Abstract
Background
Grey-white ratio (GWR) can estimate severity of cytotoxic cerebral edema secondary to hypoxic-ischemic brain injury after cardiac arrest and predict progression to death by neurologic criteria (DNC). Current approaches to calculating GWR are not standardized and have variable interrater reliability. We tested if measures of variance of brain attenuation on early computed tomographic (CT) imaging after cardiac arrest could predict DNC.
Methods
We performed a retrospective cohort study, identifying post-arrest patients treated between 2011 and 2020 at our single center. We extracted demographic data from our registry and Digital Imaging and Communication in Medicine (DICOM) files for each patient’s first brain CT. We analyzed slices 15–20 of each DICOM, corresponding to the level of the basal ganglia while accommodating differences in patient anatomy. We extracted pixel arrays and converted the radiodensities to Hounsfield units (HU). To focus on brain tissue densities, we excluded HU >60 and <10. We calculated the variance of each patient’s HU distribution and difference between the means of a two-group Gaussian finite mixture model as proxies for manually calculated GWR. We compared these metrics to existing measures of cerebral edema, then randomly divided our data into 80% training set and 20% test sets and used logistic regression to predict DNC.
Results
Of 1,133 included patients, 457 (40%) were female, mean (standard deviation) age was 58 (16) years, and 115 (10%) progressed to DNC. CTs were obtained a median [interquartile range] of 4.2 [2.8–5.7] hours post-arrest. Our novel measures correlated weakly with GWR but differed significantly between patients with and without qualitative features of cerebral edema. HU variance, but not difference between mixture model means, differed significantly between subjects with and without sulcal or cistern effacement. GWR outperformed our novel measures in predicting progression to DNC with an area under the receiver operating characteristic curve (AUC) of 0.82, compared to HU variance (AUC = 0.73) and the difference between mixture model means (AUC = 0.56).
Conclusion
There are differences in the distribution of HU on post-arrest CT in patients with qualitative measures of cerebral edema. Current methods to quantify cerebral edema outperform simple measures of attenuation variance on early brain CT. Further analyses could investigate if these measures of variance, or other distributional characteristics of brain attenuation, have improved predictive performance on brain CTs obtained later in the clinical course or derived from discrete regions of anatomical interest.
Keywords: Cardiac arrest, anoxic brain injury, hypoxic-ischemic brain injury, computed tomography, brain death, heart arrest, outcome prediction
Introduction
Annually in the United States, over 350,000 people suffer out-of-hospital cardiac arrest (OHCA) [1]. Among patients who survive to hospital admission, hypoxic-ischemic brain injury (HIBI) drives morbidity and mortality [2]. Severe HIBI results in cerebral edema with mass effect, impaired cerebral blood flow, herniation, and death by neurologic criteria (DNC) [3], which accounts for approximately 10% of all deaths among those hospitalized after OHCA [4].
Cytotoxic cerebral edema severity can be estimated based on computed tomography (CT) findings using both quantitative and qualitative approaches [5]. Under healthy conditions, cellular grey matter and myelin-rich white matter have different radiodensities [6]. The distinction between grey and white matter can be assessed globally or through measurement at well-defined anatomic regions of interest (ROI). When quantified, the difference is often expressed as the ratio of grey matter to white matter radiodensity (GWR) [7]. Cytotoxic cerebral edema after cardiac arrest is caused by neuronal metabolic failure [3]. Influx of sodium and water into neuronal bodies reduces attenuation of grey matter and thus reduces GWR. As edema progresses, mass effect results in effacement of the cisterns and sulci and may lead to brain herniation [8].
American Heart Association and European Resuscitation Council guidelines support obtaining CT imaging to estimate cerebral edema severity for neurologic prognostication [9, 10]. As the largest cause of death amongst comatose patients resuscitated from OHCA with HIBI is withdrawal of life-sustaining therapies (WLST) for perceived poor neurologic prognosis [11], objective and consistent techniques for cytotoxic cerebral edema quantification are mandatory. Current methods to calculate GWR are not standardized. Variation in ROI anatomic location produces inconsistent thresholds at which GWR serves as a specific predictor of poor outcome [12]. Furthermore, manually selecting ROIs may produce inter- and intra-rater variability [13].
