Early risk stratification for progression to death by neurological criteria following out-of-hospital cardiac arrest

Abstract

Background

Some patients resuscitated from out-of-hospital cardiac arrest (OHCA) progress to death by neurological criteria (DNC). We hypothesized that initial brain imaging, electroencephalography (EEG), and arrest characteristics predict progression to DNC.

Methods

We identified comatose OHCA patients from January 2010 to February 2020 treated at a single quaternary care facility in Western Pennsylvania. We abstracted demographics and arrest characteristics; Pittsburgh Cardiac Arrest Category, initial motor exam and pupillary light reflex; initial brain CT grey-to-white ratio (GWR), sulcal or basal cistern effacement; initial EEG background and suppression ratio. We used two modeling approaches: fast and frugal tree (FFT) analysis to create an interpretable clinical risk stratification tool and ridge regression for comparison. We used bootstrapping to randomly partition cases into 80% training and 20% test sets and evaluated test set sensitivity and specificity.

Results

We included 1,569 patients, of whom 147 (9%) had diagnosed DNC. Across bootstrap samples, >99% of FFTs included three predictors: sulcal effacement, and in cases without sulcal effacement, the combination of EEG background suppression and GWR ≤ 1.23. This tree had mean sensitivity and specificity of 87% and 81%. Ridge regression with all available predictors had mean sensitivity 91% and mean specificity 83%. Subjects falsely predicted as likely to progress to DNC generally died of rearrest or withdrawal of life sustaining therapies due to poor neurological prognosis. Two of these cases awakened from coma during the index hospitalization.

Conclusions

Sulcal effacement on presenting brain CT or EEG suppression with GWR ≤ 1.23 predict progression to DNC after OHCA.

Introduction

Even with optimal care, 10–15% of comatose survivors of out-of-hospital cardiac arrest (OHCA) progress to death by neurological criteria (DNC).[1, 2] Early reliable predictors of progression to DNC could inform clinical research, patient care, and shared decision making with surrogates. For example, most neuroprotective therapy trials in OHCA patients are neutral. Coma, defined as an inability to follow commands, is a common but imperfect inclusion criterion that is typically applied without additional stratification of brain injury severity. [3, 4] Early identification of patients that ultimately progress to DNC would allow for statistical adjustment or serve as inclusion criteria for trials targeting patients with the most severe injury. From a clinical perspective, expertise in management of severe anoxic injury and determination of DNC is not universal. Early risk stratification may aid in triage, interfacility transfer decision-making as well as early engagement of other support services for families. While early prognostic tests are imperfect, identifying high risk features could may help clinicians convey severity of brain injury to families and surrogates and potential courses. Since post-arrest patients comprise a large proportion of organ donors, reliable prediction of progression to DNC might also allow timely referral to organ procurement organizations to facilitate a smooth transition to post-mortem donor management.[1, 5]

Early brain computed tomography (CT) and electroencephalography (EEG) are established prognostic tools after cardiac arrest.[6] Cytotoxic edema can be quantified on head CT as a ratio of Hounsfield unit density of grey and white matter (GWR), or by qualitative assessment of mass effect on the ambient cisterns or sulci.[7] Reduction in GWR < 1.10 within 24 hours after arrest reliably predicts poor outcome with extremely low false positive rates.[8, 9] EEG is a rich, noninvasive neuromonitoring tool.[10, 11] Highly malignant EEG patterns including background suppression are a sign of severe brain injury.[12, 13] Prior studies have investigated clinical predictors of progression to DNC after OHCA,[14–16] but have not considered imaging or EEG. We analyzed a cohort of comatose OHCA patients to develop a clinical risk stratification tool to identify cases at high risk of progression to DNC. We hypothesized that features of brain CT and EEG obtained within 24 hours of arrest would identify patients at high risk for progression to DNC. Early identification of post arrest patients at risk for DNC criteria would facilitate control of illness severity in clinical trials, triage to centers capable of caring for patients with the most severe brain injury and expertise in brain death, guide expectations for families, and inform timely referrals to organ procurement organizations.

Methods

Patients and clinical setting

The University of Pittsburgh Human Research Protection Office approved this study as exempt from informed consent given minimal risk (protocol number 19020205). We included OHCA patients with return of spontaneous circulation treated at a single academic medical center between January 2010 and February 2020. The University of Pittsburgh Post-Cardiac Arrest Service (PCAS) managed all included patients and recorded clinical data in a prospective registry.[17–19] We excluded patients with a traumatic or neurological etiology of arrest given these are distinct risk factors for brain death compared to other arrest etiologies,[20, 21] patients who rapidly awakened within 6 hours of arrest, and patients who were transferred to our facility >24 hours after arrest.

