Neurocritical Care

, Volume 17, Issue 2, pp 240–244

Clinical MRI Interpretation for Outcome Prediction in Cardiac Arrest


    • Department of NeurologyYale University School of Medicine
    • Stroke Service, Massachusetts General Hospital
  • Patricia Scripko
    • Stroke Service, Massachusetts General Hospital
  • James Bartscher
    • Department of NeurologyHenry Ford Hospital
  • John Sims
    • Stroke Service, Massachusetts General Hospital
  • Erica Camargo
    • Stroke Service, Massachusetts General Hospital
  • Aneesh Singhal
    • Stroke Service, Massachusetts General Hospital
  • Michael Parides
    • Mount Sinai University
  • Karen Furie
    • Stroke Service, Massachusetts General Hospital
Original Article

DOI: 10.1007/s12028-012-9716-y

Cite this article as:
Greer, D., Scripko, P., Bartscher, J. et al. Neurocrit Care (2012) 17: 240. doi:10.1007/s12028-012-9716-y



In clinical practice, magnetic resonance imaging (MRI) is commonly used to assess the severity of a cardiac arrest patient’s cerebral injury, utilizing treating neurologists’ imaging interpretation. We sought to determine whether clinical interpretation of diffusion-weighted imaging (DWI) helps to determine poor outcome in comatose cardiac arrest patients.


We analyzed 80 consecutive MRIs from patients in coma following cardiac arrest. Each study was graded as “normal” or “abnormal restricted diffusion” in pre-specified brain regions by two blinded stroke neurologists. Poor outcome was defined as a modified Rankin Scale (mRS) score >4 at 3 months. Formal interpretations of neuroimaging by non-blinded neuroradiologists were compared with the blinded reviews by the stroke neurologists.


DWI abnormalities were highly sensitive (98.5 %) but only modestly specific (46.2 %) for predicting poor neurological outcome. Inter-observer reliability was moderate (kappa = 0.49 ± 0.32), with 91 % agreement between study observers, and no significant differences in study observers’ interpretations (p = 0.125). There were, however, significant differences between the study observers and the clinical neuroradiologists in identifying studies showing evidence of global hypoxic-ischemic injury (p = 0.001).


The qualitative evaluation of imaging abnormalities by stroke physicians in comatose cardiac arrest patients is a highly sensitive method of predicting poor outcome, but with limited specificity.


MRIComaCardiac arrestPrognosisHypoxic-ischemic injury


Although a large proportion of patients who suffer a cardiac arrest (CA) have a poor neurological outcome, it remains difficult to identify which comatose survivors are destined for a poor outcome. Patients who awaken rapidly following CA commonly have little or no disability, and thus do not present a diagnostic dilemma. Conversely, those who remain comatose with absent pupillary or corneal reflexes after 72 h or absent bilateral cortical somatosensory evoked potentials (SSEPs) are highly likely to have a poor outcome [1]. Unfortunately, there is a large indeterminate population whose clinical exams fall between these two extremes.

The clinical neurological assessment remains a standard for establishing severity of injury and predicting outcome. The 72-h coma exam, however, has practical limitations in the intensive care setting where multi-system organ failure, metabolic disarray, sedating medications, and focal neurological deficits are common. Neuroimaging shows promise for determining prognosis based on structural brain injury, and is relatively unencumbered by the confounding factors that often reduce the utility of clinical assessment. Diffusion-weighted imaging (DWI) identifies early ischemia-related changes [2] and offers great potential for providing prognostic value after CA. In animals that experience global cerebral ischemia, changes in apparent diffusion coefficient (ADC) values are immediately seen [3]. However, brain ADC values can return to near normal with successful early resuscitation, as demonstrated in animal studies [3, 4]. Studies in humans suggest DWI changes may correlate with poor outcomes post-arrest [59]. Although the technology of DWI is promising, its use as a prognostic tool in CA is not yet validated. The impacts of temporal, spatial and patient-specific factors are not known. Clinically, DWI is often utilized to provide an understanding of the extent of hypoxic-ischemic damage.

