Abstract
Background: There are no randomized trials comparing intra-arterial (IA) therapy with best medical treatment for acute ischemic stroke. To assess potential benefit from this therapeutic approach, we performed a systematic review of published IA series. Because outcome from stroke is highly dependent on baseline characteristics, we compared results against prognostic models adjusted for admission National Institute of Health Stroke Scale (NIHSS) scores and age.
Methods: We selected articles from MEDLINE and Cochrane Databases based on specified criteria that included 3-month clinical follow-up. Outcome functions from prognostic models were generated and difference from prediction calculated for each study. Best and worst mortality performers were identified and assessed for factors that distinguished them.
Results: We identified 27 reports with 30 treatment series representing 1,117 patients. Percent difference from predicted outcomes varied from −51 to +24.6% for mortality and −30.3 to +28.7% for good functional outcome. A mean overall difference in the percent that died compared with prediction was 0.25% (SD: ±3.5%; 95% CI: ±0.53) and in the percentage of those that achieved a good functional outcome compared with prediction was −0.15% (SD: ±2.7%; 95% CI: ±0.44). The quartile of better mortality performers relative to worst performers had a 4.8-point more severe NIHSS score at baseline (p = 0.028) and employed 50% lower doses of the most frequently used thrombolytic urokinase (p = 0.0034).
Conclusion: We found considerable variability and lack of evidence for a net improvement in outcome after intra-arterial therapy relative to predicted natural history, substantiating the need for a prospective comparison with best medical therapy. The features associated with better performers identified here may be useful in designing such a trial.
Intra-arterial (IA) therapy is utilized to treat patients who experienced an acute ischemic stroke and are excluded from receiving IV recombinant tissue plasminogen activator (rt-PA) or have not recanalized after IV therapy. Prolyse in Acute Cerebral Thromboembolism (PROACT) II,1 a study of IA pro-urokinase, remains the largest randomized trial of IA therapy in this setting, but the control group did not represent best medical therapy as they did not receive anti-platelet therapy and were further subjected to an IA catheter and heparin. The recent approval of the MERCI (Concentric Medical) catheter device,2 without inclusion of a best medical therapy control group, has generated controversy.3 In lieu of such data, and with the widespread use of IA approaches, we undertook a systematic review of published case series of IA therapy to better understand the range of outcomes reported. Comparison between such studies is problematic as outcome from stroke is known to be highly dependent on factors such as baseline stroke severity and age.4,5 We therefore organized these reports based on their baseline characteristics and compared these outcomes with that predicted from prognostic models of outcomes in stroke. We then identified the best and worst performers relative to the predicted outcome for a given baseline stroke severity and performed additional analysis of characteristics that differentiate them.
METHODS
Search and selection criteria.
The MEDLINE database and the Clinical Trials Registry of the Cochrane database were systematically searched with the assistance of a medical librarian using keywords “Catheter based therapy AND Stroke,” “Intra arterial thrombolysis,” “Intra arterial acute stroke,” and “endovascular treatment AND acute stroke.” Mesh word search with “Intra-arterial infusions” and Keyword search “Stroke” was also employed. Both authors independently reviewed each report.
Studies were selected for further analysis based on the following criteria: 1) The study was performed in humans and at least 10 subjects were represented. 2) The baseline National Institutes of Health Stroke Scale (NIHSS) score and patient age were included either as the median or mean for the group as whole or individual data so that we could perform the calculations. 3) Functional outcomes, mortality, or both were provided at 3- to 6-month follow-up. 4) Sufficient details on the IA methods were presented that we could confirm that patients received IA therapy and that methods could be abstracted for additional analysis. We contacted authors directly to provide additional details.
Data extraction.
For each study that fit the selection criteria, author names, number of patients, mortality percentage, and percentage achieving modified Rankin Scale (mRS) score of 0 to 2 were catalogued. When more than one treatment series was reported in a single study, we included only those series that received an active IA procedure. As the purpose of this study was to determine the relationship between actual outcomes from IA therapy to predicted natural history, for “intention-to-treat” studies, we calculated results for those patients that actually underwent IA treatment when that information was available.