From our registry, we abstracted patient demographics including age and sex; arrest characteristics including initial rhythm, witnessed status, number of defibrillator shocks, total duration of cardiopulmonary resuscitation (CPR), and number of intravenous or intraosseous epinephrine doses given during CPR. A PCAS provider examined all patients on arrival. We prospectively classified Full Outline of Unresponsiveness motor subscale on presentation, pupillary light reflex (present or absent), and illness severity, quantified as the Pittsburgh Cardiac Arrest Category.[19, 22] For this analysis, we classified arrest etiology as confirmed cardiac etiology (specifically acute coronary syndrome, primary dysrhythmia, left or right heart failure, or progression of cardiogenic shock) vs. non-cardiac etiologies.[18]

Neuroimaging and neurophysiology covariates

Our standard of care is to obtain a non-contrasted head CT shortly after presentation, typically during transport from the emergency department to the intensive care unit, using a GE Lightspeed VCT 64-channel scanner (120 kVp and 225mA). A blinded study coauthor (KF, JF, PJC) reviewed each CT to quantify GWR at the level of the basal ganglia with regions of interest placed over the caudate nucleus and posterior limb of the internal capsule. These selected regions of interest have similar sensitivity and specificity for poor neurological outcomes compared to other structures.[8] We also noted presence of sulcal and basal cistern effacement. We did not systematically manage patients with signs of cerebral edema with osmotic therapy or cerebrospinal fluid diversion.

Neurodiagnostic technologists applied gold-plated cup electrodes to the scalp in standard 10–20 International System of electrode placement. We started monitoring shortly after intensive care unit arrival and record data using XLTech Natus Neuroworks digital video/EEG systems (Natus Medical, Pleasanton, CA). For this analysis, we classified background activity for the first artifact free segment of EEG as continuous or discontinuous; burst suppression; or suppression using 2012 American Clinical Neurophysiology Critical Care EEG guidelines.[23] We calculated suppression ratio over the first 60 minutes of EEG monitoring using Persyst v13 at a voltage threshold of 2μV using the software’s native artifact reduction algorithms, and determined whole brain median values during this epoch.[24] We chose these measures as opposed to considering epileptiform or highly malignant characteristics because although these features predict outcome they also imply preservation of at least some cortical connectivity, function and blood flow.[10] By contrast, background suppression after cardiac arrest is associated with laminar cortical necrosis, persistently inadequate or absent intracranial blood flow and severe cytotoxic edema, all of which occur after severe primary brain injury and may result in progression to brain death.[25, 26] We calculated and report time from collapse to head CT and EEG initiation.

Determination of death by neurological criteria

Our hospital policy for determination of DNC followed the American Academy of Neurology and American Academy of Pediatrics guidelines.[27, 28] Our policy required determination of DNC >24 hours after both cardiac arrest and normothermia (when hypothermic temperature control is used), or earlier when an ancillary blood-flow study such as a four-vessel angiogram demonstrates absence of intracranial blood flow. In pediatric cases (age ≤ 18), examinations must be separated by an observation period of at least 12 hours and a confirmatory EEG is required.[29] For this analysis, we classified outcome for each patient as brain dead or non-brain dead (including survival to hospital discharge, rearrest or intractable shock and multisystem organ failure, withdrawal of life sustaining therapies (WLST) due to non-neurological reasons, or WLST for anticipated poor neurological prognosis).

Statistical analysis

We summarized cohort characteristics using mean with standard deviation, median with interquartile range, or number with corresponding percentage. We used multiple imputation with chained equations and generated 100 complete data sets.[30] We used a bootstrap to randomly partition cases into 80% training and 20% test sets 10,000 times for each of the following analyses and reported test set performance. First, we used a Fast and Frugal Tree (FFT) analysis to create a clinical risk stratification tool.[31] We reported the most common elements and configuration of trees over training samples. Second, we compared this clinical risk stratification tool to a regularized logistic regression with a ridge penalty using all available predictors.[32] We reported sensitivity and specificity at Youden’s point for both modeling approaches. For the ridge regression, we also reported pooled model coefficients and range, normalized to the standard deviation of each variable, sensitivity at 95% specificity and specificity at 95% sensitivity. To explore circumstances where FFTs consistently erred, we reported characteristics of subjects falsely predicted to progress to brain death in >20% of test sets. Finally, post hoc we repeated both FFT and ridge regression analysis with the addition of toxicological etiology of arrest and use of mechanical circulatory support (including venous arterial extracorporeal membrane oxygenation, Impella, or intra-aortic balloon pump). We used R (R Foundation for Statistical Computing, Vienna, Austria) for ridge regression (glmnet package) and FFT analysis (FFTrees package).[31, 33, 34]