We previously demonstrated that specific computer-generated temporal and regional changes on neuroimaging are both sensitive and specific for poor outcome in patients who suffer hypoxic-ischemic cerebral injury from a CA [7]. However, computer-assisted quantitative measures are not commonly performed in clinical practice. A more typical clinical scenario is a qualitative evaluation by consulting neurologists, especially stroke physicians. We sought to evaluate whether this approach was predictive of a poor outcome. We hypothesized that stroke neurologists’ interpretations of imaging findings after CA are highly sensitive for predicting poor outcome and would correlate with the formal non-blinded interpretations by neuroradiologists at the time of the study’s performance.


Using a database previously reported [7], we analyzed data acquired from 2000–2007 as part of an institutional review board–approved, Health Insurance Portability and Accountability Act compliant, single-center prospective observational study of 500 patients with non-traumatic coma. Within this cohort, 200 patients (40 %) were comatose after CA. A modified Rankin Scale (mRS) score was obtained by telephone interview at 3 months.

Because this was an observational study, the decision to perform neuroimaging was at the discretion of the treating clinicians. A total of 80 patients (40 % of the total post-cardiac arrest population) underwent magnetic resonance imaging (MRI). Fourteen patients were treated with therapeutic hypothermia. In patients who underwent more than one MRI, the initial MRI was used for this study.

Imaging Studies

Patients underwent MR imaging as previously described (15) using 1.5-T imaging units (GE Medical Systems, Milwaukee, WI). Owing to the long duration of this study over 8 years, the DWI protocols changed over time. Stroke neurologists reviewed images using a regional approach. Two stroke neurologists assessed each MRI, interpreting whether there was restricted diffusion in pre-specified brain regions: frontal, parietal, temporal and occipital cortex, caudate, putamen, globus pallidus, thalamus, cerebellar hemispheres, cerebellar vermis, white matter and brain stem. Six stroke neurologists participated in the study, all blinded to the clinical data, including the clinical course and outcome. Raters were instructed to note only abnormalities that were felt to be acute. To ensure compliance with our criteria, we held a training session for all six raters. If there was disagreement between the two reviewers for a specific brain region on an individual scan, a third stroke neurologist reviewed the image and arbitrated.

Statistical Analysis

The primary objective was to estimate the utility of imaging abnormalities to predict poor outcome (mRS score >4 at 3 months post-cardiac arrest). The Cochran test was used to determine if there were significant differences in the frequency of disagreement between reviewers of individual regions. Cohen’s Kappa coefficient was calculated to assess inter-observer reliability between the two reviewers for each set of imaging. We also analyzed the frequency of disagreement between the blinded stroke neurologists’ interpretations and those of the non-blinded clinical neuroradiologists who were involved with the patient’s clinical care in real-time. We retrospectively reviewed patient charts, noting neuroradiology reports of only hypoxic-ischemic injury attributed to the patient’s recent CA as “abnormal.” For comparison, a scan with any abnormal area on the stroke neurologists’ review was considered “abnormal.” The McNemar test was used to detect significance in differences between the neuroradiologists’ and stroke neurologists’ radiographic assessments. The McNemar test was also used to determine the significance of differences between the demographics and co-morbidities of patients who did versus did not have an MRI performed. Inter-observer reliability between these groups was assessed with the calculation of Cohen’s Kappa coefficient. Sensitivities and specificities were calculated from 2 × 2 tables. All statistical analyses were performed using SAS version 9.1.3.


Patient demographics and clinical characteristics for the 80 patients who underwent MRI are compared with those of the 121 patients who did not in Table 1.
Table 1

Patient demographics and clinical characteristics

Patient characteristic

MRI performed (n = 80)

No MRI performed (n = 120)

p value


57 (range 23–81)

62 (range 20–93)



61 % (49) male

62 % (75) male


Coronary artery disease

27 % (22)

31 % (38)


Congestive heart failure

14 % (11)

19 % (23)


Diabetes mellitus

21 % (17)

25 % (30)


Chronic obstructive pulmonary disease

23 % (18)

21 % (25)


Renal insufficiency

9 % (7)

19 % (23)


Hepatic insufficiency

9 % (7)

6 % (7)


Tobacco use

16 % (13)

17 % (21)


Alcohol use

18 % (14)

12 % (14)


Drug use

20 % (16)

10 % (12)


Day of MRI

2 (0–10)



Of the 80 patients included in this study, there were no significant differences in outcome between patients who underwent hypothermia and those who did not [7]. The median time between CA and imaging was 2 days (range 0–10). All but two patients (day 8 and day 10) underwent imaging within 1 week of arrest. Sixty-seven patients had a poor outcome defined as a mRS score >4.

