP) and are defined according to the standardised EEG terminology of the American Clinical Neurophysiology Society (ACNS) [7, 8]. The 2021 ERC-ESICM guidelines were based on an extensive systematic review showing that HMEP almost invariably predict poor neurological outcome, especially if they are recorded beyond 24 h after CA [9]. However, the certainty of the supporting evidence is low [6]. Knowledge gaps regarding the prognostic accuracy of EEG include the potential interference of hypothermia or sedation [10] and the value of EEG reactivity, defined as a change in the EEG background to external stimuli, for instance, sound and pain. Results of the systematic review informing the 2021 ERC-ESICM guidelines suggest that an unreactive EEG background is less accurate for predicting poor neurological outcome than HMEP [9]. However, it is unclear what the added value of absent EEG reactivity is in combination with HMEP.

The present study aimed to evaluate the prognostic ability of HMEP recommended by the 2021 ERC-ESICM guidelines in a large patient cohort and to investigate the added value of the absence of reactivity combined with HMEP. The secondary aim of this study was to assess whether sedation, hypothermia treatment or time point affected the prognostic reliability of EEG.

Methods

This is a pre-planned substudy of the international “Targeted Hypothermia versus Targeted Normothermia after Out-of-hospital Cardiac Arrest. A Randomised Clinical Trial”, TTM2-trial, in which adult comatose patients resuscitated after out-of-hospital CA of presumed cardiac cause were randomised to temperature control to 33 °C versus early treatment of fever (\(\ge\) 37.8 °C) [11]. The trial randomised 1900 patients between November 2017 and January 2020.

The ethics committees in participating countries approved the trial protocol (ClinicalTrials.gov NCT02908308) [12]. Consent was obtained from a legal representative and each patient regaining mental capacity.

Participants allocated to TTM at 33° C were rapidly cooled with a device and maintained at the targeted temperature until 28 h after randomisation, and then they were rewarmed for 12 h. In the normothermia arm, the aim was early treatment of fever (≥ 37.8 °C). Sedation was mandatory until 40 h after randomisation in both allocation arms; afterwards, it was stopped or tapered to minimal levels. There was no defined protocol for sedation and analgesia, but short-acting drugs or volatile anaesthesia were recommended, and the Richmond Agitation–Sedation Scale (RASS) of minus four was targeted. If the sedation was stopped before prognostication, the time point was reported in the electronic case report form (eCRF). This information was used to determine whether sedation was ongoing during the EEG. The cumulative dose and type of sedatives up to 72 h after CA were reported, but data regarding the dosage at the exact time of the EEG were unavailable.

Based on the TTM2 protocol, EEG recording was mandatory in patients who were still unconscious (not obeying commands) between 48 h and 96 h after CA, corresponding to a time interval beyond the intervention period when sedation was stopped or kept as low as possible. If this recommended time interval coincided with a weekend, the EEG was performed immediately afterwards. Instructions for performing and interpreting EEG, either routine EEG or continuous EEG monitoring, were prespecified (ESM) and included in the TTM2 protocol. The EEG recordings were assessed by local reviewers who were not blinded to clinical data in the EEG referral. The treating team was not blinded to the local EEG report. EEG results and time points were prospectively reported in the eCRF by the investigator team, and the sites were instructed to use the following classification defined according to the ACNS [7]:

• HMEP (yes/no):

  • Suppressed background (< 10 µV the entirety of the record) with or without superimposed periodic discharges.

  • Burst-suppression pattern with or without superimposed discharges with suppression periods (< 10 µV) constituting ≥ 50% of the recording.

• EEG background reactive to any external stimuli (yes/no):

  • Sound stimuli (calling the patient’s name and clapping hands) repeated at least two times.

  • Painful stimuli (both distal and proximal) repeated at least two times.

At 96 h or later after randomisation, a physician blinded to the target temperature intervention performed a multimodal prognostication. Comatose patients with Glasgow Coma Scale (GCS) motor scores 1–2 without confounding factors, such as severe metabolic derangement or lingering sedation, were eligible for prognostication. The criteria for predicting a poor prognosis according to the trial protocol were fulfilled if at least two of the following predictors were present: bilateral absence of pupillary and corneal reflexes, status myoclonus, unreactive HMEP, computed tomography (CT) or magnetic resonance imaging (MRI) of the brain with signs of global ischaemic injury, high neuron specific enolase (NSE) levels and bilaterally absent cortical somatosensory evoked potentials (SSEP) N20 (negative peak at 20 ms)-responses. Details are reported in the trial protocol [12]. The result of multimodal prognostication was prospectively reported; “likely poor prognosis” (Yes/No). The blinded physician who performed the prognostication could share the information relating to the neurological prognosis with the treating physicians but was not allowed to recommend WLST; this latter decision rested with the treating clinical team. If the prognosis was uncertain, active intensive care was continued, and patients were re-evaluated daily.

