Quantitative Serial CT Imaging-Derived Features Improve Prediction of Malignant Cerebral Edema after Ischemic Stroke

Abstract

Introduction:

Malignant cerebral edema develops in a small subset of patients with hemispheric strokes, precipitating deterioration and death if decompressive hemicraniectomy (DHC) is not performed in a timely manner. Predicting which stroke patients will develop malignant edema is imprecise based on clinical data alone. Head computed tomography (CT) imaging is often performed at baseline and 24-hours. We determined the incremental value of incorporating imaging-derived features from serial CTs to enhance prediction of malignant edema.

Methods:

We identified hemispheric stroke patients at three sites with NIHSS ≥ 7 who had baseline as well as 24-hour clinical and CT-imaging data. We extracted quantitative imaging features from baseline and follow-up CTs, including CSF volume, intracranial reserve (CSF/cranial volume), as well as midline shift (MLS) and infarct-related hypodensity volume. Potentially lethal malignant edema was defined as requiring DHC or dying with MLS over 5-mm. We built machine-learning models using logistic regression first with baseline data and then adding 24-hour data including reduction in CSF volume (ΔCSF). Model performance was evaluated with cross-validation using metrics of recall (sensitivity), precision (predictive value), as well as area under receiver-operating-characteristic and precision-recall curves (AUROC, AUPRC).

Results:

Twenty of 361 patients (6%) died or underwent DHC. Baseline clinical variables alone had recall of 60% with low precision (7%), AUROC 0.59, AUPRC 0.15. Adding baseline intracranial reserve improved recall to 80% and AUROC to 0.82 but precision remained only 16% (AUPRC 0.28). Incorporating ΔCSF improved AUPRC to 0.53 (AUROC 0.91) while all imaging features further improved prediction (recall 90%, precision 38%, AUROC 0.96, AUPRC 0.66).

Conclusion:

Incorporating quantitative CT-based imaging features from baseline and 24-hour CT enhances identification of patients with malignant edema needing DHC. Further refinements and external validation of such imaging-based machine learning models are required.

INTRODUCTION

Cerebral edema is one of the most important complications of acute ischemic stroke. The majority of those suffering a stroke exhibit increases in brain volume due to water accumulation in and around the lesion [1,2]. However, it is only in a minority that cerebral edema results in significant midline shift (MLS) and neurological deterioration [3]. This malignant edema has an extremely high mortality unless decompressive hemicraniectomy (DHC) surgery is performed before cerebral herniation [4–6]. In fact, malignant edema is the leading cause of death and deterioration in the first week after stroke [7]. Nonetheless, less than 5% of all strokes and only 20–30% of those with large hemispheric infarction will develop malignant edema [8].

Identifying which stroke patients are at high-risk for deterioration and triaging these to surgery or other early aggressive interventions has been impeded by lack of accurate early predictors [9]. Age, stroke severity (measured by National Institutes of Health Stroke Scale, NIHSS) and other baseline variables alone cannot provide sufficient discriminative information to select patients. Lesion volume measured on early MRI can predict deterioration with reasonable accuracy but is not widely available or routinely performed [10]. Furthermore, the lesion volume does not distinguish the stroke itself from the swelling that causes herniation. There is increasing interest in CT-based imaging measures for quantification and prediction of edema [11,12]. For example, measuring baseline intracranial reserve (i.e. CSF volume as a proportion of cranial volume) may be more informative than patient age in predicting which patients are more likely to develop MLS and deterioration [13,14]. We have recently demonstrated that serial measurements of CSF volume (i.e. applying reduction in CSF volume as a surrogate of the brain volume increase from edema) can provide quantitative data on edema severity [15]. Additionally, we demonstrated that incorporating follow-up imaging data within 24-hours significantly improves prediction of potentially lethal cerebral edema (death or need for DHC) [16]. However, for that risk prediction tool, we only employed crude manual review of imaging. It is likely that there is valuable quantifiable imaging features that can further improve prediction of edema. Here we aim to develop a multivariable machine-learning model for predicting malignant cerebral edema using a combination of clinical and quantitative imaging variables extracted from CTs at baseline and 24-hours. Our hypothesis was that integrating baseline and follow-up imaging features would significantly improve detection of those developing malignant edema.

