Deep learning of early brain imaging to predict post-arrest electroencephalography

Abstract

Introduction:

Guidelines recommend use of computerized tomography (CT) and electroencephalography (EEG) in post-arrest prognostication. Strong associations between CT and EEG might obviate the need to acquire both modalities. We quantified these associations via deep learning.

Methods:

We performed a single-center, retrospective study including comatose patients hospitalized after cardiac arrest. We extracted brain CT DICOMs, resized and registered each to a standard anatomical atlas, performed skull stripping and windowed images to optimize contrast of the gray-white junction. We classified initial EEG as generalized suppression, other highly pathological findings or benign activity. We extracted clinical information available on presentation from our prospective registry. We trained three machine learning (ML) models to predict EEG from clinical covariates. We used three state-of-the-art approaches to build multi-headed deep learning models using similar model architectures. Finally, we combined the best performing clinical and imaging models. We evaluated discrimination in test sets.

Results:

We included 500 patients, of whom 218 (44%) had benign EEG findings, 135 (27%) showed generalized suppression and 147 (29%) had other highly pathological findings that were most commonly (93%) burst suppression with identical bursts. Clinical ML models had moderate discrimination (test set AUCs 0.73 – 0.80). Image-based deep learning performed worse (test set AUCs 0.51 – 0.69), particularly discriminating benign from highly pathological findings. Adding image-based deep learning to clinical models improved prediction of generalized suppression due to accurate detection of severe cerebral edema.

Discussion:

CT and EEG provide complementary information about post-arrest brain injury. Our results do not support selective acquisition of only one of these modalities, except in the most severely injured patients.

Introduction

Consensus guidelines recommend providers use multiple diagnostic modalities to estimate neurological prognosis in comatose survivors of cardiac arrest.^1,2 Common tests include brain imaging with computerized tomography (CT) and electroencephalography (EEG).³ While strongly prognostic, EEG acquisition and interpretation requires equipment and expertise that are not universally available. Structural and functional measures of hypoxic-ischemic brain injury are correlated after cardiac arrest, such that common patterns of injury across diagnostic modalities co-occur.⁴ On CT, early hypoxic-ischemic brain injury is typically described by qualitative or quantitative measures of cerebral edema and resulting mass effect.⁵ On EEG, post-arrest brain injury often manifests as generalized suppression or patterns on the ictal-interictal continuum including burst-suppression with identical bursts.^6,7 There is strong collinearity between simple measures of brain edema and generalized EEG suppression, but structural damage that results in highly pathological patterns may be inapparent on clinical interpretation of early imaging.^4,7,8 Thus, standard interpretation of brain CT does not eliminate the role of EEG in multimodality post-arrest prognostication.

Deep learning technologies have achieved remarkable success in medical image-based applications.^9,10 In particular, deep learning may identify complex but informative patterns within data that are not easily captured using traditional approaches to image feature extraction. We investigated associations between brain CT images acquired early after resuscitation from cardiac arrest and initial EEG patterns using deep learning analysis. We tested the hypothesis that deep learning of brain CT could reliably predict initial EEG.

Methods

We performed a single-center, retrospective cohort study including comatose adult patients hospitalized after resuscitation from in- and out-of-hospital cardiac arrest from August 2011 to December 2018. We excluded patients who were awake, those who did not undergo both brain CT and EEG within 24 hours of arrest (for example, due to delayed presentation after interfacility transfer or rearrest before prognostic testing), those who arrested secondary to a neurological cause or trauma,¹¹ and those for whom CT imaging was contrast enhanced or motion degraded. The University of Pittsburgh Human Research Protection Office approved this study with a waiver of informed consent.

CT image acquisition and preprocessing

We routinely acquire brain CT imaging as soon as feasible in comatose post-arrest patients, typically within 2–6 hours of collapse,⁴ using a GE Lightspeed VCT 64-channel scanner (120 kVp and 225mA) which were reconstructed in the axial plane with a slice thickness of 5mm using standard algorithms. We extracted Digital Imaging and Communication in Medicine (DICOM) files from the electronic medical record. We converted these DICOMs to NIfTI format for analysis, then registered each to a standard anatomical atlas using the SPM12 (https://www.fil.ion.ucl.ac.uk/spm). Because variations in bony skull anatomy are not expected to be informative, we then performed skull stripping to remove these structures. We then resized the image volumes to a consistent size of 256×256×15. Finally, we windowed the images with a width and level of 80 and 40 HU, respectively, to optimize contrast of brain tissue and the gray/white junction. We performed image preprocessing using MATLAB software, version-R2019a (The MathWorks, Natick, MA).