Results

Subject characteristics

We treated 2,087 OHCA patients during the study period of whom 518 met exclusion criteria, and included 1,569 patients in our analysis (Supplemental figure 1). Mean age was 59 (standard deviation (SD) ± 16) years, 41% were female and 28% had a confirmed cardiac etiology of arrest (Table 1). Overall, 1,309 (83%) had a head CT acquired a median of 4.2 [interquartile range (IQR) 2.8 – 5.8] hours post-arrest. We acquired EEG in 940 patients (60%) initiated a median 9.5 [IQR 7.6 – 12.2] hours post-arrest. Overall, 9% of data were missing (Supplemental table 1). Additional details of subjects where CT brain or EEG was and was not obtained are listed in Supplemental table 2. A total of 147 subjects (9%) were diagnosed with brain death a median of 2 [IQR 1–3] days post-arrest.

Table 1.

Cohort characteristics.

Characteristic	All Subjects n= 1569	Survived to discharge n=369	Non-Brain Death n=1053	Brain Death n=147
Age, mean SD	59 (± 16)	56 (±16)	61 (± 16)	45 (± 15)
Female sex	644 (41%)	140 (38%)	428 (41%)	76 (52%)
Cardiac etiology	432 (28%)	179 (49%)	242 (23%)	11 (7%)
Initial rhythm
VT/VF	471 (30%)	212 (57%)	243 (23%)	16 (11%)
PEA	480 (31%)	85 (23%)	358 (34%)	37 (25%)
Asystole	523 (33%)	56 (15%)	387 (37%)	80 (54%)
Unknown	95 (6%)	16 (7%)	65 (6%)	14 (10%)
CPR duration, median IQR	22 [13 – 34]	14 [8 – 20]	24 [15 – 36]	32 [22 – 46]
Defibrillator shocks, median IQR	0 [0–2]	1 [0 – 3]	0 [0 – 2]	0 [0 – 1]
Epinephrine, median IQR	3 [2–5]	2 [0 – 3]	3 [2 – 6]	4 [3 – 6]
Pittsburgh cardiac arrest category
II	57 (23%)	219 (59%)	127 (12%)	11 (7%)
III	156 (10%)	64 (17%)	91 (3%)	1 (1%)
IV	930 (59%)	57 (16%)	743 (72%)	130 (89%)
Unknown	126 (8%)	29 (8%)	92 (3%)	5 (3%)
Non-reactive pupils	773 (50%)	64 (17%)	587 (55%)	122 (84%)
No motor response to pain	866 (59%)	104 (28%)	629 (60%)	133 (91%)
GWR, median IQR	1.30 [1.21–1.36]	1.35 [1.30 – 1.40]	1.29 [1.2 – 1.36]	1.10 [0.97 – 1.22]
Sulcal effacement	287 (22%)	4 (1%)	174 (20%)	109 (79%)
Basal cistern effacement	157 (12%)	0 (0%)	79 (9%)	78 (57%)
Initial EEG background
Discontinuous or continuous	315 (34%)	205 (79%)	109 (18%)	1 (1%)
Burst Suppression	398 (42%)	29 (11%)	352 (58%)	17 (23%)
Suppression	227 (24%)	25 (10%)	147 (24%)	55 (76%)
Suppression ratio, median IQR	57 [17 – 83]	11 [1 – 39]	67 [37 – 86]	84 [66 – 93]

Abbreviation: CT- computed tomography; CPR- cardiopulmonary resuscitation; EEG- electroencephalogram; GWR- grey white ratio; IQR- interquartile range; SD; standard deviation.

