EEG pattern with spectral analysis can prognosticate good and poor neurologic outcomes after cardiac arrest

Abstract

Introduction

To investigate the prognostic value of a simple stratification system of electroencephalographical (EEG) patterns and spectral types for patients after cardiac arrest.

Methods

In this prospectively enrolled cohort, using manually selected EEG segments, patients after cardiac arrest were stratified into 5 independent EEG patterns (based on background continuity and burden of highly epileptiform discharges) and 4 independent power spectral types (based the presence of frequency components). The primary outcome is cerebral performance category (CPC) at discharge. Results from multimodal prognostication testing were included for comparison.

Results

Out of a total of 72 patients, 6 had CPC 1 – 2 by discharge, all of whom had mostly continuous EEG background without highly epileptiform activity at day 3. However, for the same EEG background pattern at day 3, 19 patients were discharged at CPC 3, and 15 patients at CPC 4 – 5. After adding spectral analysis, overall sensitivity for predicting good outcomes (CPC 1 – 2) was 83.3% (95% confidence interval 35.9 – 99.6%), and specificity was 97.0% (89.5 – 99.6%). In this cohort, standard prognostication testing all yielded 100% specificity but low sensitivity, with imaging being the most sensitive at 54.1% (36.9 – 70.5%).

Conclusion

Adding spectral analysis to qualitative EEG analysis may further improve the diagnostic accuracy of EEG and may aid developing novel measures linked to good outcomes in post-cardiac arrest coma.

Introduction

For patients who remain in prolonged coma after cardiac arrest, guiding families and healthcare proxies in goals of care decisions presents a major challenge. With the advent of targeted temperature management (TTM), post-arrest mortality and morbidity have notably improved but remain high^1–3. Most patients die due to signs of hypoxic-ischemic brain injury (HIBI) and subsequent withdrawal of life-sustaining treatment (WLST), and for those who survive, a significant portion remains comatose or severely impaired^4–6.

Currently, the best predictors of outcome include:

• serial neurologic exams, primarily involving the detection of absent brainstem reflexes,

• malignant EEG patterns and status epilepticus,

• somatosensory evoked potentials (SSEP),

• serum biomarkers, namely neuron-specific enolase (NSE), and

• neuroimaging, computed tomography (CT) or magnetic resonance imaging (MRI).

When performed properly with the correct timing, abnormalities in these prognostication tests can be used to diagnose HIBI. Each abnormal test can have high specificity for predicting poor outcomes, generally ~100%, but these tests all have low sensitivity for good outcomes, at ~50% or even less^6–8. Predicting good outcomes is particularly difficult because these tests are designed for detecting signs of severe neurologic injury.

Due to the limitations of prognostication testing, especially in the first several days, there is the risk of premature WLST. Our recent studies have shown that EEG features suggesting preserved corticothalamic circuitry were correlated to better outcomes. Patients who exhibited faster brain wave activity were able to emerge from prolonged comas and to regain significant function, even if their EEG was initially in burst suppression pattern (BSP)^9–11. Traditionally, BSP is viewed as a highly malignant pattern, though the appearance of intra-burst activity may be a factor^6,12,13.

In contrast to earlier methods of pure visual EEG inspection, more recent studies now utilize more quantitative and spectral analysis, which opens the door to powerful machine learning methods for predicting outcomes. There have been important findings regarding background features such as reactivity^14,15, continuity¹⁶, and spectral power^17–19, yielding sensitivity as high as ~80% and specificity 80%−100% for predicting poor outcomes with a trade-off function. These studies generally avoid EEG with epileptiform activity, and not all address discontinuity.

In this study, we performed qualitative and spectral analysis of EEG data to identify common features in comatose patients after cardiac arrest. We aimed to apply a stratification system of EEG patterns and spectral types that may be helpful for standardized interpretation of EEG. EEG patterns with discontinuity and highly epileptiform activity were included. Correlations between outcomes and EEG features were assessed, and then compared to the other standard post-arrest prognostication tests.