Previous analyses have attempted to address these limitations using brain imaging segmentation to identify ROIs objectively and consistently [14, 15]. It is unclear how well segmentation algorithms perform on images with anatomic changes secondary to HIBI or images obtained from different scanners [16]. Additionally, this method is time-consuming and computationally intensive [17], limiting its use in acute care. Accordingly, we sought to identify simple, objective, and interpretable methods to quantify cytotoxic cerebral edema on early brain CT in comatose patients resuscitated from OHCA.
It is plausible that a radiographically normal brain would yield two distinct distributions of Hounsfield units (HU) corresponding to grey and white matter. In contrast, the distributions of HU in a brain with edema secondary to HIBI are expected to be less variable. We hypothesized that descriptive statistics or a two-group Gaussian finite mixture model could distinguish edematous from radiographically normal brain, and that these novel radiographic markers of cytotoxic cerebral edema would correlate with GWR and predict progression to DNC (Figure 1).
Figure 1.
Hypothesized and actual Hounsfield unit distributions.
We hypothesized that a radiographically normal brain would have two distinct distributions of HU (A). Alternatively, an edematous brain would have less-distinct distributions (B). We hypothesized that these differences could be quantified using variance (C, D) and the difference between distribution means in a finite mixture model (E, F). Example patient data are shown in (G, H), which deviate from our expected distributions.
Methods
Patients and Clinical Setting
The University of Pittsburgh Human Research Protection Office approved this research (STUDY19020205). We performed a retrospective cohort study including comatose survivors of OHCA treated at a single academic medical center between January 2011 and March 2020. We screened all OHCA patients in our registry during the study period and excluded those for whom initial imaging was performed more than 24 hours after arrest, those who underwent imaging after contrast injection (e.g., imaging after coronary angiography) expected to alter brain attenuation, and those with a traumatic or neurologic etiology of arrest for whom imaging changes might reflect acute pathology other than HIBI. We did not exclude patients with pre-existing brain pathologies systematically but excluded one patient with outlying attenuation metrics found post hoc that were found to be secondary to extreme hydrocephalus. We defined coma as a presenting Full Outline of UnResponsiveness (FOUR) motor score of 0–3 [18]. We abstracted patient demographics, arrest characteristics, and proximate cause of death -- defined as withdrawal for perceived poor neurological prognosis, withdrawal for none-neurological reasons, brain death, or re-arrest – from our prospective registry [19–21].
Neuroimaging
A Post-Cardiac Arrest Service provider evaluates OHCA patients on arrival to our emergency department, provides ongoing resuscitation, and guides the early diagnostic evaluation. Our standard workup includes a non-contrasted brain CT, typically obtained upon disposition from the emergency department. A GE Lightspeed VCT 64-channel scanner (120 kVp, 225 mA, 5mm slice thickness) was used throughout the entire study period. We recorded the time from arrest, as documented on the emergency medical services patient care report, to brain CT. As part of our prospective registry, trained examiners review each brain CT to measure GWR manually and qualitatively assess for sulcal and cistern patency. We calculate manual GWR from a single CT slice at the level of the basal ganglia with ROIs placed over the caudate nucleus and posterior limb of the internal capsule (Supplemental Figure 1).
Image Processing
We collected source Digital Imaging and Communications in Medicine (DICOM) brain CT image files for each patient. In cases where multiple studies were obtained, we included the study obtained at our medical center most proximate to the cardiac arrest. We selected slices 15–20 to include the level of the basal ganglia while accommodating individual differences in patient anatomy. We extracted pixel arrays from the DICOMs and converted grayscale intensities to HU. To capture relevant brain tissue while excluding tissue types that were not of interest (i.e., bone and cerebrospinal fluid) from the individual subject HU arrays, we selected one radiographically normal brain and recreated CT images with different ranges of HU (Supplemental Figure 2). Two co-investigators (J.E. and P.J.C.) visually inspected the ranges and selected densities > 10 and < 60 HU as including brain tissue of interest and excluded other non-brain tissue types for analysis. We filtered all subject HU arrays for densities > 10 and < 60. We used the Pydicom package in Python (Python Software Foundation, https://www.python.org) for image processing.