METHODS

Study participants and clinical data

Patients diagnosed with ischemic stroke at three international stroke centers were enrolled in a prospective observational cohort (the Genetics of Neurological Instability after Ischemic Stroke, GENISIS study) between 2008 and 2017. All participants provided informed consent for data collection, including acute stroke imaging. Participants were excluded if stroke onset time was unknown, if the final diagnosis was not stroke or if stroke was located in the brainstem or cerebellum. All those whose imaging was available for centralized analysis were included in this sub-study. Participants were excluded if baseline CT was not performed within 12 hours of stroke onset or already showed a clear stroke-related hypodensity, or if follow-up CT was either not available or not performed within 48 hours of stroke. NIHSS was obtained at both baseline and at 24-hours. To analyze those at highest risk for edema, we limited our analysis to those with NIHSS of seven or greater at baseline [17]. Glucose levels were measured on admission to the emergency department. All participants were followed prospectively for neurologic deterioration due to cerebral edema. Our primary endpoint was the development of malignant cerebral edema leading to either DHC and/or resulting in death in the presence of MLS of 5-mm or greater [16]. Such potentially lethal edema is an extreme endpoint but we felt that identification of these most severe cases requiring surgery or otherwise dying represented the most relevant target group where early decision-making is critical. We estimated that 5-10% of stroke patients would develop malignant edema. We aimed to collect data on 500 stroke patients in order to collect 25 edema-related events. As this is a convenience sample from a larger stroke study, we were limited to those enrolled who met all our eligibility criteria.

Imaging analysis

Both baseline and 24-hour CT imaging were processed to extract quantitative imaging variables, as described previously, using algorithms developed in-house using TensorFlow (version 1.13.1) and extracted using MATLAB (2019a, Natick, Massachusetts: The MathWorks Inc.) [18], This included first extraction of the intracranial supratentorial space using k-means clustering for skull removal followed by registration of each baseline image to atlas templates for removal of infratentorial and non-cranial structures. The follow-up images were then coregistered to the baseline for each participant to ensure similar brain volumes were analyzed. These cranial regions then underwent automated segmentation into CSF and brain compartments [19], If more than one scan was performed at a time point (e.g. normal axial brain and thin slices), we selected scans with 3 to 5-mm slice thickness, for consistency. The intracranial reserve was calculated as the proportion of cranial volume comprised by CSF [14], Percent change in CSF volume from baseline to 24-hours was calculated as ΔCSF. Midline shift was measured manually by a single trained investigator at the level of the septum pellucidum. Infarct-related hypodensity was manually outlined slice-by-slice to provide lesion volume (a combination of infarct and edema).

Machine learning model building

The dataset was complete except for a few missing data points: for NIHSS (3) and glucose (4). Values were imputed for each missing feature using the three closest data points, using the K-Nearest Neighbor (KNN) algorithm. All features were then standardized to improve stability of model training. The dataset was partitioned into ten folds and 10-fold cross validation was employed to train and internally validate each model. In each often rounds, we oversampled the cases of malignant edema to balance the cases and controls within the 9-folds used as the model training dataset. Training size was further increased using data augmentation (i.e. by adding three percent noise to the smallest principal components).

We applied logistic regression to provide probabilistic binary outputs using a one layer neural network with softmax activation, implemented within the Keras machine-learning platform. Several regression models were trained, sequentially increasing the features that were included. The first model incorporated only baseline clinical variables that were associated with our endpoint in univariate analysis. The second added baseline intracranial reserve to these variables. The third added ΔCSF extracted automatically from 24-hour imaging. The final regression model incorporated all baseline and 24-hour data, including all quantitative imaging measures.

Model performance:

After training each model on the 9-fold data, its performance was tested on the 10^th (validation) fold to provide unbiased results for that round and minimize over-fitting. This validation fold comprised the original, unaltered data without augmentation and had not been seen by the algorithm during training. This process was repeated ten times such that each fold was used once for validation. This allowed the construction of an overall matrix showing true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) for prediction of all cases and controls in our dataset. As our outcome of interest is relatively rare (i.e. many ‘negatives’ who do not develop malignant edema), accuracy and specificity would be poor metrics that overestimate performance for prediction of edema cases. Even area under the receiver-operating-curve (AUROC, which plots sensitivity vs 1-specificity), although commonly presented to summarize model performance, does not accurately reflect prediction when the outcome of interest is uncommon. Preferred metrics in such situations are recall: the sensitivity to detect cases (i.e. TP/TP+FN) and precision: the positive predictive value (TP/TP+FP), indicating what proportion of predicted positives actually turn out to have malignant edema (versus how many are FP). Overall model performance requires balancing precision and recall and is best captured by the area under the precision-recall curve [20].