Methods

Patient selection and records review

In total, 72 patients were enrolled in this study. This study included all patients 18 years of age or older who were admitted from November 2013 to Oct 2019 at the New York Presbyterian Weill Cornell Medical Center for cardiac arrest (either in- or out-of-hospital events) and who were placed on continuous EEG monitoring per clinical protocol. Patients who qualify were prospectively enrolled for the study, though data analysis was done offline and in a retrospective fashion. Family members and/or healthcare proxies were notified of the patient’s participation, with the option to opt-out. The study was approved under our Institutional Review Board.

Patients who were enrolled received standard-of-care treatments provided by the inpatient staff, including TTM and seizure management when appropriate, without any interference from the research team. The duration of EEG monitoring and the decision to pursue further prognostication testing were not affected by our study, and the research team was not involved in WLST discussions. Electronic medical records of these patients were systematically reviewed, and relevant demographic and clinical data were collected.

CPC score at time of discharge was used as the measure of neurologic outcome. For outcome analysis, a CPC of 1 – 2 representing good outcome. A score of 3 – 5 is considered a poor outcome, but for display purposes, a score of 3 will be separated from 4 – 5.

EEG recording and data selection

EEG data were acquired using the Natus Connex or EMU40 amplifier and the Xltek software (Natus Medical Incorporated, Oakville, Ontario, Canada). High-pass filter cutoff was set at 0.1Hz, low-pass cutoff at 70Hz, and sampling rate at 200Hz. Notch filter was set at 60Hz. Electrodes were placed using the standard 10–20 system.

Data were screened in Xltek viewer and then exported to a customized Matlab software, which was developed in house with selected scripts from the open-source EEGLAB package. All EEG data were reviewed in common reference montage. Data contaminated by heavy noise or artifact were not included. Post-acquisition digital filtering was not necessary for the most part.

If available, EEG data from the first 3 days (~72 hours) of recordings were selected for analysis. While somewhat arbitrary, this period covers the TTM treatment and rewarming window. The entire EEG recording was not reviewed in detail for this study. If available, epochs around the time of morning rounds (9AM) and bedtime (9PM) were selected to capture potential different states of stimulation and arousal, which is used to assess for variability. No formal reactivity data was used.

To limit potential bias, EEG epochs were selected separately and then analyzed by two independent investigators, and any significant discrepancies were discussed and rectified. Prior to completing all EEG analysis, investigators were blinded to clinical data, such as identifying information, demographics, medical history, clinical course, and outcomes.

EEG and spectral analysis

Each EEG epoch was first visually inspected. Then, 30 3-second segments were extracted to generate power spectral density (PSD) plots. In general, zenith leads were chosen for spectral analysis to minimize effects of factors such as external noise, myogenic or movement artifacts, and intrinsic and extrinsic sources of asymmetry.

PSDs were estimated using the Thompson multitaper method, with a time-half bandwidth product of 3. For epochs with BSP, including those with predominant epileptiform activity within bursts, only bursts ~1.5 seconds or longer were used for PSD analysis (similar methods as in^9–11). The raw EEG was reviewed prior to generating PSDs, the data were analyzed separately but were available during this process.

For correlation with outcomes, we used EEG data from day 3 of recording, if available. If EEG was disconnected before day 3, due to the patient waking up, WLST, or logistic reasons, data prior to disconnection were used instead. Data from day 1 and day 2 were used primarily for assessing variability or major changes in patterns.

Stratification of EEG patterns and spectral profiles

EEG data were grouped based on visual assessment into the following patterns:

• I)mostly continuous without highly epileptiform activity,
• II)BSP without highly epileptiform activity,
• III)BSP with highly epileptiform activity,
• IV)continuous highly epileptiform activity, and
• V)minimal or no activity.