FFT and ridge regression models

Across training sets, >99% of FFTs included three predictors of DNC: presence of sulcal effacement, and in cases without sulcal effacement, the combination of EEG background suppression and GWR ≤ 1.23 (Figure 1). FFTs had a mean test set sensitivity and specificity of 87% (SD ± 6%) and 81% (SD ± 2%), respectively (Figure 2). With all available predictors, ridge regression had better performance (91% sensitivity (SD ± 5%), 83% specificity (SD ± 4%)) (Table 2, Figure 2). At 95% specificity, ridge regression had moderate sensitivity (77%, SD ± 1); at 95% sensitivity, specificity was poor (59%, SD ± 1).

Figure 1.

Fast and frugal tree with sample test set data performance.

CT- computed tomography; EEG- electroencephalograph; FFT- fast and frugal tree

Figure 2.

Bootstrap sensitivity and specificity of fast and frugal tree and ridge regression models.

Blue dotted line- mean value

Table 2:

Ridge regression model coefficients.

	Coefficient			Odds Ratio
Characteristic	Mean	Min	Max	Mean	Min	Max
Age	−0.0021	−0.0025	−0.0015	0.9652	0.9575	0.9753
Female sex	0.4703	−0.2016	0.9491	1.2604	0.9056	1.5951
Cardiac etiology	−0.4821	−1.4796	0.0586	0.8063	0.5163	1.0265
Initial rhythm
VT/VF	ref	-	-	-	-	-
PEA	0.2863	−0.1535	0.9657	1.1445	0.9302	1.5761
Asystole	0.3622	−0.4173	0.8775	101895	0.8188	1.5225
CPR duration	0.0008	0.0002	0.0012	1.0139	1.0038	1.0237
Defibrillator shocks	−0.0350	−0.0589	−0.0071	0.9226	0.8729	0.9837
Epinephrine	−0.0365	−0.0592	−0.0174	0.9027	0.8468	0.9522
Pittsburgh cardiac arrest category
II	ref	-	-	-	-	-
III	−2.9967	−3.2567	−1.5186	0.3935	.03628	0.6232
IV	0.0955	−0.8161	0.7320	1.0469	0.6767	1.4196
Non-reactive pupils	0.2696	−0.1182	0.9797	1.1444	0.9426	1.6320
No motor response to pain	1.2072	0.5650	1.8222	1.8151	1.3218	2.4591
GWR *	−1.2152	−1.7000	−0.3089	1.0045	1.0002	1.0113
Sulcal effacement	1.539	1.0897	3.1032	1.8741	1.5600	3.5478
Basal cistern effacement	3.0241	1.9627	4.3805	2.6097	1.8636	4.0121
Initial EEG background
Continuous or discontinuous	ref	-	-
Burst suppression	0.1139	−0.3083	0.6389	1.0577	0.8593	1.3692
Isoelectric suppression	2.3459	1.8913	2.8816	2.9136	2.3682	3.7192
Suppression ratio	0.0001	0.00007	0.0003	1.0045	1.0002	1.0113

Abbreviation: CPR- cardiopulmonary resuscitation; EEG- electroencephalogram; GWR- grey white ratio.

^*coefficient per 0.10 increase in GWR.

FFT model outliers

Overall, 202 subjects were falsely predicted as likely to progress to DNC in >20% of FFT test sets. Most (196 subjects) died in the hospital: 30% (n=59) rearrested or succumbed to multisystem organ failure prior brain death testing; 62% (n=122) had WLST due to poor neurological prognosis prior to brain death testing; and the remaining (8%, n=15) had WLST due to pre-existing advanced directives. Median length of stay for non-survivors in this group was 1 [IQR 1–2] day. Greater than 70% of patients that survived to hospital day 1 had nonreactive pupils, absent cough reflex, and no response to pain or myoclonus at 24 hours post-arrest. Six of 202 frequently misclassified subjects survived to hospital discharge, of whom two awakened from coma (Supplemental table 3, 4, supplemental figure 2). Post hoc model predictors, coefficients and model performance did not meaningfully differ from our main models (data not shown).