Follow-up was performed face-to-face or by telephone interview 180 days after CA. Poor outcome was defined as a modified Rankin Scale (mRS) score of 4–6 (moderate-severe disability, severe disability, or death) [13].

We used SPSS version 28 for the statistical analysis. We included the first EEG performed between 24 h and 14 days after CA. We calculated the ability of the HMEP to predict poor outcome (specificity, sensitivity, positive predictive value, and negative predictive value). To investigate the added value of unreactive background in combination with HMEP, we used the McNemar test. We performed Chi-square tests to assess whether the predictive ability of HMEP was similar between patients with or without ongoing sedation and between patients in the hypothermia vs the normothermia group. We used McNemar’s and Fisher’s tests when comparing the prognostic ability of HMEP at various time-windows (0-24 h, 24-48 h, 48-72 h, 72-96 h, 96-120 h, and > 120 h). We used logistic regression to calculate odds ratios (OR). We calculated 95% confidence intervals according to Wilson’s method.

Results

Patients

Among 1900 patients randomised in the TTM2-trial, 1029 were still comatose and alive at the time point of prognostication and eligible for an EEG according to protocol. EEG was not performed in 110 of these patients (Fig. 1, ESM Table E1). Among the remaining 919 patients, 14 were excluded because the EEG results were missing in the eCRF. In addition, 60 patients were excluded from the primary analysis (reported separately in the ESM Table E2) because their EEG was recorded before 24 h from CA. Thus, 845 patients (678 [80%] males; mean age 65 years) at 59 trial sites were included in the primary analysis. Their baseline characteristics are summarised in Table 1.

Fig. 1
figure 1

Flow chart of study inclusion

Full size image
Table 1 Patient characteristics
Full size table
Table 2 Prognostic value of EEG
Full size table

EEG was recorded at a median of 71 h after CA (interquartile range [IQR] 52-93 h; range 24 h–13.3 days) (ESM Fig. E2). Prognostication was performed in 611 (72%) patients, of whom 412 underwent WLST. In 304 of these patients, the reason for WLST was a presumed poor neurological prognosis. Prognostication and WLST approach per country is presented in ESM Table E3. At 6-month follow-up, 579 (69%) patients had poor neurological outcome.

Table 3 Characteristics of patients with a falsely pessimistic EEG-based prediction
Full size table

The 110 missing patients without an EEG had a shorter time to ROSC (p = 0.003), fewer clinical seizures (p < 0.001), fewer rates of WLST (p = 0.011), and a better long-term outcome (p < 0.001) (ESM Table E1).

Predictive value of highly malignant EEG patterns

Of 845 patients, 307 (36%) had HMEP (Table 2). HMEP beyond 24 h predicted poor outcome with 50% (confidence interval [CI] 46–54%) sensitivity and 93% (CI 90–96%) specificity.

Eighteen patients with HMEP on EEG had a good outcome at 6 months. Their characteristics are described in Table 3. The distribution of false-positive patients was similar across the participating countries (ESM Table E3).

For the subgroup of patients in whom WLST for neurological reasons was not performed the sensitivity and specificity to predict poor outcome were 40% (CI 34–46%) and 93% (CI 90–96%), respectively (ESM Table E4).

Detailed follow-up data (mRS score) is presented in ESM Table E5.

Background reactivity

Results of reactivity testing to external stimuli were available in 821 (97%) patients, of whom 298 (36%) had HMEP. Among these, 268 (90%) had an unreactive EEG background. The combination of HMEP and unreactive EEG was significantly more specific (97% vs 93%; p = 0.008) and less sensitive (46% vs 50%; p < 0.001) than HMEP regardless of reactivity (Table 2).

Unreactive background per se, without considering whether the EEG fulfilled criteria for an HMEP, had 60% specificity to predict poor outcome.

Thirty patients with HMEP had a reactive EEG background. Of these, eight (27%) had a good outcome.

Sedation

In the study cohort, 730 (86%) patients received propofol and 395 (47%) midazolam during the first 72 h in the ICU. Data on whether sedation was ongoing during the EEG recording were available in 600 (71%) patients, of whom 402 (67%) were sedated. No difference in the specificity of the HMEP was observed comparing patients with and without ongoing sedation (p = 1.000).

