|Home | About | Journals | Submit | Contact Us | Français|
The early assessment of the likelihood of neurological recovery in comatose cardiac arrest survivors remains challenging. We hypothesize that quantitative non-contrast CT, combined with neurological assessments, are predictive of outcome.
We analyzed datasets acquired from comatose cardiac arrest patients who underwent CT within 72 hours of arrest. Images were semi-automatically segmented into anatomical regions. Median Hounsfield units (HU) were measured regionally and in the whole brain (WB). Outcome was based on the 6-month modified Rankin Scale (mRS) score. Logistic regression was used to combine Glasgow Coma Scale score measured on Day 3 post-arrest (GCS_Day3) with imaging to predict poor outcome (mRS>4).
WB HU (P=0.02) and the ratio of HU in the putamen to the posterior limb of the internal capsule (PLIC) (P=0.004) from 175 datasets from 151 patients were univariate predictors of poor outcome. Thirty-three patients underwent hypothermia. Multivariate analysis showed that combining median HU in the putamen (P=0.0006) and PLIC (P=0.007) were predictive of poor outcome. Combining WB HU and GCS_Day3 resulted in 72% [61–80%] sensitivity and 100% [73–100%] specificity for predicting poor outcome in 86 patients with measurable GCS_Day3. This was an improvement over prognostic performance based on GCS_Day3≤8 (98% sensitive but 71% specific).
Combining density changes on CT with GCS_Day3 may be useful for predicting poor outcome in comatose cardiac arrest patients who are neither rapidly improving nor deteriorating. Improved prognostication with CT compared to neurological assessments can be achieved in patients treated with hypothermia.
Accurate prediction of neurological recovery after cardiac arrest remains problematic. Prognostic techniques have traditionally relied on clinical examinations, electrophysiological measurements or biochemical changes1, 2. However, these are most effective only for patients either rapidly recovering who will have a good outcome, or severely injured who will have a poor outcome. Although some suggest that a Glasgow Coma Scale (GCS) score <5 within the first 48 hours is predictive of poor outcome3, 4, others have demonstrated that GCS at admission was a poor predictor5. At 72 h, GCS<6 was 67% sensitive at 85% specificity6 for predicting poor prognosis. The American Academy of Neurology (AAN) Practice Parameters evidence-based review reported that clinical findings associated with poor outcome 3 days after resuscitation were absent pupillary light response (PLR) or corneal reflexes (CR), or extensor posturing or no motor response to pain7. The guidelines concluded there was insufficient evidence to support or refute the use of brain imaging for prognostication.
Diffusion-weighted imaging (DWI) is sensitive to brain injury after transient global ischemia. Two large independent studies showed that severe reductions in the apparent diffusion coefficient (ADC), a measurement of tissue water diffusivity sensitive to cytotoxic edema, are linked to poor long-term outcome8, 9. Although these MRI findings are extremely promising, MRI can be difficult to acquire in critically ill patients due to the necessary monitoring and treatments10, 11. This has led some to suggest using changes in non-contrast computed tomography (NCCT) for prognostication12. Studies showed that patients with good neurological recovery had a higher gray matter to white matter Hounsfield unit (HU) ratio than those who did poorly13–15; however these studies had relatively small sample sizes.
Our goals were (1) to determine whether early changes in NCCT predict poor functional recovery in comatose post-cardiac arrest patients in a large patient cohort, (2) to examine spatio-temporal patterns in the evolution of brain injury and recovery and (3) to determine whether imaging combined with clinical findings provides better prediction of poor clinical outcome vs. using the clinical findings alone.