For each patient, we calculated two novel metrics. First, we pooled HU values across slices and calculated the variance of the HU distribution. Second, we fit a two-group Gaussian finite mixture model to the HU distribution for each patient using the package scikit-learn. We then calculated the difference between distributions’ means.
Primary and Secondary Outcomes
Previous studies have evaluated GWR to predict survival or functional status [12]. Withdrawal of life-sustaining therapies (WLST) may lead to poor outcome in patients with radiographically normal brains and recoverable HIBI. Moreover, clinicians may choose WLST based on results of neuroimaging, resulting in models that reinforce existing clinical practice patterns [22]. Consequently, we selected DNC as our outcome to identify cases where the ground truth of severe, irrecoverable neurological injury was certain. Patients that died due to other causes or survived were considered non-DNC. Because patients with severe HIBI who with time would progress to DNC may die first from multisystem organ failure or have WLST (i.e., DNC is not observed despite imaging that is consistent with non-survivable injury), we performed a secondary analysis including only the subset of patients for whom a ground truth outcome was known with certainty: patients with DNC and those who awakened from coma to follow verbal commands. Determination of DNC was performed according to hospital policy consistent with American Academy of Neurology guidelines [23]
Novel Measure Simulations
It is likely that our novel measures are influenced by factors other than grey and white matter. To quantify the degree to which unmeasured factors may influence these measurements, we performed simulations to determine the behavior of these measures under ideal conditions and compared them to values derived from DICOM images.
First, we produced a model with only two discrete values: 35 (corresponding to the average HU of grey matter) and 25 (corresponding to the average HU of white matter). For this initial model, we generated 100,000 total values, with 40% equal to 35 and 60% equal to 25. Next, to include the natural variability of grey and white matter densities, we created a second model using two normal distributions. Of 100,000 total values, 60% represented grey matter with mean centered at 35 and standard deviation of 3, with the remaining 40% representing white matter centered at 25 with standard deviation of 3. We considered this our “radiographically normal” model. Finally, to model a brain with “severe edema”, we performed a similar procedure as our second model but with the grey matter distribution being centered at 29. We report variance and difference in finite mixture model means under each of these conditions.
Statistical Analysis
We calculated descriptive statistics to summarize baseline characteristics and outcomes. We report mean and standard deviation (SD) for normally distributed data, median and interquartile range [IQR] for continuous data, and number with corresponding percentage for categorical data. We calculated the Pearson correlation coefficient of our novel metrics with manually measured GWR and used t-tests to compare all HU distribution metrics between patients with and without qualitative features of edema such as effacement of the basal cisterns or sulci.
To predict DNC, we randomly partitioned the cohort into 80% training and 20% test sets and used logistic regression to quantify the association of each HU distribution characteristic and qualitative feature with outcome. We evaluated model performance in the test set data as the area under the Receiver Operator Characteristic (AUROC) curves. We compared the AUROCs of our novel measures to manual GWR and qualitative features of edema. We performed the same procedure to predict DNC in the secondary analysis cohort and visually compared the AUROCs to the those from the primary analysis. We used R for these simulation and statistical analyses (R Foundation for Statistical Computing, Vienna, Austria) [24].
Results
Subject Characteristics
We treated 2,556 patients during the study period, of whom 1,133 were included (Figure 2). Mean age was 58 (SD +/− 16) years, and 457 (40%) were female (Table 1). Brain CTs were acquired a median of 4.2 (IQR 2.8 – 5.7) hours post-arrest. Overall, 265 subjects (23%) awakened in hospital, 277 (24%) survived to hospital discharge, 550 (49%) had WLST, 191 (17%) rearrested or died due to multisystem organ failure, and 115 (10%) progressed to DNC.
Figure 2.
STROBE diagram.
Table 1.