Data analysis.
Mortality model (Model I).
A model4 for mortality based on median baseline NIHSS score and mean age and a corresponding sine-square function (Model I) was used for mortality comparisons. In this model, developed from 16 placebo arms of randomized control trials, baseline NIHSS score accounted for 91% of the variance in predicting 3-month mortality. Addition of age to NIHSS score accounted for 95% of the predictive variance.4 These 16 trials had median NIHSS score of >4 and <20.4
Functional outcome model (Model II).
Nearly all of the IA reports selected for inclusion used the mRS score as the long-term functional outcome measurement, whereas none of the prognostic models used mRS as the outcome variable. The TOAST (Trial of ORG 10172 in Acute Stroke Treatment)6,7 study used the 20-point Barthel Index (t-BI) and the Glasgow Outcome Scale and provided outcomes related to stroke subtype and baseline NIHSS (n = 1,281).6,7 We selected the natural history data derived from the TOAST outcome database to compare with the IA data. As the trial showed no statistically significant differences in outcomes between treated and placebo arms, they collapsed the groups together. Note, however, that preplanned subgroup analysis did show that treated patients in the large vessel category had statistically significant better outcomes.6
It was necessary to perform additional data manipulation to make TOAST data7 compatible with mRS. The t-BI was multiplied by 5 (to generate a 100-point scale compatible with the BI), and “good outcome” defined as t-BI of 12 to 20 was therefore changed to 60 to 100. A relationship between the mRS and the 100-point BI has been established.8 Inspection of the relationship between BI and mRS8 showed that a BI of 60 to 100 is virtually superimposable over mRS of 0 to 2, and therefore we compared TOAST7 “good outcomes” with studies using mRS 0 to 2 as their functional outcome measure without additional correction factors.
TOAST did not publish equations to represent their “good outcome” data, and so we generated a mathematical representation of their “good outcome” curve. As IA therapy is generally applied to patients with known or suspected larger vessel occlusion, the “nonlacunar” plot was selected. A fourth degree polynomial function was fit to the curve (figure E-1 on the Neurology Web site at www.neurology.org) representing good outcome (Model II).
Although these functions are meant to represent patients with NIHSS scores ranging from 0 to 42, midway through the 6-year recruitment period, patients with NIHSS scores >15 were excluded and 162 (out of 1,275) TOAST patients had scores between 16 and 42.6
Data management, programming, calculations, and plotting.
MATLAB version R2006b was used for all programming tasks of modeling, data management, calculation, plotting, and exporting of data. Curve fitting and Statistical Toolboxes available as extensions to MATLAB were used. Statistical calculations were verified with Sigmaplot 9.
Each study was assigned a letter based on baseline median NIHSS score (or mean NIHSS score when median was not available). Based on the study NIHSS and mean age, we computed the predicted mortality from Model I.4 Absolute and relative differences (absolute difference/predicted) between the reported mortality and the predicted mortality and good outcome were computed. A modified funnel plot10 was generated to assess for systematic publication bias related to small sample size in the selected studies. To calculate means and CIs, two additional transformations were performed. First, we calculated arcsine-square root transformations, as suggested for proportional data.11 Mean, SD, and SEM of these results were then transformed back from degrees to percentages.11 Untransformed data can be found in appendix E-1, whereas transformed results are presented in the text. Owing to wide variability of sample sizes, we also report a “weighted mean” that was calculated by multiplying the difference in angles by sample size and dividing by total number of subjects.
Additional analysis of best and worst mortality performers.
After the above calculations were performed, we segregated the studies into quartiles based on the magnitude of difference from predicted natural history for further analysis. We selected mortality as the outcome because it required no data manipulation. We analyzed which variables were associated with the quartile of studies that included the best and worst mortality performers. The factors analyzed were baseline NIHSS score, age, arterial territory, date of publication, type of IA, and dose of thrombolytic agents. These values were compared between groups and differences assessed with Student t test (two tailed) or χ2.