BSP was defined as >50% suppression. Mostly continuous background would encompass continuous, nearly continuous, and discontinuous patterns. Highly epileptiform activity included generalized periodic discharges or frequent to abundant discharges, with morphology of sharps, spikes, polyspikes, spike and wave, polyspike and wave, or a mix of these discharges. Highly epileptiform discharges often were accompanied by a suppressed background. Rhythmic slow activity (either generalized or lateralized or more focal) without the described highly epileptiform discharges were not considered epileptiform and would be considered as I or II. Minimal or no activity referred to records with very low voltage <20 μV or seemingly flat, respectively.

This stratification system is largely based on patterns reported in other studies^{12,13,20–22}. Please see Figure 1 for representative EEG data. Including all BSP, patterns II – V would be considered malignant.

Figure 1: Example EEG patterns.

Data are shown directly from Xltek without digital processing, formatted for print. Channels are displayed as per Xltek output in common reference montage. Data from two different headboxes are shown. Each tick on the X-axis marks 1 second; amplitude scaled for visualization purposes. A) Pattern I – continuous background with mixed frequency components including some beta activity. B) Pattern I – continuous background but with diffuse, high-amplitude alpha activity. C) Pattern III – burst suppressed background with slow activity within bursts. D) Pattern IV – burst suppressed background with spikes and polyspike-wave discharges within bursts. E) Pattern V – very suppressed background with GPDs. F) Pattern V – minimal activity with electrocardiac artifact.

The power spectra of these EEG data were then reviewed. Any potential peaks in the delta (0 – 4Hz), theta (4 – 8Hz), and alpha (8 – 12Hz) ranges were identified. For EEG epochs containing BSP, only the signal segments within the bursts were extracted, similar to techniques in our earlier publications^9–11. The presence of beta activity was not included in analysis, mainly because it was very rarely the predominant rhythm.

Spectral data were grouped based on common features into the following types:

• 0)containing physiologic alpha peak,
• 1)theta peak,
• 2)delta peak or normal spectral floor, and
• 3)non-physiologic peaks only or abnormal spectral floor.

A spectral floor was considered normal if it follows roughly the 1/frequency power law. For EEG with minimal electrical activity, the floor would be notably lower in power throughout, which reflects the absence of brain activity and the presence of only ambient noise. For EEG with mostly epileptiform activity, the floor was usually much higher in power throughout, with high power in the higher frequencies and aliased peaks. EEG patterns III, IV, and V would yield abnormal spectral floors.

For EEG with low-amplitude slow activity, delta peaks were not always discernible from the spectral floor; therefore, spectra with normal floors with or without clear delta peaks were grouped together. Features such as anterior-posterior gradient and variability, seen either in the EEG tracings or the power spectra, were used to distinguish physiologic alpha peaks from pathologic ones. Non-physiologic peaks include persistent diffuse alpha activity as seen in alpha coma, noise or artifact peaks, and peaks from epileptiform activity. Please see Figure 2 for representative spectral data.

Figure 2: Example power spectra types.

Only zenith channels are shown, all with the same frequency and power scale. From left to right: Type 0 – containing alpha peak most prominent in posterior leads, without clear theta or delta peaks but with evidence of beta activity. Type 1 – no alpha but prominent theta peak. Type 2 – no clear peaks. Type 3 (a) – high-power spectra with abnormal noise floor. Type 3 (b) – low-power spectra with abnormal noise floor. Note, these spectra correspond to the EEG epochs shown in Figure 1, except alpha coma which is shown separately in Supplementary Figure 1.

Statistical analysis

Statistical testing of EEG and outcomes, which involved the non-parametric Kruskal Wallis test and multiple comparison using Tukey’s method, was performed using Matlab’s statistical package. Confidence interval approximation for sensitivity and specificity was done using the Wilson score.