Cohort characteristics.
| Characteristic | All Subjects n = 1133 |
Survived to discharge n = 277 |
Non-Brain Death n = 1018 |
Brain Death n = 115 |
|---|---|---|---|---|
| Age, mean SD | 58 (± 16) | 57 (± 17) | 60 (± 16) | 43 (± 15) |
| Female sex | 457 (68%) | 110 (40%) | 400 (39%) | 57 (50%) |
| Race | ||||
| White | 822 (73%) | 204 (74%) | 740 (73%) | 82 (71%) |
| Black | 123 (11%) | 47 (17%) | 112 (11%) | 11 (10%) |
| Asian | 6 (0.5%) | 2 (0.5%) | 5 (0.5%) | 1 (0.9%) |
| American Indian or Alaskan Native | 8 (0.5%) | 2 (0.5%) | 8 (0.5%) | 0 (0%) |
| Native Hawaiian or Other Pacific Islander | 0 (0%) | 0 (0%) | 0 (0%) | 0 (0%) |
| Unknown | 174 (15%) | 22 (8%) | 153 (15%) | 21 (18%) |
| Ethnicity | ||||
| Hispanic or Latino | 2 (0.1%) | 1 (0.4%) | 1 (0.09%) | 1 (0.9%) |
| Non-Hispanic or Latino | 895 (79%) | 223 (81%) | 802 (79%) | 93 (81%) |
| Unknown | 236 (21%) | 53 (19%) | 215 (21%) | 21 (18%) |
| Cardiac etiology | 301 (26%) | 126 (45%) | 293 (29%) | 8 (7%) |
| ACS | 147 (13%) | 66 (24%) | 143 (14%) | 4 (3%) |
| Intrinsic dysrhythmia | 28 (2%) | 16 (6%) | 26 (3%) | 2 (2%) |
| Dysrhythmia secondary to cardiomyopathy | 76 (7%) | 36 (13%) | 76 (7%) | 0 (0%) |
| Structural heart disease | 3 (0.3%) | 2 (0.7%) | 3 (0.3%) | 0 (0%) |
| LV failure | 12 (1%) | 1 (0.4%) | 12 (1%) | 0 (0%) |
| RV failure | 35 (3%) | 5 (2%) | 33 (3%) | 2 (2%) |
| Initial rhythm | ||||
| VT/VF | 321 (28%) | 145 (52%) | 310 (30%) | 11 (10%) |
| PEA | 341 (30%) | 68 (25%) | 312 (31%) | 29 (25%) |
| Asystole | 397 (35%) | 48 (17%) | 334 (33%) | 63 (55%) |
| Unknown | 74 (7%) | 16 (6%) | 62 (6%) | 12 (10%) |
| Witnessed status | ||||
| Unwitnessed | 319 (28%) | 48 (17%) | 257 (25%) | 62 (54%) |
| Layperson witnessed | 547 (48%) | 143 (52%) | 509 (50%) | 38 (33%) |
| EMS witnessed | 152 (13%) | 39 (14%) | 142 (14%) | 10 (9%) |
| In ED arrest | 115 (10%) | 47 (17%) | 110 (11%) | 5 (4%) |
| CPR duration, median IQR | 20 [11–34] | 13 [7–20] | 20 [11–32] | 32 [19–41] |
| 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 [2–5] | 4 [3–6] |
| Pittsburgh cardiac arrest category | ||||
| II | 254 (22%) | 159 (57%) | 248 (24%) | 6 (5%) |
| III | 108 (10%) | 44 (16%) | 107 (11%) | 1 (0.9%) |
| IV | 692 (61%) | 49 (18%) | 588 (58%) | 104 (90%) |
| Unknown | 79 (7%) | 25 (9%) | 75 (7%) | 4 (3%) |
| Initial FOUR motor score, median IQR | 0 [0–3] | 2 [0–3] | 0 [0–3] | 0 [0–0] |
| Hours from arrest to CT, median IQR | 4.2 [2.8–5.7] | 3.7 [2.1–5.4] | 4.2 [2.8–5.7] | 4.5 [3.3–5.6] |
| GWR, median IQR | 1.30 [1.22–1.37] | 1.30 [1.35–1.40] | 1.31 [1.24–1.37] | 1.22 [0.96–1.54] |
| Sulcal effacement | 229 (20%) | 3 (1%) | 137 (13%) | 92 (80%) |
| Basal cistern effacement | 121 (11%) | 0 (0%) | 56 (6%) | 65 (57%) |
| HU variance, median IQR | 92 [82–104] | 94 [85–107] | 93 [83–105] | 83 [75–95] |
| Difference between mixture model means, median IQR | 11.1 [9.4–12.6] | 11.1 [9.7–12.9] | 11.1 [9.4–12.7] | 10.9 [9.1–12.4] |
| Awakened in hospital | 265 (23%) | 44 (16%) | 265 (26%) | 0 (0%) |
CPR – Cardiopulmonary Resuscitation; FOUR – Full Outline of UnResponsiveness; CT – Computed Tomography; GWR – Grey-White Matter Ratio
Novel Measure Simulations
The HU variance in our simulated models was significantly lower in the model of severe edema (13) than in the normal model (33). The difference between mixture model means was equal to the difference in the means of the simulated distributions (Discrete values model = 10, Radiographically normal model = 10, Severe edema model = 4).