RESULTS
One hundred sixty-five articles were identified by the initial search criteria, of which 27 reports with 30 different treatment series fit the selection criteria listed in Methods. These reports represented 1,117 patients. Table 1 includes the letter identifications, the corresponding reference numbers, and other characteristics of the studies analyzed. PROACT II1 and PROACT21 heparin “placebo arms” were not included in any analysis because they did not represent an intent to treat. When the same series was published more than once, only the later publications2,16,23,27,28,31,38,40 or a more complete series33,39 were considered. The 27 selected reports were published from 1998 to April 2006 and came from eight different countries (Canada, Germany, Italy, Japan, Korea, Latvia, Switzerland, and USA). IA therapies encompassed thrombolysis with and without adjunctive agents, angioplasty, mechanical thrombectomy, and intravascular ultrasound augmentation. Funnel plot did not suggest a systematic bias of better mortality outcomes related to smaller sample size (figure 1).
Table 1 Series identification, baseline characteristics, mortality, and functional outcomes
Figure 1 Funnel plot of all studies
Funnel plot of all studies showing odds ratios (ORs) vs sample sizes. The OR for a study was calculated as p/(1 − p) divided by q/(1 − q), where p is the percent mortality reported by study and q is the percent mortality predicted by the model. Horizontal lines indicate the extent of the CI. Vertical line is the pooled OR (0.84). Pooled OR was calculated by ([Σnp] * [Σn(1 − q)])/([Σnq] * [Σn(1 − p)]), where n is the sample size in each study.
Mortality.
Three-month mortality was reported in 28 series representing 1,047 patients. Figure 2 shows mortality as reported by different studies superimposed on a three-dimensional mortality surface plot predicted by Model I.4 Percent difference from predicted mortality varied from −51 to +24.6%. The mean difference in the percent of patients in all series that died compared with prediction was 0.25% (SD: ±3.5%; 95% CI: ±0.53; untransformed data presented in appendix E-1, table E-1, and figure E-3). The weighted mean difference in mortality from prediction was 0.19%.
Figure 2 Series mortality superimposed on Model I mortality surface
Color surface represents mortality predicted from Model I [Mortality = (sin(0.035 × NIHSS + 0.011 × Age − 0.76))2; from Uchino et al.4] Letters (A through Z, a through d) show mortality of a series plotted at the series median baseline National Institutes of Health Stroke Scale (NIHSS) score and mean age. Surface was made semitransparent to view the lighter letters indicating lower than predicted mortality below the surface (figure E-2 has been redrawn from a different viewing perspective to better visualize the relationship between each series and the model prediction).
Functional outcome.
Twenty-six of 30 series reported outcomes in terms of mRS 0 to 2 representing 989 patients. These outcomes were plotted against the model of good outcome predicted by Model II. Figure 3 is the plot showing the percentage of patients achieving mRS of 0 to 2 at 3 months on the Model II good outcome curve. Again, there was considerable variability among outcomes. Percent difference from predicted good functional outcome varied from −30.3 to +28.7%. The mean difference in percent of patients in all series that had a good functional outcome compared with prediction was −0.15% (SD: ±2.7; 95% CI: ±0.44; untransformed data in table E-1 and figure E-4). The weighted mean difference in outcome from prediction was 0.35%.
Figure 3 Model II fit to TOAST7 “good outcome” data and superimposed series good outcome
Line represents the model fitted to TOAST7 “good outcome” data. Percentage of subjects in each series achieving an mRS score of 0 to 2 plotted against median baseline National Institutes of Health Stroke (NIHSS) Scale score.
Additional analysis.