An Institutional Review Board-approved single-center prospective observational study of 500 patients with non-traumatic coma was performed from 2000 to 2007. Critically ill patients admitted to intensive care units who had an admission diagnosis of coma, or had an ensuing diagnosis of coma during admission were eligible for inclusion. To meet the definition of coma by Levy et al16, patients failed to open their eyes either spontaneously or in response to noise, expressed no comprehensible words, and neither obeyed commands nor moved extremities appropriately to localize or resist painful stimuli. Patients were excluded who were found to have a traumatic cause for coma. We excluded patients less than 18 years of age, as younger age groups may have different outcomes. Patients who were kept in a pharmacological coma (e.g. post-operatively), or who were found to be in a comatose state upon discontinuation of medications were also excluded on the basis that there would be no method for dating the inception of the comatose state, and daily clinical data would be of unclear benefit. Within this cohort, 200 patients were comatose secondary to hypoxic-ischemic brain injury. The decision to acquire additional testing, including neuroimaging, was at the discretion of the treating clinicians. Patients underwent routine clinical evaluations, including detailed neurological assessments, on days 0, 1, 3 and 7, and a modified Rankin Scale (mRS) score was obtained at 3 and 6 months in surviving patients. Neurological assessments included GCS, PLR, CR and GCS motor score (M). Of the 200 patients, 166 underwent CT imaging within 7 days of arrest. Patients were excluded if hemorrhage or a previous stroke was noted, if imaging was of poor quality, or if the first imaging study was performed >72 hours from arrest. This resulted in 151 patients. Some patients underwent repeat imaging, resulting in a total of 175 imaging studies performed within 3 days of arrest.
Due to the observational nature of the study, the imaging parameters of the 175 studies varied, consisting of 5 mm (N=169), 3 mm (N=5) and 2 mm (N=1) thick slices with median [interquartile range (IQR)] in-plane resolution of 0.43×0.43 mm2 [0.43×0.43–0.45×0.45 mm2]. NCCT data sets were non-linearly co-registered17 to the ICBM-452 T1 5th Order Polynomial Warps Atlas18. Using the ICBM probabilistic atlases19, 20, masks were generated for the following regions: cerebellum, frontal lobe (FL), insula, occipital lobe (OL), parietal lobe (PL) and temporal lobe (TL), caudate nucleus (CN), putamen, thalamus and white matter (WM). Further segmentation of white matter regions into the posterior limb of the internal capsule (PLIC), corpus callosum (CC) and corona radiata (CorRad) was performed using the JHU ICBM-DTI-81 white matter atlas21, 22 distributed as part of FMRIB (Functional MRI of the Brain) Software Library (FSL)23. Only tissue with > 50% probability was included for analysis (see Figure 1). Median Hounsfield units (HU) were measured in each region and the whole brain (WB). To exclude cerebral spinal fluid and artifacts, analysis was limited to HU between 15 and 100.
Spatial differences among the regions were examined using a one-way analysis of variance (ANOVA) with post-hoc Student-Newman Keuls test. Clinical outcome was defined by modified Rankin Scale (mRS) as either good outcome (mRS≤4) or bad outcome (mRS>4). Differences in patients with good outcome were compared (two-tailed Wilcoxon rank-sum test) to patients with poor outcome . We also compared differences in demographics (two-tailed Wilcoxon rank-sum test) between patients who were included in this study and the patients in the database who were excluded. For categorical variables, a two-sided Fisher's Exact Test was used. Since the AAN Practice Parameters (AANPP) require a neurological exam to be performed on Day 3, we did subset analysis between patients for whom GCS could be measured on Day 3 (Group I) and patients for whom GCS could not be obtained (due to death or “comfort measures only” status) on Day 3 (Group II). Subset analysis examining differences between early and late imaging was done for patients imaged within 24 hours and beyond 24 hours. Logistic regression was performed using the bias-reduction method24, 25 to investigate the relationship between poor or good outcome and GCS at baseline (GCS_Day0) (acquired at admission or Day 1), GCS_Day3, age, sex, time-to-CT (days from last seen well), whether the patient had a shockable rhythm, whether the patient was treated with hypothermia and WB HU. We also investigated the prognostic performance of poor recovery using AANPP7, in which poor outcome was classified by absence of CR, PLR or M2. For multivariate analysis, the decision regarding which parameters to include in a forward stepwise method was based on the Akaike Information Criterion (AIC), a measure that is a function of both training error and complexity26. Multivariate analysis was performed to evaluate the contribution of anatomical regions for prediction of outcome. Sensitivity and specificity estimates were calculated using jack-knifing or leave-oneout approach27 for which the model coefficients for a particular imaging dataset were calculated using all other datasets in order to avoid evaluating the model performance on the same data upon which it was trained. Sensitivities were calculated for the case of minimum false positives, and maximum number of true positives. 95% confidence intervals were calculated according to the efficient-score method28, 29. The sensitivities of the different models at the same specificity were compared using a McNemar test30.
Logistic regression analysis was repeated excluding hypothermia-treated patients to evaluate the influence of hypothermia on patient prognostication. Hypothermia was adopted at our institution in 2002. The reasons some of our patients did not undergo hypothermia included being outside of the time window and being too unstable from a cardiovascular standpoint.