Reliability of Qualitative MRI

Percentages of disagreement among raters for each region are shown in Table 2. These percentages ranged from 11 (caudate, putamen) to 25 % (frontal cortex). No significant differences were observed in frequencies of disagreement among regions on DWI (p = 0.08). Agreement between the two observers for each set of imaging was moderate for “any abnormality” analysis (kappa = 0.49 ± 0.32, p < 0.001), with agreement occurring in 91 % of cases and no significant differences found between observers’ conclusions (p = 0.125). There was a significant difference in qualitative analyses for “any abnormality” on imaging between our study’s blinded observers and the non-blinded neuroradiologists who were involved with the patient’s clinical care (p = 0.001). Agreement on the presence of injury attributable to CA was moderate between these two groups (kappa = 0.50 ± 0.24, p < 0.001). In all 11 cases of disagreement, neuroradiology found no evidence of global hypoxic-ischemic injury, while our study readers did. Nine of these 11 patients had a poor outcome.
Table 2

Percentages of disagreement between readers for each region

Frontal cortex

Parietal cortex

Temporal cortex

Occipital cortex




Globus pallidus

White matter

Cerebellar hemisphere

Cerebellar vermis


25 %

12.5 %

22.5 %

12.5 %

12.5 %

11 %

11 %

19 %

19 %

21 %

22.5 %

20 %

MRI as a Screening and Prognostic Tool by Sensitivity and Specificity

A restricted diffusion abnormality on DWI in any region had a high sensitivity for predicting poor outcome of 98.5 % (95 % CI: 90.9–100 %) but only a modest specificity of 46.2 % (95 % CI: 20.4–73.9 %) (panel a in Table 3). The region of DWI change did not carry any relevance; with any DWI abnormality, a poor outcome occurred. Sixty-six of the 67 poor outcomes were correctly predicted by DWI, whereas, of the 13 patients with a good outcome, only 6 of these 13 had absence of hypoxic-ischemic injury on DWI. Specificity was not improved when only abnormalities in the cerebellum, basal ganglia or cortex were used (panels b and d in Table 3).
Table 3

Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) for any imaging abnormality or specific brain regions


Estimate (%)

95 % CI

a. Any imaging abnormality














b. Basal ganglia abnormalities














c. Cortical abnormalities














d. Cerebellar abnormalities














Of note, one patient who was misclassified by our neuroimaging criteria had a large embolic stroke at the time of CA. This patient was noted to have extensive injury and a poor prognosis because of this simultaneous event; there was no evidence on his neuroimaging of global hypoxic-ischemic injury that could be attributed to his CA. Because our readers were instructed to only dictate a set of images as abnormal if the abnormalities were felt to be because of the arrest, this patient’s images were determined in this strict regard as “normal.” If this patient had not been included in the analysis, the sensitivity of DWI abnormalities for predicting poor outcome in CA would have reached 100 %.


Neuroimaging is a commonly utilized tool for determining the degree of global hypoxic-ischemic injury following CA. Its role in predicting morbidity and mortality, however, has yet to be clearly defined. Although previous studies support the potential utility of neuroimaging in predicting poor outcome after CA, interpretation of these data is limited by small sample sizes, timing of imaging and heterogeneous populations [811].

To date, the two largest and most promising studies of MRI in comatose patients after CA were both quantitative analyses of DWI, and are thus limited by the impracticality of implementation into clinical practice [5, 7]. Unfortunately, most studies noting the generalizable, more clinically relevant, qualitative MRI findings are purely descriptive. Arbelaez et al. [10] retrospectively reviewed the MRIs of ten patients who experienced a prolonged CA, grouping these patients by the timing of their imaging. This study did not draw correlations between imaging characteristics and functional outcome. Kawahara et al. [11] described serial MRIs of three patients, all of whom had an admitting GCS of 3 or 4 and died or remained vegetative, precluding an analyses of imaging changes associated with poor versus good outcomes. Wijdicks et al. examined a subgroup of patients who had an unclear prognosis based on more standard tests [6]. Their prospective study involved ten patients who were comatose after CA and had normal bilateral SSEPs, intact pupillary reflexes and non-specific EEG findings. Patients were imaged once within 1 week of the arrest at different time points, and seven patients had DWI with quantitative measurements of regional ADCs. Two of their patients had focal, minimal imaging abnormalities and a meaningful recovery, whereas the eight who had diffuse extensive imaging changes remained vegetative or died.