Eleven (65%) of the false-positive patients (Table 3) received midazolam. The cumulative dose of midazolam up to 72 h was higher in these 11 patients (median 386 mg, IQR 16–648) compared to the remaining 384 patients who received the drug (median 150 mg, IQR 27–341).

Level of targeted temperature management

Among 845 patients, 442 (52%) were randomised to hypothermia. There were no differences between the hypothermia and normothermia groups regarding the prevalence (p = 1.000) and predictive ability (sensitivity p = 0.934; specificity p = 0.149) of HMEP. Since the intervention period lasted until 40 h after randomisation, only 39 (9%) patients had ongoing hypothermia during the EEG recording.

Time point of electroencephalogram

The prognostic accuracy of HMEP recorded after 24 h did not change across the various time points, while it was significantly lower when the EEG was recorded before 24 h (ESM Table E2).

Multimodal prognostication and HMEP in relation to WLST

Prognostication was performed in 611 (72%) patients, of whom 244 fulfilled the multimodal trial criteria for a “likely poor prognosis” at the time point of prognostication (96 h) (Table 1). The odds ratio of performing WLST in the patients with an HMEP compared to those without an HMEP was 4.6 (CI 3.4–6.3), and the odds ratio of performing WLST for the patients with and without a “likely poor prognosis” was 14.3 (CI 9.7–21.5).

Discussion

To the best of our knowledge, this prospective international multicentre study on the accuracy of EEG after CA is one of the largest ever conducted and the one with the broadest geographical representation, involving 59 sites in Europe, USA, Australia, and New Zealand. Its results show that the false-positive rate of HMEP recorded after 24 h from CA is 7%, which is higher than that reported in previous smaller studies. In addition, it showed that combining HMEP with the absence of reactivity significantly improved specificity.

The 2021 ERC-ESICM guidelines [6] and the recently published guidelines from the American Neurocritical Care Society [14] recommended HMEP as a predictor of poor outcome in patients with hypoxic–ischaemic brain injury based on previous investigations showing a specificity close to 100% [9] and substantial agreement among raters [15]. However, this evidence was based on smaller studies where EEG assessment was often centralised and carried out by a limited number of experts [8, 9, 16,17,18,19]. Conversely, in our multicentre study, local reviewers with diverse backgrounds and experiences assessed the EEG. Although we provided the local investigator teams with instructions (ESM Appendix) on how to record and classify the EEG according to ACNS [7] the EEG assessment by these teams may have been suboptimal, which may explain the lower specificity we observed. Nevertheless, even if the pragmatic design of our study may have been associated with reduced accuracy, it improves its generalisability, showing the results of real-life implementation of the ERC-ESICM guidelines in a wide geographical area. The reduced accuracy of the HMEP shown by the present study underlines the importance of adopting a multimodal approach to prognostication after CA.

The representativeness of our sample is confirmed by the 11% rate of patients with missing EEGs, which is in line with previous literature [20,21,22]. Even though the reasons for missing EEGs were not documented, the patients without EEG had more benign baseline characteristics and better outcome compared to the study cohort. This suggests that our included sample is representative of patients with more uncertain neurological outcome, eligible for prognostication. Although this may potentially represent selection bias, the patients with missing EEGs underwent multimodal prognostication based on other tests (ESM Table E1).

In our study, lack of EEG background reactivity per se was only inconsistently associated with poor outcome after CA, in line with current evidence [9, 23]. However, the presence of an unreactive EEG background significantly increased the specificity of HMEP, with only a 3% false-positive rate. Conversely, 8/30 (27%) patients with HMEP and a reactive background achieved neurological recovery in our cohort. This finding aligns with a previous study [24] and suggests that, although reactivity is a rare occurrence with HMEP [25, 26], it may have an added value to reduce the risk of falsely pessimistic predictions and should be actively searched for.

Sedation affects EEG in a dose-dependent manner [27]. In a previous study from our group, the only patient with good outcome and burst-suppression had EEG recorded during sedation [16]. A recent study investigated the impact of sedation and found that a sub-type of burst-suppression, with so-called identical bursts predicted poor outcome also during ongoing sedation [28]. However, this pattern is typically transient and disappears in median 36 h after CA, thus before most of the EEGs in our study. In the present study, the false-positive patients received twice as high cumulative dose of midazolam during the first 72 h compared to the rest of the study population (Table 3). This may suggest a potential interference from midazolam on the prognostic accuracy of EEG. However, the small size (n = 11) of this subgroup prevents a meaningful statistical analysis. Specifically designed studies will be needed to address this point.