Table 1 shows the demographics of the 151 patients with usable imaging performed within 72 hours of arrest, dichotomized by good or poor neurological outcome at 6 months. There was no significant statistical difference (P>0.05) between included patients and those who were excluded for not having usable CT scans (N=49) in age (61.5±16.3), female sex (37%), shockable rhythm (35%), treatment with hypothermia (12%), GCS_Day3 (3 [range 3–15]), absent PLR (22%), absent CR (22%), M≤2 (72%) and 6 month mRS (6 [range 0–6]), Patients who did not undergo imaging had significantly lower (P=0.02) GCS_Day 0 scores (3 [range 3–7]) and a significantly lower (P=0.02) proportion of patients for whom GCS_Day3 could be measured (37%). For the included patients, there was a significantly higher proportion of cases with shockable rhythms in the good vs. poor outcome group. There was a significant difference in terms of the percentage of patients treated with hypothermia in the good outcome group. GCS_Day0 was measured on Day 0 for 125 patients, and on Day 1 for 26 patients. There was no significant difference in GCS_Day0 between outcome groups, but, for the subset of patients (N=86) who had GCS_Day3 measured (Group I), there was a significant difference in GCS_Day3 between good and poor outcome groups. For Group I, there was still no statistically significant difference (P>0.05) in GCS_Day0 between poor and good outcome groups. There was also a significant difference in the distribution of patients on Day 3 with absent CR, GCS motor score M≤2 or predicted to have poor outcome according to the AANPP clinical signs. For Group II, patients for whom neurological findings on Day3 could not be obtained (N=65), 59 died before Day 3 and 6 were made comfort measures only (but had not yet died) by Day 3. Group II patients had significantly lower (P=0.01) admission GCS_Day0 scores (3 [range 3–8]) compared to Group I (4 [range 3–8]), and a lower incidence of shockable rhythm (19% vs 41%, P=0.007). All Group II patients had a 6 month mRS=6. Age (60.5±14.1 vs 57.9±19.3), female sex (34% vs 46%) and hypothermia treatment (27% vs 15%) were not statistically different (P>0.05) between Groups I and II.
CC and CorRad exhibited significantly lower median HU (P<0.001) while TL and OL exhibited significantly higher median HU (P<0.01) than the other regions. A significant difference (P<0.05) was found between all regions except the insula for patients with good versus poor outcome (Table 2). The ratio of putamen/PLIC was significantly different between the two outcome groups. No statistically significant difference was found between Group I (N=102) and Group II (N=73) for whole brain HU (P=0.09), and white matter (P=0.07), CC (P=0.21), CorRad (P=0.25), FL (P=0.23), insula (P=0.07) and CN (P=0.10), but differences were found for the other regions (P<0.05).
119 NCCT scans were performed within 24 h (Early), and 42 scans between 24–72 h (Late). For 14 scans, the time of arrest could not be determined well enough to discriminate between whether the patient was scanned within 24 h. Regionally, only the putamen showed significant (P=0.03) reductions in the Late group (29.5 [27.7–31.9]) compared to the Early group (31.0 [29.0–33.0]). For all other regions, no difference was found. For Group I, 59% of the studies were performed within 24 h, 15% between 1–2 days, 15% between 2–3 days. For Group II, 81% were done within 24 h, 12% between 1–2 days and 4% between 2–3 days. Time from the event could not be determined accurately for the remaining datasets. For the Early cohort, there was a significant difference in median HU between outcome groups for WB, WM and all regions except the insula, and PL (Table 3). In contrast, in the Late cohort, for which there were only 3 scans associated with good outcome, no significant difference was found between outcome groups. However, a significant difference was found between Putamen/PLIC for the two groups.