Correlation between Novel Measures and Manual GWR
HU variance and manual GWR were weakly correlated (Pearson’s r = 0.22, p < 0.01). The difference between mixture model means and manual GWR were very weakly correlated (Pearson’s r = 0.11, p < 0.01) (Supplemental Figure 3).
Association of Novel Measures with Qualitative Features of Cerebral Edema
Subjects with sulcal effacement had significantly lower GWR (1.06 (0.17) vs. 1.32 (0.12), P < 0.01) and HU variance (86 (18) vs. 94 (18), P < 0.01). Similarly, patients with basal cistern effacement had lower GWR (1.04 (0.15) vs. 1.29 (0.15), P < 0.01) and HU variance (86 (17) vs. 94 (18), P < 0.01). The difference between mixture model means was not significantly different between patients with sulcal effacement (10.7 (2.9) vs. 10.9 (2.3), p = 0.12) or basal cistern effacement (10.8 (2.8) vs. 10.9 (3.0), p = 0.37) (Figure 3).
Figure 3.
Association of Novel Measures with Qualitative Features of Cerebral Edema. Subjects with qualitative features of cytotoxic cerebral edema had significantly lower GWR and HU variance but did not have a significant divergence in the difference between mixture model means.
Prediction of Progression to Death by Neurologic Criteria
GWR had a higher AUC when predicting progression to DNC in test data than HU variance or the difference between mixture model means (GWR AUC = 0.82, HU variance AUC = 0.73, mixture model mean difference AUC = 0.56) (Figure 4). The presence of sulcal effacement (AUC = 0.80) and cistern effacement (AUC = 0.76) were associated with DNC. Results were comparable when we restricted the cohort to only subjects who were diagnosed with DNC or awakened (data not shown).
Figure 4.
Prediction of Progression to Death by Neurologic Criteria.
Discussion
Using simple statistics, we developed novel biomarkers of cytotoxic cerebral edema severity from brain CTs obtained from comatose OHCA patients. These approaches did not outperform manually calculated GWR in predicting progression to DNC.
While manual GWR calculation was performed on a single slice with pre-defined regions of interest, we calculated metrics using multiple slices. The larger number of HU observations could obscure smaller changes in radiodensities in specific anatomic regions that would otherwise be apparent using manually placed ROIs. Although we selected HU ranges to include grey and white matter, other tissues captured in this density range certainly contribute to the overall distribution of HU. This is likely why our simulated values for these novel measures under ideal circumstances diverged from the observed measures and did not perform as well as GWR.
Our work suggests more sophisticated methods are needed for fully automated methods to outperform current approaches. Kawai and colleagues demonstrated that a convolutional neural network performed similarly to GWR at predicting outcome in their cohort [25]. While further work is required to evaluate the generalizability of this method, such black-box methods are not interpretable by clinicians and are rarely adopted in clinical practice [26].