In the present study, treatment of clinical or electrographic seizures was not protocolised. We note, however, that in the recent TELSTAR trial [29] that investigated effects of antiseizure treatment, all patients with a highly malignant EEG background, e.g. periodic discharges over a suppressed background, had a poor outcome regardless of intervention arm.

Hypothermia affects neurotransmitter release and may cause depression of the EEG background [30], depending on temperature levels. Our results showed that the prognostic ability of HMEP was not affected in patients managed at 33 °C, in line with a previous study from our group [8]. Evidence informing the 2021 ERC-ESICM guidelines shows that the specificity of HMEP for predicting poor outcome is greater when they are recorded after 24 h from CA[9], which aligns with the data from the early EEG (< 24 h) subgroup in our study (ESM Table E2). However, there is limited available evidence on the prognostic accuracy of EEG recorded later than 96 h after CA [9, 19]. Results of our study showed similar prognostic ability of the HMEP between 24 h and 14 days after CA (ESM Table E2). Interestingly, the sensitivity of HMEP remained high over time, in contrast with other studies showing a rapid decrease of their prevalence after 24–36 h from arrest [9, 31]. The reasons for this may include differences in case mix or sedation protocol across studies. These results should be interpreted with caution since the cohort gradually changed over time due to awakening and deaths.

An important limitation of our study was that clinicians were not blinded to the EEG. HMEP was part of our multimodal prognostication protocol and associated with higher probability for WLST compared to patients without HMEP indicating a potential self-fulfilling prophecy bias which may have overestimated the specificity of HMEP [32, 33]. However, studies conducted in communities where WLST was uncommon reported higher HMEP specificities than our study [34, 35]. We further note that in the large subgroup of patients (n = 541) who did not undergo WLST due to a presumed poor neurological prognosis, the specificity of the HMEP to predict poor outcome was equal to that of the whole study cohort (ESM Table E4). Self-fulfilling prophecy bias is difficult to avoid in prognostication studies, especially for predictors like the EEG that cannot be concealed from the treating team because they are essential for clinical management (e.g. to detect and treat nonconvulsive seizures) [6].

To limit the risk of an inappropriate WLST, our trial protocol stated that no decision on treatment withdrawal could be based on a single predictor, as recommended by the current ERC-ESICM guidelines, and that the time point of prognostication should be no earlier than 96 h. In patients with HMEP, prognostication was based on a median of three prognostic tests, and WLST was performed after a median of 47 h after EEG. Importantly, meeting the trial criteria for poor prognosis was a significantly stronger predictor of WLST than HMEP alone. These findings suggest that the trial sites used multiple prognostic tools in addition to HMEP during multimodal prognostication, but do not rule out a self-fulfilling prophecy bias for EEG. We emphasise the importance of a conservative approach after cardiac arrest, avoiding early WLST. We note that patients with falsely pessimistic test results in our study recovered consciousness at a median time of one week. In a French observational study [36], awakening from post-anoxic coma took up to 12 days after CA. Finally, although WLST tends to occur earlier in patients with prolonged disorders of consciousness due to anoxic brain injury compared with other causes of acquired brain injury [37] recent evidence shows that some of patients with anoxic brain injury may eventually recover at long-term follow-up [38].

A second limitation of our study was that the EEG results were dichotomised into presence or absence of HMEP in the eCRF. Consequently, the relative proportions of the HMEP subtypes (suppression vs burst-suppression) and their respective accuracies are unknown. Third, although the type and the cumulative dose of sedatives up to 72 h after CA were reported, their exact dose at the time of EEG recording was not available. Finally, this study is focussed on the ability of EEG patterns only to predict poor outcome. The specific EEG features predicting good outcome [39] were not investigated. Data concerning the predictive power of EEG in combination with other prognostic predictors are under investigation and will be reported in future studies.

Conclusions

In our study, the HMEP recommended in the 2021 ERC-ESICM guidelines for post-resuscitation care predicted poor outcome after CA with 93% specificity when assessed by local EEG reviewers. This is lower than that reported in studies with centralised EEG review by experts, suggesting that the large-scale implementation of these guidelines may be associated with reduced accuracy. The combination of the HMEP with an unreactive EEG background significantly improves prognostic performance. As with most prognostic studies, a self-fulfilling prophecy bias likely contributed to our results. However, the specificity of HMEP was similar in the two subpopulations of patients in whom WLST did and did not occur.