Of age, sex, GCS_Day0, time-to-CT, shockable rhythm, hypothermia and WB, WB (P=0.02), lack of hypothermia treatment (P=0.003) and non-shockable rhythm (P=0.01) were significant univariate predictors of poor outcome. GCS_Day3 (P<0.0001), CR absence (P=0.02), and M≤2 (P<0.0001) were significant univariate predictors of poor outcome, while PLR absence was not (P=0.14) in the 102 datasets analyzed from 86 patients. Prognosis of poor outcome using AANPP was also significant (P<0.0001). Multivariate analysis using only clinical parameters found that not receiving hypothermia treatment and GCS_Day3 were significant predictors (P<0.05) of poor outcome. WB (Spearman's ρ=−0.25, P=0.0009) and GCS_Day3 (Spearman's ρ=−0.36, P=0.0002) significantly inversely correlated with 6 month mRS. Performance of WB, GCS_Day3, PLR, CR and M2 as univariate and multivariate predictors of poor outcome in Group I are shown in Table 4 and Figure 2. Our sensitivity and specificity results on Day 3 using a GCS cutoff of 6 and PLR are similar to those described in the AAN Practice Parameters7. No threshold of GCS_Day3 was 100% specific for predicting poor outcome in all patients. Sensitivity of absent CR is greater than that reported in the AAN Practice Parameters7. However, we did not obtain a 0% false positive rate for M≤2. Excluding 29 datasets from 23 patients treated with hypothermia resulted in no false positives for M≤2 criterion and, by extension, the AANPP (Table 4) in the 73 remaining datasets. Excluding hypothermia-treated patients also dramatically increased the prognostic performance of GCS_Day3 to 100% sensitivity at 100% specificity when using a cutoff of 8.
The ratio of putamen/PLIC was a significant univariate predictor (P=0.004), while the ratio of CN/PLIC was not (P=0.56). Using AIC for selecting which tissue regions to include in a multivariate model, the optimal model included the putamen (P=0.0006) and PLIC (P=0.007). Combining GCS_Day3 with WB, putamen and PLIC or putamen/PLIC increased sensitivity. Adding WB to M≤2, or AANPP improved the performance of M≤2 prediction for specificity=100%, and by extension AANPP, but not PLR absence or CR absence. Based on the minimum AIC, when including GCS_Day3 and different brain regions, the optimal model included GCS_Day3 (P=0.0002), cerebellum (P=0.02), insula (P=0.09), frontal lobe (P=0.06) and WB (P=0.12), resulting in significantly higher sensitivity (P<0.001), in predicting poor outcome while maintaining the same specificity as GCS_Day3+WB model (Table 4). When excluding hypothermia patients, the optimal model consisted of GCS_Day3 (P=0.01), resulting in 100% sensitivity at 100% specificity. However, this model does not reach 100% specificity when including all patients.
Our results show that neuroimaging combined with the neurological exam can improve prediction of clinical outcome for patients who survive more than 24 hours after cardiac arrest. Consistent with MRI studies that showed that global ADC reductions are predictive of poor outcome8, 9, our study shows that changes in the brain due to cytotoxic edema resulting in decreases in median whole brain HU are also predictive. Not surprisingly, the initial rhythm of cardiac arrest (i.e. whether it was shockable or not) and GCS_Day3 were also found to be significant predictors.
Torbey et al showed that combining neuroimaging (ratio of CN/PLIC) with clinical parameters (reversed GCS and duration of arrest (DAR)) can improve patient prognostication31 in a cohort of 32 patients. Our study did not include DAR because it could not be determined accurately in the majority of our patients since it was not prospectively recorded. Our study also included only patients with a maximum admission GCS of 8, while the Torbey et al study included patients with GCS scores up to 15, suggesting our patients consisted of a more critically ill population. In addition, the CN HU in our study (27 [25–29]) was lower than that reported in previous studies13, 14, again indicative of a more severely injured group. Our study strictly included only truly comatose patients, which may help to explain the high mortality rate in comparison to other studies which may have included post-cardiac arrest patients who, in distinction, had a depressed level of consciousness but who were not comatose. This led to the limited number of patients with good outcomes that in turn resulted in the large confidence intervals in our specificity for predicting poor outcome. Further studies are therefore necessary, involving greater number of patients to more accurately estimate the specificity of tools for prognosticating poor outcome in comatose cardiac arrest patients before they can be used for clinical decision-making. Such studies should be limited to those patients in whom the prognosis is truly in question, excluding those rapidly awakening (and therefore with higher GCS scores) or brain dead.