EEG acquisition and interpretation

We routinely obtain continuous EEG monitoring in comatose post-arrest patients. EEG technologists place 22 gold-cup electrodes in standard International 10–20 system placement shortly after patient arrival in the intensive care unit, generally within 12 hours of collapse.⁴ We record digital video/EEG using XLTech Natus Neuroworks systems (Natus Medical, Inc). EEG findings can evolve over time after cardiac arrest, but highly pathological findings reflect more severe brain injury and are less dynamic.^8,12–14 In the present study, we only considered the initial patterns observed. Two study coauthors annotated all EEG recordings, as previously described.¹⁴ For the present analysis, we focused on the prediction of generalized suppression or other highly pathological findings that are prognostic after cardiac arrest (Table 1).⁶ Because we evaluated early EEG, we grouped burst-suppression, discontinuous, near continuous and continuous EEGs without highly pathological findings together, since in the absence of burst-suppression with identical bursts EEG background generally becomes continuous in the first 24 hours post-arrest.¹²

Table 1:

Classification of EEG outcomes

Outcome category	Included findings
Generalized suppression	Isoelectric suppression
	Burst suppression with identical bursts
Other highly pathological findings	Seizures
	Generalized or other periodic discharges
	Continuous or near continuous background activity
Benign activity	Discontinuous or burst suppression without identical bursts, seizures, or periodic discharges
	No discharges or sporadic epileptiform discharges only

Clinical covariates

We extracted standard demographic and clinical information available on patient presentation from our prospective registry. Variables included age; sex; arrest location (in-hospital vs. out-of-hospital); presenting rhythm (ventricular tachycardia/fibrillation (VT/VF), pulseless electrical activity (PEA), asystole); arrived from the scene of collapse or via interfacility transfer; witnessed status; immediate CPR (none, layperson or professional); arrest duration; number of rescue shocks administered; number of epinephrine doses administered; any confirmed cardiac etiology of arrest, classified as previously described;¹¹ and Charlson Comorbidity Index. We also extracted presenting neurological examination: pupillary light reflex (bilaterally present, unilaterally present, bilaterally absent, unable to assess); cough reflex and gag reflex (present, absent, unable to assess); motor exam (localizing to pain; flexion/withdrawal; extension; myoclonus; none; unable to assess). Finally, we extracted Pittsburgh Cardiac Arrest Category, a validated measure of global presenting post-arrest illness severity.¹⁵

EEG outcomes

Based on their clinical relevance and hypothesized structure-function correlates, we focused on three primary EEG outcomes of interest: generalized suppression vs. other highly pathological activity vs. benign activity. We operationalized our outcome in this manner instead of using other validated approaches for classification of post-arrest EEG (e.g. that developed by Westhall, et al.) because we hypothesized patients with generalized suppression would likely have distinct radiographic findings compared to those exhibiting other “highly malignant” findings.⁶ We explored several ways to structure the prediction task. First, we treated EEG as a multinomial (i.e., 3-level categorical) outcome. Thereafter, we considered each possible pairwise combination to discriminate EEG states: generalized suppression vs. other highly pathological activity; generalized suppression vs. benign activity; and other highly pathological activity vs. benign activity.

Statistical analysis, deep learning network architecture and model training

We summarized cohort clinical characteristics and outcomes using descriptive statistics and report mean with standard deviation or median with interquartile range, as appropriate. We inspected clinical variables for frequency and patterns of missingness, then used multiple imputation with chained equations to create complete data sets using predictive mean matching for continuous variables and multinomial logistic regression for categorical variables.

First, we used three machine learning methods: logistic regression, support vector machine (SVM), and a multi-layer perceptron neural network with two hidden layers (30 nodes in the first hidden layer and 10 nodes in the second hidden layer) to predict EEG outcomes from clinical covariates (not including brain imaging). We used 5-fold cross validation to evaluate the performance of the model. We calculated and reported mean area under the curve (AUC) across the five splits and its standard deviation. For the multinomial classification task, we calculated AUC using a one-vs-rest approach.