RESULTS

Of 1799 people enrolled in the cohort study, 759 had no imaging available. Of those with imaging, 336 had either no follow-up CT or CT was beyond 48 hours after stroke. An additional 37 had a non-stroke diagnosis (7), a cerebellar or brainstem stroke (24), or stroke onset time was unknown (6). Baseline CT was not available for 24 while 18 of the remaining 616 were excluded due to image processing problems. Those excluded were predominantly milder strokes (median NIHSS of 5) without malignant edema. This left 598 subjects with baseline and follow-up CT close to 24-hours. None of those with NIHSS of six or less developed malignant edema. Of the 361 with NIHSS of seven or above, 20 (6%) developed malignant cerebral edema requiring DHC (12) or resulting in death (8). Median time to DHC was 2.7 days (IQR 1.7-5) and to death was 5 days (IQR 4-6). All imaging was performed prior to DHC.

Time from stroke onset to baseline CT was a median of 1 hour and 33 minutes (IQR 1-3 hours) and time to follow-up CT was 25 hours (IQR 19-28). Those with malignant edema had higher baseline NIHSS and glucose and were non-significantly younger and more likely to be nonwhite (TABLE 1). They also had significantly lower intracranial reserve (11.5% vs. 14%, p=0.002). Nonetheless, review of the distribution of data demonstrated significant overlap between those with and without edema for all baseline variables (Supplemental Figure 1). CSF reserve was correlated with age (r=0.72) but even within a given age range, the reserve was lower in those progressing to develop malignant edema (see Supplemental Figure 2).

Table 1:

Comparison of baseline and follow-up data in cases with malignant edema and controls

Variable	Malignant Edema (n=20)	No Malignant Edema (n=341)	P value
Age, years	65.3 (15)	69.4 (13)	0.26
Sex, female	13 (65%)	153 (46%)	0.13
Race, non-white	5 (25%)	39 (11%)	0.08
NIHSS, baseline	19.5 (15.5-21)	13 (10-18)	0.0002
Serum glucose, mg/dl	151 (121-232)	122 (108-147)	0.004
Received tPA	15 (75%)	272 (80%)	0.78
Time to baseline CT	1 hr 24 mins	1 hr 33 mins	0.66
Intracranial reserve	11.5% (7-13)	14% (10-18)	0.002
Time to FU CT	20 hours (15-29)	25 hours (19-28)	0.38
NIHSS at 24-hours	20.5 (19-26)	7 (3-14)	< 0.0001
Change in NIHSS	+3 (+7 to −1)	−5 (−1 to −10)	< 0.0001
Midline shift at 24-hours, mm	4.7 (3-9.7)	0 (0-0)	< 0.0001
Reduction in CSF volume	53% (37-70)	15% (6-27)	< 0.0001
Infarct volume, ml	252 (181-322)	4 (0-37)	< 0.0001

All continuous data are presented as medians (interquartile range) except for age which is presented as mean (standard deviation).

All quantitative measures obtained at 24-hours were strongly associated with malignant edema. Not only was 24-hour NIHSS higher, but those progressing to malignant edema exhibited a worsening of their score (by a median of 1 point) over time, compared to a net improvement of five points in the NIHSS amongst controls. Midline shift was greater and all except three in the edema group had measurable MLS by 24-hours while most of the 341 controls did not. However, in absolute numbers, there were just as many with MLS amongst those not ultimately requiring surgery or dying. Similarly, infarct volume was much greater on average in the malignant edema group, but significant overlap existed and there were three cases without measurable (i.e. visible) hypodensity at 24-hours who still went on to develop malignant edema. All these cases still had significant measurable reduction in CSF volume (median 74% from baseline) despite no measurable infarct volume.

The model with baseline variables alone had poor ability to predict malignant edema (AUROC 0.59). There were several false negatives (i.e. recall of only 60%), many false positives (precision of 7%, i.e. more than ten for every one case identified), resulting in AUPRC of only 0.15. When intracranial reserve was added, recall improved to 80%, AUROC to 0.82 but precision remained low at 16% and AUPRC 0.28 (FIGURE 1). 46% of errors from the baseline model were correctly reclassified with knowledge of intracranial reserve (primarily false positives that were now identified as not at risk for malignant edema, i.e. true negatives).

Figure 1:

Comparison of performance between four models to predict malignant cerebral edema

Each of the 24-hour imaging biomarkers was first evaluated in isolation as a predictor of malignant edema. The AUROC for infarct volume was 0.86 with a sensitivity of 85% with 85% specificity at a threshold volume of 87-ml. Similarly, MLS alone had an AUROC of 0.89 with a sensitivity of 85% and specificity of 90% using a cutoff of 2-mm. ΔCSF had the highest AUROC (0.91) and could achieve 95% sensitivity with specificity of 74% at threshold above 27%. This compares to the AUROC for change in NIHSS of 0.81 (0.72-0.90) with a sensitivity of 70% and specificity of 76% if NIHSS worsened at 24-hours relative to baseline. When ΔCSF was added to baseline variables in the multivariate model, performance improved, reducing the number of false positives (i.e. higher precision with AUPRC almost doubling to 0.53). When all 24-hour variables were included, the final model was able to correctly identify 90% of cases (AUROC 0.96) with many less false positives (precision up to 0.38) and AUPRC of 0.66 (FIGURE 2). There was an 80% net reclassification of errors from the baseline model.