Results

Clinical variables and outcomes

Relevant clinical data from the patient cohort were compiled, shown in Table 1. The 6 patients with good outcomes were younger overall, but several of the youngest patients had poor outcomes as well. Most patients (63 out of 72) received TTM, and 26 of the 72 patients underwent WLST. Median days of stay was shortest for patients in the CPC 4 – 5 group, likely reflecting WLST.

Table 1:

Clinical data grouped by outcomes

	CPC 1–2	CPC 3	CPC 4–5	Total
Number of patients	6	21	45	72
Male gender	5	15	28	49
Median years of age (range)	44.5 (32 – 54)	63 (29 – 84)	67 (19 – 89)	-
Median days of stay post-CA (range)	21.5 (10 – 30)	23 (9 – 125)	12 (1 – 140)	-
TTM	5	18	40	63
WLST	0	0	26	26

Prognosticators and outcomes

Results from prognostication testing are summarized in Table 2, similarly grouped by CPC scores. Included are the pupillary reflex (post-rewarming), SSEP, imaging, and NSE values, with abnormal results tallied. EEG results will be covered in more detail in the following sections.

Table 2:

Prognostication data grouped by outcomes

	CPC 1–2	CPC 3	CPC 4–5	Total abnormal	Total tested
Number of patients	6	21	45	-	72
Absent pupillary reflex (after rewarming)	1	1	12	14	72
Absent bilateral N20	0	0	9	9	29
CT or MRI suggestive of HIBI	0	0	20	20	63, 40*
First NSE level > 61 ng/mL	0	0	8	8	49

^*63 patients had CT scans. 40 had MRI scans.

All 72 patients had serial exams and EEG as part of this study, but not all underwent the remaining tests. The presence of myoclonus was excluded as a prognosticator, because myoclonus without epileptiform activity is not strongly correlated with poor outcomes^7,8,23. Only 29 patients had SSEP testing, and a report of absent bilateral N20 on the interpretation was scored as abnormal. 63 patients had CT scans, and 40 of these patients had follow-up MRI scans. Radiology reports consistent with HIBI on either CT or MRI were scored as abnormal.

The timing in NSE sample collection was highly variable in this study. Of the 72 total patients, 49 had at least one NSE collected. Of these 49 patients, 19 patients had NSE collected before 24hrs, with some samples done within few hours of cardiac arrest. For the remaining 30 of the 49 patients, some were collected >72hrs, and a few were collected even over 1 week from the initial event. 22 patients had an additional NSE value collected at varying intervals, and 13 had at least 3 NSE values. To simplify interpretation, a threshold of 61 ng/mL for the first NSE level was used to signify an abnormal result.

The diagnostic accuracy measures (sensitivity and specificity with 95% confidence intervals) of these prognosticators are summarized in Table 3. For imaging, only MRI results are shown due to higher sensitivity. Overall, abnormal results are highly specific for poor outcomes (CPC 3 – 5), but none of these tests are useful for predicting good outcomes.

Table 3:

Diagnostic accuracy of prognosticators for poor outcomes

	Sensitivity	Specificity
Absent pupillary reflex (after rewarming)	19.7% (10.9 – 31.3)	83.3% (35.8 – 99.6)
Absent bilateral N20	32.1% (15.9 – 52.4)	100.0% (2.5 – 100.0)*
MRI suggestive of HIBI	54.1% (36.9 – 70.5)	100.0% (29.2 – 100.0)
First NSE level > 61 ng/mL	17.8% (8.0 – 32.1)	100.0% (39.8 – 100.0)

^*Only 1 patient in CPC 1–2 group had SSEP tested

EEG data and outcomes

The CPC outcome scores for the different EEG patterns and spectral types are summarized in Table 4. For cases when the morning and night epochs were stratified into different patterns or types, the better of the two numbers was chosen, assuming the lower number is more reassuring.