The seven studies with worst relative mortality with respect to Model I were identified with the letters (beginning with worst) F, P, C, B, G, T, and Z. The studies with the best relative mortality rates when compared with Model I were (beginning with best) b, W, c, A, R, S, O, d, and J. Series A and R were intention-to-treat studies.12,25 In series R we were unable to determine outcomes separately for those patients that actually received IA. In series A, mean age of treated patients could not be determined. These studies were therefore excluded from this analysis, leaving seven to be compared in the best and worst quartiles. Relative to predicted outcomes, the best seven performers (without series A and R) would have prevented 44 deaths of 205 patients (21%), whereas the worst seven performers would have resulted in an excess of 29 deaths of 181 patients (16%).
Better outcomes with respect to mortality relative to model prediction was associated with a more severe baseline NIHSS score (table 2; 21.1 ± 4.08 vs 16.3 ± 3.14; p = 0.028). To examine this finding further, baseline NIHSS score was plotted against difference from predicted outcome. There was a significant association for mortality and functional outcome in both models (Model I absolute difference: r2 = 0.31; p = 0.002; Model I relative difference: r2 = 0.18; p = 0.024; Model II absolute difference: r2 = 0.14; p = 0.059; Model II relative difference: r2 = 0.15; p = 0.047).
Table 2 Analysis of best performing mortality quartile compared with worst performing mortality quartiles in relation to Model I4
Urokinase was the most frequent thrombolytic used in the selected studies. A lower median (or mean) dose of urokinase was associated with a better outcome (p = 0.0034; table 2). Indeed, the worst performers employed nearly 50% higher median/mean dose. Neither mean age of the subjects, use of adjunctive mechanical disruption of the thrombus, nor date of publication differed between the best and worst performers. Admission glucose was reported in only two studies, and time to treatment was reported inconsistently and could not be compared among most studies. There was no apparent difference between arterial territories represented, with basilar artery territory represented in similar percentage of the best and worst performers. Two of the best performers used glycoprotein IIb/IIIa antagonists in some fashion while they were not mentioned in the worst performers. The only report that used fixed dose of IA thrombolytic without mentioning recanalization as a goal was in the best performing group.36
DISCUSSION
We compared the results of IA case series with predicted outcome using models of mortality and functional outcome. CIs indicated no substantive difference from natural history.
Given the lack of randomized trials, there have been prior efforts to assess the impact of IA therapy. A review of treatment specifically for basilar artery occlusion41 concluded that there was no clear benefit for IA compared with IV thrombolytic therapy beyond the 3-hour time window, but lacked a control population to determine if either approach is better than no thrombolytic treatment. An earlier systematic review of IA therapy42 assessed studies that largely predated the NIHSS and hence could not systematically control for baseline stroke severity, but did compare IA results with a pooled control population. This review suggested IA therapy may be beneficial.42 However, the majority of their control patients were treated with an IA catheter as part of the placebo arm of the two PROACT studies, and it is not clear how representative this group is to patients treated with “best medicine.”
Although no method short of a randomized trial can match cohorts for multiple variables, the method employed here utilizes the major predictive factors related to outcome. Note that virtually all published individual case series compare their results with prior case series. However, without continuous functions, these studies cannot adequately adjust for the impact of baseline stroke severity or age.
The accuracy of our results depends critically on the relevance of the models to the populations represented in these IA trials. It is likely that the mortality results are most reliable, as the mortality outcome model was similarly based on multiple studies.4 The functional outcome results are less certain, given they are based on individual patient data7 rather than multiple clinical series. Although we would have preferred to use the same approach for functional outcome as used previously,4 those authors did not find a relationship between median NIHSS score and good outcome, but the number of studies available to them for this measure was small.4 As a result, we restricted our major secondary analysis of features associated with good outcome to mortality.