Next, we developed our imaging-based models. We used the 9 center slices (slice numbers 4–12 out of 15) as clinically relevant input to the model (Figure 1). We used three state-of-the-art approaches to build multi-headed deep learning models using similar model architectures. We combined multiple convolutional layer components in parallel and built three multi-headed deep learning models using VGG16,¹⁶ ResNet50,¹⁷ and GoogLeNet¹⁸, respectively, as the backbones. Each multi-headed model has an input of 9 slices preprocessed patient images (256 × 256 × 9). We concatenated the outputs of the multiple convolutional channels and passed them to fully connected layers. We pretrained each model on ImageNet,¹⁹ a large natural image dataset, using Adam algorithm for optimization.²⁰ We implemented these deep learning networks using Keras on a computer system with the following specifications: Intel® Core^™ i7-2670QM CPU@2.20GHZ with 8 GB RAM and a Titan X Pascal Graphics Processing Unit (GPU). We again used five-fold cross validation to train and test model performance, respectively and report average AUC and its standard deviation. Finally, we performed each prediction task combining the best performing clinical and imaging models (Figure 2). We will publicly share our code via GitHub after acceptance of the paper.

Figure 1:

Each co-registered brain CT consisted of 15 slices, for which we selected the nine center slices (slice numbers 4–12) as clinically relevant model inputs.

Figure 2:

Example of multi-headed deep learning network architecture. VGG16 backbone is shown here as an example. In our imaging-only models, no clinical data were included. In the combined models, clinical data entered directly into a dense layer that was then combined with imaging to make a single output prediction (red border).

Results

Overall, 500 patients met inclusion and exclusion criteria (Figure 3). Mean (SD) age was 58 (17) years, 205 (41%) were female and 448 (90%) arrested out-of-hospital (Table 2). Benign EEG findings were observed in 218 (44%) patients, generalized suppression was observed in 135 (27%) and 147 (29%) had other highly pathological findings that were most commonly (93%) burst suppression with identical bursts (Supplemental Table 1). Overall, 2% of clinical covariates were missing and we imputed five complete data sets.

Figure 3:

STROBE diagram

Table 2:

Cohort clinical characteristics and outcomes.

Clinical characteristic	Overall cohort (n = 500)
Age, years	58 (17)
Female sex	205 (41)
Out-of-hospital arrest	448 (90)
Interfacility transfer	338 (68)
Initial rhythm
VT/VF	147 (29)
PEA	172 (34)
Asystole	151 (30)
Unknown	(30)
Witnessed collapse	284 (57)
CPR prior to ALS
Layperson	125 (25)
Professional	171 (34)
Shocks	0 [0 – 2]
CPR duration	20 [11 – 30]
Epinephrine doses	3 [2 – 5]
Cardiac etiology of arrest	133 (27)
Charlson Comorbidity Index	l [0 – 2]
Presenting exam findings
Pittsburgh Cardiac Arrest Category
II	124 (25)
III	48 (10)
IV	287 (57)
Pupillary light reflex
Bilaterally present	259 (52)
Bilaterally absent	218 (44)
Gag reflex present	190 (38)
Cough reflex present	173 (35)
Motor response
Localizing	26 (5)
Flexion/withdrawal	81 (16)
Extension	11 (2)
No response	247 (49)
Myoclonus	109 (22)

Data are presented as raw number with corresponding percentage, mean (standard deviation) or median [interquartile range].

Overall, logistic regression and neural network models consistently outperformed SVM predicting EEG from clinical characteristics and had moderate discrimination (average test set AUCs 0.73 – 0.80) (Table 3). Image-based deep learning models performed worse than models from clinical predictors (test set AUCs 0.51 – 0.69), particularly discriminating benign from highly pathological findings. Combining clinical and imaging models never outperformed logistic regression or neural network-based clinical models alone in terms of overall discrimination. However, incorporation of image-based deep learning models did improve sensitivity at 0% false positive rate for prediction of generalized EEG suppression compared to clinical models alone due to accurate detection of severe diffuse cerebral edema (Supplemental Figures). For example, sensitivity improved from 0.09 to 0.26 in models predicting generalized suppression vs benign EEG.

Table 3:

Performance of clinical, imaging and combined models predicting initial EEG state among comatose survivors of cardiac arrest.