The quality of evidence for neurological prognostication in comatose patients resuscitated from cardiac arrest is low. The American Heart Association has offered a scientific statement recommending standards for the design of future studies of neurological prognostication [27]. Among weaknesses identified, simple dichotomization of observed patient outcome as good or poor is key. Most deaths after OHCA are preceded by WLST informed by perceived neurologic outcome [11]. In acutely brain injured patients, WLST almost always results in death and the counterfactual outcome (i.e., recovery potential given continuation of life-sustaining therapies) is not observed. In this scenario, prognostic models may perpetuate self-fulfilling prophecies [28].
Evidence suggests clinicians act upon prognostic tests in real time. Beekman and colleagues performed a retrospective analysis of changes in post-arrest treatments after early brain imaging [22]. Almost half of patients had a change in treatment plan based solely on interpretation of the imaging data, often leading to limitation or WLST. The investigators compared severity of HIBI between the interpreting neuroradiologist and three board-certified neurointenstivists blinded to clinical outcome. There was only slight agreement between neuroradiologist interpretation and neurointenstivists as to the presence of HIBI (kappa = 0.198). Agreement on severity of HIBI was noted between all three neurointenstivists in only one in four of brain CTs.
To address this potential source of bias, our outcome of interest was DNC versus all other potential outcomes, including survival and death from other causes. We selected this outcome to identify cases where the ground truth of severe brain injury is known. In this cohort, 10% of the patients progressed to DNC, consistent with prior research [4]. Use of DNC as an outcome also has important limitations. We have shown that patients predicted to be at high risk for DNC often succumbed to multisystem organ failure or WLST prior to DNC determination [29]. Thus, some patients without DNC observed are likely to have initial neuroprognostic findings suggestive of irrecoverable HIBI. We performed a secondary analysis limited to patients whose recovery potential was known with certainty: those who awakened from coma and those with DNC. Future studies of prognostic tools should consider the impact of these inherently flawed outcomes on model performance [30].
Limitations
In addition to an imperfect outcome measure, our study has important limitations. Subjects were treated at a single center where standard practice included early CT imaging. As edema evolves over time, early imaging may not capture the full extent of HIBI. Analysis of these measures on brain CTs obtained later in a patient’s course may yield difference results. While our manual GWR calculations and assessments of sulcal and cistern patency are performed by trained reviewers, these measures have not been assessed for inter- and intra-rate reliability. The standard workflow for neuroimaging analysis employs several pre-processing steps that require knowledge of radiographic data manipulation and statistical computing [31]. We sought to identify a method of quantifying the severity of post-arrest HIBI on brain CT that would be both easy to obtain with open-source software and interpretable by a bedside clinician. Therefore, we elected to use simple descriptive statistics as a novel method of brain CT image analysis. This method has not been described previously and our use of a standard preprocessing workflow may not have yielded optimal inputs for analysis. Moreover, the HU range used for analysis was selected by visual inspection of the source imaging files, rather than a data-driven approach to select the optimal HU range.
Conclusions
Current methods for manual quantification of cytotoxic cerebral edema severity outperform simple measures of variance of early brain CT data. Further analyses could investigate if these measures of variance, or other distributional characteristics of brain attenuation, have improved predictive performance on brain CTs obtained later in the clinical course or derived from discrete regions of interest.
Supplementary Material
Table 2.
Associations and Performance of Novel Measures.