There are factors that may cause the model derived from data in TOAST7 to overestimate good outcome, thereby causing the IA studies to appear worse than they may actually be. Both the treated and the untreated groups were included, yet preplanned subgroup analysis indicated benefit from treatment in the large artery group.7 The 24-hour time window (mean 15.7 hours) of entry into TOAST is longer than many of the IA series. There appears to be a reduction in median NIHSS scores in series with longer time windows.4 Therefore, a systematic shift to the right of the TOAST function is conceivable, and projection back to shorter time windows might yield a less optimistic outcome model. Note also that the TOAST functional outcome included a combined mRS and Glasgow Outcome Scale,7 while no IA study employed that combined measure. We are not certain what effect the combined score would have when compared with NIHSS alone, but it seems unlikely that more strict criteria would identify more patients with a good outcome.
Because it is based on a relatively small number of studies, our secondary analysis of factors associated with good outcome has to be considered tentative. However, the association of the severity of baseline NIHSS with difference from predicted outcome was a consistent finding across models. Poorer functional outcomes and higher mortality relative to that predicted were seen in those series with the lower median NIHSS scores. Although less severely affected patients likely do better overall than more severely affected patients, less severe patients had a worse outcome relative to that predicted. The importance of this finding relates to selection of patients to undergo IA therapy, suggesting that it should be reserved for more severe patients who will do poorly without intervention.
The better performers were also characterized by considerably lower doses of urokinase. Although most studies employed some element of mechanical thrombus disruption, the only study that used fixed dose thrombolytic without regard for success or extent of recanalization36 was represented in the better performers, whereas the MERCI mechanical thrombectomy trial28 was the eighth worst study. This combination of characteristics suggests that more aggressive efforts at complete recanalization lead to poorer outcomes relative to natural history.
There are similar findings in the recent cardiac literature, where procedures to enhance recanalization are associated with worse clinical outcomes.43–48 Several explanations for this apparent “paradox” have been proposed, but there is a consistent suggestion that the increased time that these procedures require or adverse effects of adjunctive agents may contribute. Although, with some exceptions, most IA series for acute stroke demonstrate that those patients who recanalize did better than those who did not, this finding does not in itself mean that aggressive recanalization should be the goal of IA. There exists the possibility that those patients that do not readily recanalize and require more aggressive efforts may actually have been harmed by the efforts,3,49 a possibility that only comparison with a medical control group can address.50 The benefit of imaging criteria to identify those patients who may benefit from intervention at longer time intervals has been suggested retrospectively,51 and prospective studies are under way.52 However, it is also possible that these imaging characteristics may predict benefit from late IV approaches equally well to IA approaches. Also, ongoing technical improvements may improve IA outcomes. However, we did not see a relationship between outcomes and date of publication, but this may not take into account when the patients were actually treated.
ACKNOWLEDGMENT
Lynn Burke, medical librarian at the University of Texas Medical Branch, assisted in the literature search. Drs. Gilbert Hillman and Roderic Fabian provided input into successive drafts. Earlier versions of the database were maintained by William Dalmeida, Grant Niccum, and Sundeep Mandava.
Footnotes
-
Supplemental data at www.neurology.org
Disclosure: The authors report no conflicts of interest.
Received June 22, 2006. Accepted in final form February 2, 2007.
REFERENCES
Letters: Rapid online correspondence
REQUIREMENTS
If you are uploading a letter concerning an article:
You must have updated your disclosures within six months: http://submit.neurology.org
Your co-authors must send a completed Publishing Agreement Form to Neurology Staff (not necessary for the lead/corresponding author as the form below will suffice) before you upload your comment.
If you are responding to a comment that was written about an article you originally authored:
You (and co-authors) do not need to fill out forms or check disclosures as author forms are still valid
and apply to letter.
Submission specifications:
- Submissions must be < 200 words with < 5 references. Reference 1 must be the article on which you are commenting.
- Submissions should not have more than 5 authors. (Exception: original author replies can include all original authors of the article)
- Submit only on articles published within 6 months of issue date.
- Do not be redundant. Read any comments already posted on the article prior to submission.
- Submitted comments are subject to editing and editor review prior to posting.
You May Also be Interested in
Hemiplegic Migraine Associated With PRRT2 Variations A Clinical and Genetic Study
Dr. Robert Shapiro and Dr. Amynah Pradhan