	Predicted outcome(s)
Model/Inputs	Benign vs. suppressed	Benign vs. highly pathological	Suppressed vs. highly pathological	Benign vs. suppressed vs. highly pathological
Clinical covariates
LR	0.73 (0.02)	0.80 (0.04)	0.74 (0.08)	0.74 (0.04)
SVM	0.69 (0.04)	0.70 (0.05)	0.63 (0.03)	0.66 (0.03)
NN	0.73 (0.02)	0.80 (0.05)	0.75 (0.04)	0.75 (0.04)
Imaging
VGG16	0.69 (0.04)	0.51 (0.03)	0.66 (0.05)	0.60 (0.04)
ResNet50	0.64 (0.04)	0.55 (0.06)	0.62 (0.02)	0.61 (0.02)
GoogLeNet	0.66 (0.04)	0.51 (0.07)	0.62 (0.04)	0.58 (0.04)
Combined model	NN + VGG16	NN + ResNet50	NN + VGG16	NN + ResNet50
Performance	0.73 (0.05)	0.61 (0.05)	0.67 (0.06)	0.66 (0.03)

Data are presented as area under the curve with standard deviations.

Discussion

Guidelines recommend providers use multiple diagnostic modalities when predicting outcome in comatose survivors of cardiac arrest.^1,2 This approach is only sensible insofar as data from each modality are not perfectly correlated. We hypothesized that deep neural learning applied to brain CT images acquired early after cardiac arrest could be used to predict subsequent EEG patterns. Strong discrimination in this predictive task might obviate the need to acquire both CT and EEG, which is appealing insofar as CT is more widely available than urgent EEG. Overall, our results do not support the feasibility of this approach. Our deep neural learning models demonstrated only modest discrimination predicting EEG, suggesting these modalities assess largely independent aspects of early hypoxic-ischemic brain injury. A notable exception was the specific case in which generalized suppression could be accurately predicted from models that identified severe cerebral edema. This association is biologically plausible and has been demonstrated using simpler metrics of edema such as grey matter to white matter attenuation ratio or qualitative assessment.⁴ Taken together, this suggests that in patients with severe cerebral edema acquisition of EEG does not improve prognostication, and that complex deep neural learning approaches do not capture additional prognostic information in this application. By contrast, clinical models had moderate discrimination predicting EEG. This is not surprising given many established relationships between these factors. For example, many (but not all) patients with post-anoxic myoclonus demonstrate burst suppression with identical bursts,²¹ and longer arrest duration predicts more severe brain injury and thus risk of generalized suppression.

Our deep neural learning could not discriminate between benign and highly pathological EEG findings on initial EEG. Virtually all highly pathological findings in this cohort were generalized suppression with periodic discharges or burst suppression with identical bursts. Generalized periodic discharges may develop in patients with only mild hypoxic ischemic brain injury.²² Patients with burst suppression with identical bursts have been described on postmortem histopathology to demonstrate widespread and severe neuronal damage.^23,24 These studies have relied on quantification of selective eosinophilic neuronal death, a process that is delayed hours to days after ischemia.²⁵ Early severe cerebral edema is mechanistically distinct and develops rapidly after devastating global brain ischemia.^26,27 It may be that analysis of delayed brain imaging would identify anatomic findings that correlate with highly pathological findings, a hypothesis we could not explore with this data set.

Our study has several limitations. First, included patients were treated at a single academic medical center, limiting the generalizability of our findings. Moreover, we performed preprocessing that excluded contrasted or motion-degraded imaging, a further threat to generalizability. In addition, while the deep learning models constructed using three representative backbones (VGG16, ResNet50, and GoogLeNet) showed similar effects in the imaging analysis, it is not impossible that more advanced modeling techniques may show different performance, which merits further evaluation and investigation in future work. Both structural correlates of post-arrest brain injury and EEG findings evolve in the hours to days after resuscitation from cardiac arrest.^12,26–28 Thus, it may be that associations between these two modalities also vary over time, such that, for example, findings from CT and EEG acquired 24 hours post-arrest are better correlated than earlier diagnostic results.

In conclusion, except in cases of severe early cerebral edema after cardiac arrest, brain CT imaging and EEG are complementary diagnostic modalities. Advanced analytical approaches cannot extract anatomic information from early imaging that reliably predicts prognostic neurophysiological findings identified on EEG.