|
Sulcal Effacement n = 229 |
No Sulcal Effacement n = 904 |
p-value | |
|---|---|---|---|
| GWR | 1.06 (± 0.17) | 1.32 (± 0.12) | < 0.01 |
| HU variance | 86 (± 18) | 94 (± 18) | < 0.01 |
| Difference between MM means | 10.7 (± 2.9) | 10.9 (± 2.3) | 0.18 |
|
Basal Cistern Effacement n = 121 |
No Basal Cistern Effacement n = 1012 |
p-value | |
|---|---|---|---|
| GWR | 1.04 (± 0.15) | 1.29 (± 0.15) | < 0.01 |
| HU variance | 86 (± 17) | 94 (± 18) | < 0.01 |
| Difference between MM means | 10.8 (± 2.8) | 10.9 (± 3.0) | 0.55 |
| AUC (95% CI) | OR (95% CI) | p-value | |
|---|---|---|---|
| GWR | 0.82 (0.72–0.92) | 0.002 (0.00–0.01) | < 0.01 |
| HU variance | 0.73 (0.61–0.85) | 0.98 (0.96–0.99) | < 0.01 |
| Difference between MM means | 0.56 (0.44–0.69) | 0.97 (0.90–1.04) | 0.38 |
| Culcal effacement | 0.80 (0.70–0.90) | 26.38 (15.33–47.47) | < 0.01 |
| Cistern effacement | 0.76 (0.65–0.86) | 19.04 (11.63–31.60) | < 0.01 |
GWR – Grey-White Matter Ratio; HU – Hounsfield Unit; MM – Mixture Model; AUC – Area under the Receiver Operating Characteristic Curve; CI – Confidence Interval; OR – Odds Ratio
Acknowledgements:
We acknowledge Valerie Arias, MD, Joanna Fong-Isariyawongse, MD, Mark Schmidhofer, MD, and Amy Wagner, MD, who provided care to patients during this study. Figure 1 was created with BioRender.com.
Disclosures:
Dr. Elmer recieves research funding from the National Institutes of Health to study brain recovery after cardiac arrest (R01NS124642; R01NS119825). Dr. Callaway recieves research funding from the National Institutes of Health to study emergency treatments after cardiac arrest (U24NS100659; U24NS100656
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Declaration of Interest
Two co-investigators, C.W.C. and J.E., are part of the Editorial Board of Resuscitation.
Declaration of Interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Bibliography
- 1.Tsao CW, et al. , Heart Disease and Stroke Statistics—2023 Update: A Report From the American Heart Association. Circulation, 2023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Perkins GD, et al. , Brain injury after cardiac arrest. The Lancet, 2021. 398(10307): p. 1269–1278. [DOI] [PubMed] [Google Scholar]
- 3.Sandroni C, Cronberg T, and Sekhon M, Brain injury after cardiac arrest: pathophysiology, treatment, and prognosis. Intensive Care Medicine, 2021. 47(12): p. 1393–1414. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Sandroni C, et al. , The rate of brain death and organ donation in patients resuscitated from cardiac arrest: a systematic review and meta-analysis. Intensive Care Med, 2016. 42(11): p. 1661–1671. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Keijzer HM, et al. , Brain imaging in comatose survivors of cardiac arrest: Pathophysiological correlates and prognostic properties. Resuscitation, 2018. 133: p. 124–136. [DOI] [PubMed] [Google Scholar]
- 6.Weinstein MA, Duchesneau PM, and MacIntyre WJ, White and Gray Matter of the Brain Differentiated by Computed Tomography. Radiology, 1977. 122(3): p. 699–702. [DOI] [PubMed] [Google Scholar]
- 7.Metter RB, et al. , Association between a quantitative CT scan measure of brain edema and outcome after cardiac arrest. Resuscitation, 2011. 82(9): p. 1180–1185. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Hayman EG, et al. , Cerebral Edema After Cardiopulmonary Resuscitation: A Therapeutic Target Following Cardiac Arrest? Neurocrit Care, 2018. 28(3): p. 276–287. [DOI] [PubMed] [Google Scholar]
- 9.Panchal AR, et al. , Part 3: Adult Basic and Advanced Life Support: 2020 American Heart Association Guidelines for Cardiopulmonary Resuscitation and Emergency Cardiovascular Care. Circulation, 2020. 142(16_suppl_2): p. S366–S468. [DOI] [PubMed] [Google Scholar]
- 10.Nolan JP, et al. , European Resuscitation Council and European Society of Intensive Care Medicine Guidelines 2021: Post-resuscitation care. Resuscitation, 2021. 161: p. 220–269. [DOI] [PubMed] [Google Scholar]