Our findings suggest that stroke neurologists’ interpretation of brain DWI MRI as showing evidence of global hypoxic-ischemic injury has a high sensitivity (98.5 %) for predicting poor outcome and can be a useful screening tool (Table 3). It has a poor specificity, however, limiting its overall usefulness. The finding that 54 % of good outcome patients had hypoxic-ischemic changes on neuroimaging warns against a practice of assuming all patients with such abnormalities will do poorly. This may be particularly true for younger, healthier patients and perhaps in patients whose imaging changes are less marked and extensive. As ADC values, and thus diffusion-weighted abnormalities, can return to near normal with successful early resuscitation, this may also have contributed to the low specificity [3, 4]. It is difficult to interpret reliability in this study, as it was likely skewed by the overwhelming majority of patients with a poor outcome, because Cohen’s kappa tends to underestimate agreement when there is a high prevalence of one category. Furthermore, the significant differences found between the interpretations by the study observers and formal neuroradiologists’ reports may reflect the impact of knowing a patient’s clinical history. Clinicians should be wary of using the non-blinded neuroradiologist report of “global hypoxic-ischemic injury” as a basis for predicting a patient’s outcome.

Patients with diffuse multifocal cortical abnormalities have previously been shown to be more likely to have a poor outcome [9]. In our study, this was also the case; no patients with diffuse cortical abnormalities achieved greater than a persistent vegetative state. A novel finding in this study, however, was that in patients with a poor outcome, nearly all had diffusion-weighted changes, the only exception being a patient who died of another cause.

Although our study adds to previous work because of its large sample size and pragmatic qualitative approach, it has several limitations. Despite being several times larger than similar studies, this study remains limited by having too few patients with a good outcome. Less than half of the patients with hypoxic-ischemic coma underwent MR imaging, limiting generalizability. The high prevalence of poor outcome in the patients we studied (84 %) is reflective of the expected prevalence in the population of patients who are comatose after CA, but it may not be reflective of the patient population who will most benefit from having neuroimaging performed for prognostic purposes. The timing of imaging may have also been suboptimal, as a recent study suggests a smaller window of 49–108 h proves best for differentiating between patients who have good and poor neurological outcomes after CA [5]. It is conceivable that the reversibility of DWI abnormalities in CA patients may have led to a lower specificity, given that the cortex is usually most affected given the insufficient blood flow, whereas the deep white matter is relatively resistant to hypoxic-ischemic insults [3, 4]. It is unclear whether the formal imaging interpretations by neuroradiologists created a “self-fulfilling prophecy” with the clinical staff caring for the patients. Neuroradiologists were not blinded to the patients’ clinical data, and thus their interpretations of the radiographic data may have been subject to bias, as opposed to the blinded stroke neurologists.

Three particular strengths of our study bear mentioning. (1) The patients who were included in this study were strictly comatose, and thus their outcome certainly remained in question. Other studies of CA patients with MRI have included patients who were not comatose, and who thus were likely destined for a better neurological outcome. (2) Although the overall number of patients was less robust than one would hope for in a clinical study, it is nevertheless the largest cohort to date. (3) This study emphasized what is done in actual clinical practice; rather than using a computer-generated output of apparent diffusion coefficient depression, we approximated what is done on a daily basis by practicing stroke specialists, including a comparison of the formal neuroradiologists’ impressions that occurred in real-time. These qualities enhance the generalizability and impact of the results.

In conclusion, there is clearly importance to experienced vascular neurologists interpreting neuroimaging in comatose CA patients, as is widely done in clinical practice. As we move toward establishing neuroimaging as a validated tool for prognostication, future studies should emphasize this useful and real-time aspect of clinical care and its potential for introducing bias because of its inherent subjectivity.

Copyright information

© Springer Science+Business Media, LLC 2012