Discussion

We developed a simple risk stratification tool to identify patients likely to progress to DNC after OHCA. The incidence of DNC in our cohort was 9%, similar to prior published studies.[1] Patients falsely predicted to progress to DNC by our risk stratification tool most often died from multisystem organ failure and early WLST based on presumed poor neurological outcome or preexisting advanced directives. It is likely that some of these patients would ultimately have been pronounced DNC if they had not first succumbed to other causes, which would have improved model specificity. Two patients predicted to progress to DNC based on initial CT imaging and EEG awakened from coma and survived to hospital discharge. This is consistent with consensus guidelines that assert that no predictors of poor outcome short of death itself are perfectly specific in the first days after cardiac arrest.[35, 36] We advocate for aggressive intensive care and management even in cases at high probability of progression to DNC, provided this care aligns with patient values and preferences, and triage of patients at high risk of DNC to centers with expertise in its determination.[37]

FFT predictors also had large coefficients in our regression models and are biologically plausible indicators for severe hypoxic-ischemic injury. Sulcal effacement observed on head CT reflects displacement of cerebral spinal fluid and compression of brain to the bony skull secondary to cerebral edema with mass effect. After cardiac arrest, background suppression signifies severe global cortical dysfunction, synaptic failure, and deranged neuronal metabolism. Influx of water into neurons reduces their radiodensity, effectively narrowing the measured GWR. These findings correlate with diffuse supratentorial and infratentorial neuronal necrosis on brain autopsy,[12] though both EEG and CT have limitations that reduce both sensitivity and specificity predicting poor or favorable outcome as confounders may also affect results absent structural brain injury. While rare, sustained hypercapnia can precipitate cerebral vasodilation and diffuse cortical edema that can be reversable.[38] ] At high doses, sedatives, anesthetics and other neuroactive agents can induce reversible generalized EEG suppression and neurological examination findings mimicking brain death.[39] Despite these limitations, the predictors identified by our FFT have some advantages to those explored in past studies. Prior work has focused on demographics, arrest characteristics, measures of shock severity, and etiology of arrest.[14, 40] Arrest characteristics may not be known reliably on presentation.[41, 42] Given the heterogeneity of post-arrest illness, vasopressor requirement (a predictor in some models) is expected to be less specific than direct markers of brain injury. Thus, a risk stratification tool based on objective measures acquired early during hospitalization is appealing.

Our risk stratification tool may be useful to guide families through the anticipated clinical course of severe post-arrest brain injury. Clinicians may expect hemodynamic and other organ instability associated with progression to DNC.[43] Availability of organs for transplant is outpaced by demand and cardiac arrest patients with irrecoverable brain injury comprise a large and growing population of potential organ donors.[5, 44] Proactive management of brain death physiology may prevent cardiovascular collapse and rearrest, and in cases where organ donation is pursued, increase organ yield and improve post-transplant graft function.[45] Moreover, early notification for potential to progress to DNC may invite earlier involvement of an organ procurement organization and optimize timing of approach for organ donation consent.[46]

While regression modeling had better performance, the model was much more complex and used variables that may be difficult verify within the first day after arrest. When considering implementing in a clinical setting, we favor our risk stratification tool for simplicity at the cost of mild decrement in sensitivity and specificity. Our regression model with highly sensitive cutoff was only somewhat specific, but may be useful for screening patients for low-risk interventions such as organ procurement organization engagement. A highly specific cutoff may be useful in the setting such as considering a patient for escalating mechanical circulatory support or other invasive interventions.

Our work has limitations. Study authors adjudicated and calculated CT features. These measures have a degree of inter-rater disagreement.[37, 47] We measured GWR with regions of interest in the caudate nucleus and posterior limb of the internal capsule. Our approach may be falsely pessimistic in cases of caudate infarction but preserved GWR in other regions, though these regions have similar sensitivity and specificity for predicting poor outcome compared to other regions.[8] Moreover, Automated review of non-contrast CTs have been an active area of acute stroke imaging research,[48] and calculation of GWR is possible though require further clinical validation.[49, 50] Future algorithms may utilize automated interpretations of GWR, sulcal and cistern effacement, or uncover other signatures of severe anoxic injury of prognostic value. We reviewed studies on a continuous basis and utilized the 2012 American Clinical Neurophysiology Society guidelines for EEG interpretation and quantified background continuity with suppression ratio.[51] Quantitative EEG analysis may not perfectly correlate with expert review.[52] We included OHCA in the current study, and so it is unclear how this tool would perform for in hospital cardiac arrest. This is also a single center study, limiting generalizability. Finally, we did not have granular detail of aggressiveness of osmotic or other therapies aimed at reducing intracranial pressure.

Conclusions

Early brain CT and EEG allowed for reliable prediction of progression to DNC after OHCA when compared to a model with many additional characteristics. These tests are widely available and are used commonly in post-arrest prognostication. Our risk stratification tool can be useful to anticipate the likely course of severe anoxic brain injury, identify potential organ donation opportunities and control for illness severity in future clinical trials.