Regionally, we found that decreased HU's in the putamen and the PLIC were significant predictors of poor outcome, not surprising given their important roles in motor function and the emphasis of mRS on motor recovery32. Further, the early selective sensitivity of the basal ganglia to global ischemia has been well documented12. We found that the ratio of the putamen/PLIC was predictive of poor outcome, but not CN/PLIC, consistent with a previous study by Torbey et al13. Another study14 by Choi et al found that CN/PLIC within 24 h in 28 patients was predictive of poor outcome. Differences in our findings are likely due to differences in outcome measures, patient cohorts and timing of CT. Subset analysis of patients imaged early or late showed that putamen/PLIC was significantly associated with poor recovery only for patients scanned >24 hours post-arrest. In comparison, the other regions no longer exhibited significant differences in the late group, despite being significantly different within 24 hours (Table 3), but this is likely due to limited number of patients with good recovery (N=3) in the late group. Of all the regions examined, only the putamen demonstrated a significant decline between early and late CT HU values.
Our regional results are consistent with our MRI study in comatose post-cardiac arrest patients8. In the MR study, for 48 patients imaged within 72 h, ADC values tended to be higher in the patients who achieved good outcomes (N=6) compared to those with poor outcomes (N=42). These differences were not statistically significant for all regions, likely due to the limited sample size, which resulted in the MRI study being underpowered for regions that did not demonstrate severe ADC reduction. Further, there are temporal differences in the two studies. In the MRI study, 21% of the studies were performed on Day 0, while for this CT study 68% were performed on Day 0. We speculate that pseudonormalization in ADC as a result of reperfusion, as demonstrated in animal models of global ischemia33, may diminish detectable differences in ADC values between patients with good and poor outcomes the later the MRI is obtained. To properly compare the strengths and weaknesses of the two imaging modalities, and to understand the physiological implications of differences in findings, one would need a prospective study in which both modalities were acquired at similar time points.
For patients for whom Day 3 neurological exam data was collected, our sensitivity and specificity results for the predictive performance of clinical signs for poor outcome were consistent with previously published findings7. Discrepancies in predictive performance when using either M≤2 or AAN Practice Parameter recommendations were found for patients who underwent hypothermia therapy, suggesting caution is needed when using motor response for prognostication purposes given that hypothermia-treated patients may have poor motor responses on Day 3 despite eventual good outcome2, 34. For patients for whom GCS_Day3 was obtained, combining neuroimaging data with neurological exam data improved prognostication of patient outcome over use of imaging or clinical data alone. For patients in whom prognostication is most difficult, i.e. those who are not rapidly deteriorating, imaging may therefore provide supplemental information regarding the severity of brain injury. Although sophisticated brain segmentation techniques can significantly increase prognostic performance of our models, inclusion of a single easily measured parameter, i.e. median whole brain HU, was able to offer significant improvements over the sole use of GCS measured 3-days post-arrest. Multivariate models involving regional changes, however, may be useful for providing insight into the pathophysiological processes following global hypoperfusion and resuscitation. Exclusion of hypothermia patients also improved prognostication for the multivariate models, although using GCS3 ≤8 provided 100% sensitivity at 100% specificity. This suggests that patients not treated with hypothermia who are still comatose at Day 3 will likely have a poor outcome. For patients treated with hypothermia, this may be inaccurate. In addition, not all hypothermia-treated patients had a good outcome. Of the 33 patients who received hypothermia, 79% had a poor outcome. The percentage of patients with good outcomes (21%) is much lower than what has previously been reported in two randomized studies of hypothermia treatment in survivors of cardiac arrest due to ventricular fibrillation, in which 43–55% of the patients treated had good outcomes35, 36. We suspect this may be due to the fact that only 55% (N=18) of our treated patients had shockable rhythms, and of these only 33% (N=6) had good outcomes. The addition of imaging increased the specificity of the neurological exams and maintained high specificity even for patients treated with hypothermia. This is especially important as hypothermia increasingly becomes part of standard management in these patients.
The observational nature of this study introduces potential bias in terms of the decision to perform imaging and the timing of the scan. Ideally, imaging would be obtained at admission and at 3 days. Many patients died due to withdrawal of care, a common problem in these studies, resulting in a self-fulfilling prophecy of poor outcome in patients who may have otherwise recovered2. Clinicians in this study were not blinded to the CT or any other data, and thus abnormal CTs might have swayed the treating team. However, the treating team did not do a quantitative evaluation of HUs, and thus any patient management decisions based on imaging findings would likely have been based purely on drastic gross changes, which were not common in our population. Prospective studies are clearly needed with pre-specified imaging and neurological examination times, as well as sufficient time to allow for recovery in patients for whom the outcome is in question.
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 citable 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.