- 11.Steinberg A, et al. , Frequency of Withdrawal of Life-Sustaining Therapy for Perceived Poor Neurologic Prognosis. Crit Care Explor, 2021. 3(7): p. e0487. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Zhou F, et al. , Gray-White Matter Ratio at the Level of the Basal Ganglia as a Predictor of Neurologic Outcomes in Cardiac Arrest Survivors: A Literature Review. Front Med (Lausanne), 2022. 9: p. 847089. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Caraganis A, et al. , Interobserver Variability in the Recognition of Hypoxic-Ischemic Brain Injury on Computed Tomography Soon After Out-of-Hospital Cardiac Arrest. Neurocrit Care, 2020. 33(2): p. 414–421. [DOI] [PubMed] [Google Scholar]
- 14.Hanning U, et al. , Automated assessment of early hypoxic brain edema in non-enhanced CT predicts outcome in patients after cardiac arrest. Resuscitation, 2016. 104: p. 91–4. [DOI] [PubMed] [Google Scholar]
- 15.Kenda M, et al. , Automated Assessment of Brain CT After Cardiac Arrest-An Observational Derivation/Validation Cohort Study. Crit Care Med, 2021. 49(12): p. e1212–e1222. [DOI] [PubMed] [Google Scholar]
- 16.Beekman R and Hirsch KG, Brain imaging after cardiac arrest. Curr Opin Crit Care, 2023. 29(3): p. 192–198. [DOI] [PubMed] [Google Scholar]
- 17.Fawzi A, Achuthan A, and Belaton B, Brain Image Segmentation in Recent Years: A Narrative Review. Brain Sciences, 2021. 11(8): p. 1055. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Wijdicks EF, et al. , Validation of a new coma scale: The FOUR score. Ann Neurol, 2005. 58(4): p. 585–93. [DOI] [PubMed] [Google Scholar]
- 19.Coppler PJ, et al. , Validation of the Pittsburgh Cardiac Arrest Category illness severity score. Resuscitation, 2015. 89: p. 86–92. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Chen N, et al. , Arrest etiology among patients resuscitated from cardiac arrest. Resuscitation, 2018. 130: p. 33–40. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Rittenberger JC, et al. , Outcomes of a hospital-wide plan to improve care of comatose survivors of cardiac arrest. Resuscitation, 2008. 79(2): p. 198–204. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Beekman R, et al. , Early head CT in post-cardiac arrest patients: A helpful tool or contributor to self-fulfilling prophecy? Resuscitation, 2021. 165: p. 68–76. [DOI] [PubMed] [Google Scholar]
- 23.Wijdicks EFM, et al. , Evidence-based guideline update: Determining brain death in adults: Report of the Quality Standards Subcommittee of the American Academy of Neurology. Neurology, 2010. 74(23): p. 1911–1918. [DOI] [PubMed] [Google Scholar]
- 24.Team, R.D.C., a language and environment for statistical computing. http://www.R-project.org, 2009.
- 25.Kawai Y, et al. , Explainable artificial intelligence-based prediction of poor neurological outcome from head computed tomography in the immediate post-resuscitation phase. Scientific Reports, 2023. 13(1). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Vellido A, The importance of interpretability and visualization in machine learning for applications in medicine and health care. Neural Computing and Applications, 2020. 32(24): p. 18069–18083. [Google Scholar]
- 27.Geocadin RG, et al. , Standards for Studies of Neurological Prognostication in Comatose Survivors of Cardiac Arrest: A Scientific Statement From the American Heart Association. Circulation, 2019. 140(9): p. e517–e542. [DOI] [PubMed] [Google Scholar]
- 28.De-Arteaga M and Elmer J, Self-fulfilling prophecies and machine learning in resuscitation science. Resuscitation, 2023. 183: p. 109622. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Coppler PJ, et al. , Early risk stratification for progression to death by neurological criteria following out-of-hospital cardiac arrest. Resuscitation, 2022. 179: p. 248–255. [DOI] [PubMed] [Google Scholar]
- 30.Elmer J, et al. , Time to Awakening and Self-Fulfilling Prophecies After Cardiac Arrest. Crit Care Med, 2023. 51(4): p. 503–512. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Muschelli J, Recommendations for Processing Head CT Data. Front Neuroinform, 2019. 13: p. 61. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.