Table 4:

EEG and spectral scores grouped by outcomes

	CPC 1–2	CPC 3	CPC 4–5
Number of patients	6	21	45
EEG patterns (day 1 | day 3*)
- I	6 | 6	19 | 19	14 | 15
- II		2 | 2	2 | 3
- III			13 | 8
- IV			3 | 4
- V			13 | 15
Spectral type (day 1 | day 3*)
- 0	3 | 5	2 | 2	0 | 0
- 1	2 | 1	6 | 6	7 | 7
- 2	1 | 0	10 | 13	10 | 10
- 3		3 | 0	28 | 28

^*Data from day 3 if available

The relationship between EEG pattern and CPC score is illustrated in Figure 3. Results from the Kruskal Wallis test and Tukey’s multiple comparison suggest that better CPC scores were statistically associated with EEG pattern I compared to patterns III and V. Patterns II and IV did not statistically separate from the other patterns, likely due to low numbers of patients with these patterns.

Figure 3: Boxplots of EEG data and outcomes.

CPC scores are plotted against top) EEG pattern and bottom) power spectra type. * and ** denote statistically different groups per Kruskal Wallis test and Tukey’s multiple comparison at α = 0.05.

In a similar fashion, Figure 3 also shows the relationship between spectral type and CPC, as well as the results from statistical testing. There was also evidence to support statistical correlation of spectral types with different outcomes.

In terms of predicting good outcomes, all patients with good outcomes (CPC 1 – 2) had EEG pattern I, but most patients with pattern I did not have good outcomes. Pattern I therefore had a sensitivity of 100.0% (95% confidence interval 54.0 – 100.0%) for predicting good outcomes but only specificity of 48.5% (36.0 – 61.1%). When using spectral types, sensitivity for predicting good outcomes was not as high at 83.3% (35.9 – 99.6%), but unlike for EEG patterns, specificity was much higher at 97.0% (89.5 – 99.6%).

From the perspective of predicting unfavorable outcomes (CPC 3 – 5), the malignant EEG patterns II – V had a sensitivity of 48.5% (36.0 – 61.1%) and specificity of 100% (54.1 – 100%). Using spectral types 1 – 3, sensitivity for predicting unfavorable outcomes was 97.0% (89.5 – 99.6%), and specificity was 83.3% (35.9 – 99.6%).

When combining EEG patterns and spectral types to predict good outcomes, the numbers are identical to those that result from using spectral types alone, because only EEG pattern I yielded spectral type 0.

Discussion

Limitations in interpreting outcome data

Aside from brain death, post-cardiac arrest patients die of either WLST or medical complications. In studies such as this one, it is worth simply mentioning that interpretation of outcome data will always be limited by the possibility that some patients may have died prematurely, due to early WLST or non-neurologic complications, resulting in higher numbers of poor outcomes. Outcomes can also be heavily influenced by the clinical scenario and the providers’ impression, which again can lead to higher poor outcomes. Prognostic testing can also be difficult to time properly for various logistical reasons, which can reduce the prognostication value.

Our opt-out method for patient recruitment was intended to increase inclusion and enrollment, but this type of study overall may unavoidably contain selection bias. Patients who required TTM and EEG monitoring likely already have more severe clinical courses. Patients who were opted out may also have more complex clinical scenarios.

Finally, for this study, the CPC score at time of discharge was used for assessing outcome. More long-term evaluations and more formal testing would likely give better estimations of the true level of neurologic recovery, since survivors can improve with time and rehabilitation^24–26.

Patients with good outcomes

For the group of 6 patients with good outcomes, their electronic records were reviewed in further detail (investigators were not blinded at this point). Regardless of the severity of their conditions, these patients all had delayed recovery of consciousness. EEG findings were relatively reassuring, and all had physiologic alpha activity by the end of recording with eventual good clinical recovery. In this case, the cause of delayed recovery was more likely medications and TTM, rather than brain injury.

The exam, EEG findings, and clinical course for these patients are summarized in Table 5. The initial exam is not always reliable with confounding sedation and hypothermia. To avoid premature WLST in the acute setting, it would be important to obtain EEG data earlier and exercise caution in making major decisions.