Figure 2:

Receiver operating curve and precision recall curve for four models to predict malignant cerebral edema

DISCUSSION:

In our cohort of over 350 stroke patients from three international sites, only 20 (6%) experienced the most severe form of malignant edema, i.e. requiring DHC or resulting in death. Nonetheless, this proportion is in line with other large stroke cohort studies [17]. By limiting our study to those with baseline NIHSS of seven or greater, in order to focus on larger strokes that were at greater risk for edema [21], and by requiring subjects have undergone repeat imaging within 48 hours, we eliminated milder strokes. We confirmed that those with NIHSS below seven are at very low risk; in fact, none of those amongst almost 250 patients in our cohort with mild stroke severity on admission developed malignant edema. Our remaining cohort still represents a broad sample of moderately severe strokes of varying sizes, approximating the primary population in which a concern for edema may be raised when a patient with stroke is admitted.

Our results confirm that accurately predicting which of these patients will develop midline shift, deteriorate, and require surgery is difficult at the time of admission. Although these patients tend to have higher NIHSS on admission, no specific cutoff can be applied to confidently identify those at highest risk. We confirmed that hyperglycemia is associated with more severe edema [22]. However, a model with all available baseline clinical variables had poor sensitivity to identify malignant edema and would result in ten false positives for every true case identified. We demonstrated that baseline intracranial reserve (i.e. CSF volume as a proportion of cranial space) improves prediction of those at risk [14]. While atrophy is correlated with age, this imaging surrogate of atrophy was significantly better at distinguishing those destined for malignant edema than age alone. This biomarker, extracted automatically by our image analysis pipeline, has also been demonstrated to predict outcome after endovascular treatment for large vessel stroke [23], It likely captures the space available to compensate for swelling and those with more reserve are able to tolerate larger strokes with more edema before deterioration. It also highlights how quantitative imaging biomarkers may aid in stroke prognosis beyond clinical or traditional qualitative imaging evaluation, which may not as completely or precisely capture the extent of reserve available and early edema that is developing.

Our prior study suggested that measuring reduction in CSF volume from baseline to follow-up CT at 24-hours may provide a useful biomarker of edema severity [15], We chose to evaluate CSF volume rather than brain volume primarily because it is easier to accurately segment on CT images of variable quality and represents a proxy that reciprocates brain volume increases. In this study, incorporating ΔCSF improved precision (i.e. less false positives) and significantly increased AUPRC. Precision and recall are the most relevant metrics when the goal is predicting rare but life-threatening events such as malignant edema, where accuracy and specificity would overestimate performance. Furthermore, a measurable reduction in CSF was observed in several cases where no MLS and/or no visible hypodensity were seen on follow-up imaging in those later going on to develop malignant edema. When we evaluated infarct volume as a predictor, we found that a threshold of 90-ml provided maximal sensitivity with reasonable specificity. This cutoff is very similar to that proposed using acute MRI [10], However, infarct volume measured with MRI in that prior study and with CT in ours had low predictive value.

We found that combining several imaging biomarkers on 24-hour HCT (i.e. lesion volume along with MLS and ΔCSF) in a machine-learning model significantly improved prediction beyond any single measure alone. This finding is intuitive and perhaps not surprising, as detection of a large infarct with early MLS is an important component in selecting patients for DHC. A regression model with all baseline and 24-hour variables combined had 90% recall, suggesting that one in ten cases would still be missed. This is in accordance with prior studies incorporating MRI lesion volume that demonstrated reasonable prediction but were unable to ensure 100% sensitivity [10], In addition, we believe that a CT-based approach has broader applicability than performing MRI acutely in all stroke patients. Furthermore, we have automated the quantitative analysis of CSF volumes and are currently testing automated measurement of infarct volume and midline shift into a seamless end-to-end image processing pipeline that could provide all this data on stroke patients to guide the clinician. Incorporating such data into machine-learning models could provide a probabilistic prediction of whether a patient is likely to develop worsening edema and require DHC.