Table 5:

Clinical course for patients with good outcomes

	Day 1	Day 2	Day 3	Relevant narrative
Patient 66				Age 45, man. Out-of-hospital arrest, >20mins to return of spontaneous circulation (ROSC). Regained consciousness only briefly, developed convulsions and posturing. Intubated, sedated, received TTM. Later had fevers and pneumonia. On EEG from admission for 6 days, no seizures. Weaned off sedation by day 5, emerged from coma and following commands by day 7. Extubated on day 10, discharged on day 28 with CPC 1.
- Brainstem reflexes	No	No	No
- EEG pattern	I	I	I
- Spectral type	0	0	0
Patient 70				Age 44, man. Out-of-hospital arrest, ROSC 2mins. Remained comatose and developed convulsions and rigidity, worse with stimuli. Received sedation and TTM. On EEG for 4 days, no seizures. Weaned off sedation by day 3, emerged and following commands by day 4. Extubated on day 4, discharged on day 10 with CPC 1.
- Brainstem reflexes	Yes	Yes	Yes
- EEG pattern	I	I	I
- Spectral type	2	2	0
Patient 17				Age 52, man. Out-of-hospital arrest, ROSC 20mins. Required emergent cardiac catheterization for myocardial infarction, remained sedated and received TTM. On EEG for 4 days, no seizures, and on day 4 alpha activity emerged. Self-extubated on day 8, delirious at first but later was following commands. Discharged on day 25 with CPC 2.
- Brainstem reflexes	Yes	Yes	Yes
- EEG pattern	I	I	I
- Spectral type	1	1	1
Patient 33				Age 43, man. In-hospital arrest x2 on admission, ROSC 5mins x2. Had multi-organ failure and severe shock. Intubated, sedated, received TTM. Had systemic and pulmonary emboli, received thrombolysis and anticoagulation, developed small subdural hematoma. On EEG for 2 days, no seizures. Weaned off sedation by day 7, emerged and following commands by day 8. Extubated on day 8, discharged on day 15 with CPC 2.
- Brainstem reflexes	-	Yes	Yes
- EEG pattern	I	I	-
- Spectral type	0	0	-
Patient 36				Age 54, woman. Recent in-hospital arrest, ROSC 4mins, already had tracheostomy tube. Repeat in-hospital arrest, 2mins to ROSC. Intubated, sedated, no TTM. On EEG for 2 days after second event, no seizures. Began emerging from coma while on sedation on day 2. Consistently following commands by day 4. Discharged on day 30 with CPC 2, with tracheostomy collar.
- Brainstem reflexes	-	Yes	Yes
- EEG pattern	I	I	-
- Spectral type	0	0	-
Patient 52				Age 33, man. Out-of-hospital arrest, fell from height and drowned, intubated and resuscitated in field, unclear time to ROSC. Repeat arrest en route, ROSC 4mins. Sedated, received TTM. Significant pulmonary injury and small bilateral subdural hemorrhage. On EEG for 3 days, no seizures. Emerging from coma on day 2, consistently following commands by day 3. Extubated on day 8, discharged on day 18 with CPC 2.
- Brainstem reflexes	No	Yes	Yes
- EEG pattern	I	I	I
- Spectral type	1	1	0

Patients with poor outcomes

Patients with CPC 3 – 5 are generally regarded as having poor outcomes, but what qualifies as an acceptable outcome can be more complex. A CPC of 3 encompasses a wide range of capabilities and deficits, from being ambulatory to having full paralysis, and may be viewed as a favorable outcome for some. In addition, some patients may have pre-existing disability, which is not accounted for outcome analysis.

In this study, two patients had reassuring EEG patterns and spectral types but were discharged with a CPC score of 3 (see Table 6 for clinical summary). A more detailed chart review revealed that both patients had pre-existing neurologic deficits from prior stroke, and both recovered from cardiac arrest without significant new deficits. Practically, these patients had good outcomes, but for consistency’s sake, they were grouped with the rest with a CPC of 3.

Table 6:

Clinical course for patients with spectral type 0 and CPC 3

	Day 1	Day 2	Day 3	Relevant narrative
Patient 40				Age 84, man. Admitted for femoral fracture repair, had in-hospital arrest postop, 15mins to ROSC. Intubated, initially sedated, received TTM. On EEG after arrest for 3 days, no seizures. Had prior stroke with baseline left-sided weakness. Found to have small L-sided subdural hemorrhage. Weaned off sedation rapidly, started emerging from coma on day 2 but worsened due to hypotension. Consistently following commands by day 3, extubated by day 5, discharged on day 27 with CPC 3.
- Brainstem reflexes	Yes	Yes	Yes
- EEG pattern	I	I	I
- Spectral type	0	0	0
Patient 48				Age 40, man. Out-of-hospital arrest. Multiple arrests (>50) requiring shocks. Intubated, sedated, required ECMO, no TTM. Developed frequent bilateral arm shaking while sedated, myoclonic in nature. On EEG for 5 days, no seizures. Had prior strokes with prior R-sided weakness and cognitive deficits. Weaned off sedation rapidly, emerged on day 2 and following commands. Extubated on day 6, discharged on day 17 with CPC 3.
- Brainstem reflexes	Yes	Yes	Yes
- EEG pattern	I	I	I
- Spectral type	0	0	0

Limitations in EEG analysis

This study focused on the utility of spectral analysis and potential prognostic value of stratifying EEG patterns based on frequency components. In doing so, we did not examine in detail other potentially useful features on qualitative EEG analysis, such as reactivity and anterior-posterior gradient patterns. The prognostic value of qualitative analysis in our cohort may increase with these features, though it is likely that more patients would be needed to adjust for a more complex stratification system. Also, instead of using only data from morning and bedtime samples, more longitudinal analysis may produce more representative data on the most active state.

It is important to also note that the quantitative data simply represent a transformation of the qualitative data, and therefore the quantitative features are dependent on the raw qualitative data. The resulting stratification system, therefore, is also partly dependent on features from the raw EEG. The EEG reviewers were blinded to the clinical data, but they were not blinded to the raw EEG when analyzing the spectral data. In future studies, it would be more prudent to blind the EEG reviewers to the raw EEG when analyzing spectral data, to limit further bias.

Utility of detailed EEG analysis

Full prognostication testing can be challenging to obtain in the standard clinical setting. For the patients in this study, imaging was the testing modality most pursued (see Table 2), even though imaging is the most logistically difficult to obtain. Accuracy of imaging tests depends on timing, which is also true for NSE levels and SSEP. Neurologic exams can often be confounded by medications.

EEG is unique in that it can provide bedside, continuous, real-time data that is important not only for monitoring seizures but also for prognostication. Prior studies in literature have established certain malignant patterns as reliable predictors of poor outcomes, and this study yielded similar results. In addition, we importantly identified possible features linked to good outcomes. By stratifying EEG patterns according to the continuity of the background and degree of epileptiform activity, and by further stratifying the spectral content according to frequency components, we were able to strongly correlate EEG data with CPC scores at discharge. Our accuracy results were similar and in certain aspects better than what others reported using quantitative EEG methods^14–19.

For future studies, with a higher patient sample size, including more patients with good outcomes and more with BSP patterns, we would be able to model outcomes using regression or machine learning methods. The outcomes also need not be a dichotomous good or bad variable; it may be useful to delineate a moderate outcome category, i.e. a separate CPC 3 group. With more detailed quantitative analysis, it may be possible to refine these methods and achieve higher accuracy and earlier prognostication. Further studies will be required to determine the reliability of this approach, and reliable prediction of good and bad outcomes would be valuable in caring for this special population of vulnerable patients.