Strengths of our study include our use of a rigorous internal validation to minimize the likelihood of overfitting. This is a risk when constructing and testing a model on the same small sample of data. Cross-validation avoids this bias by training the model on a different subset of data than what was used for testing its performance. We also provide proof-of-principle that quantitative imaging biomarkers can provide informative data for prediction of malignant edema beyond traditional clinical and crude imaging parameters. A combination of CSF metrics, obtained automatically from all almost one thousand CT scans in this study, quantified both baseline reserve available to compensate for edema and the degree of brain volume increase seen as a result of edema (ΔCSF). These, in concert with early midline shift, allow reasonably good prediction of malignant edema.

This quantitative modeling approach builds upon the EDEMA (Enhanced Detection of Edema in Malignant Anterior Circulation Stroke) score that our group previously developed [16], That demonstrated that ordinal ranking of simple imaging measures (effacement of basal cisterns, degree of midline shift) on 24-hour CT could enhance prediction beyond baseline variables. In this study, by quantifying displacement of CSF and volume of infarct-related hypodensity on repeat imaging, we further improved prediction. Our logistic regression models provide probabilistic outputs for risk of malignant edema; we did not construct or test a specific score as this would require categorizing continuous variables and could lose some of the power harnessed by our quantitative approach. Furthermore, we believe a score is appropriate when a predictive model is ready to be applied at the bedside and this study still only represents a preliminary proof-of-principle evaluation of quantitative imaging in the prediction of edema. Further testing in larger cohorts for external validation is required before clinical implementation can be considered. Ultimately, the implications of utilizing any such model on clinical decisionmaking and whether enhanced early prediction can improve outcomes for stroke patients has to be rigorously evaluated [24].

There remain several limitations to this study. Even in our relatively large cohort, lethal edema was relatively rare. We also could only include those patients with repeat imaging performed, representing a biased subset of those with more severe strokes and those at higher risk for deterioration. One factor mitigating selection bias is that at two of the three centers involved, all those receiving tPA (representing over three-quarters of our cohort) routinely undergo repeat CT at 24-hours. Ideally, we would like to validate our models in prospective external cohorts with routine imaging at 24-hours (or even earlier) to avoid bias and ensure generalizability. Lack of validation and limited generalizability have been major problems for stroke prognostic studies and hindered implementation of clinically applicable prediction tools [25]. We are currently collecting imaging and clinical data on an additional 2,000 stroke patients and finalizing an image-processing pipeline capable of extracting all the quantitative metrics outlined, so that we can refine and validate these prediction models. We included a broad sample of strokes in this initial study. In future studies, we may focus on large vessel strokes, as these may be at particularly high-risk for edema.

Our final model, although performing significantly better than those with only baseline data, still had incomplete precision and could not identify all cases of malignant edema. It identified two false cases for every one true case of edema. Therefore, it could at best, be used to triage high- risk candidates for early aggressive medical interventions and not select definitely which patients require surgery [26]. It is possible that additional data would further improve prediction. For example, lesional water uptake, another measure of early edema formation, could be useful in determining which patients are liable to malignant edema [2]. This can be calculated from baseline or follow-up CTs by measuring the density of the infarct-related region compared to the contralateral unaffected brain. This density ratio is a dynamic measure that has been associated with evolving midline shift and was able to resolve treatment differences from interventions to reduce edema [27]. In addition, in those with large vessel strokes, evaluation of collaterals may provide insights into risk for edema [28]. The ASPECTS score has also shown promise as a baseline imaging feature, incorporated with NIHSS and collateral profile in the malignant brain edema (MBE) score for those with large-vessel occlusions [29]. Evaluation of ASPECTS can be time-consuming and subject to inter-rater variability, but automated algorithms are emerging to facilitate this [30].

Finally, it may be that regression approaches that combine such inter-related clinical and imaging data are not optimal for prediction of complex events. Machine learning models that are able to identify more complex non-linear interactions within datasets could improve prediction, as they have shown for other stroke outcomes such as delayed ischemia after subarachnoid hemorrhage [31]. Other studies have suggested that regression models do work as well as more complex machine learning approaches to predict stroke outcome [32]. We are now exploring whether neural networks, specifically those that can integrate time-series data, can improve prediction. Such recurrent neural networks have recently demonstrated promise for predicting complications after stroke [33]. As we increase our dataset, we will test similar approaches to predicting malignant edema using serial imaging phenotypes. We hope that such a non-linear model that integrates the evolution of biomarkers over time can identify all cases of edema with fewer false positives. In addition, rather than using pre-specified imaging features, a deep learning approach that takes the raw imaging data as input could provide an even more powerful and truly agnostic approach to predicting edema.