Abstract
Background: The diagnosis and management of acute ischemic stroke are limited by the lack of rapid diagnostic assays for use in an emergency setting. Computed tomography (CT) scanning is used to diagnose hemorrhagic stroke but is relatively ineffective (<33% sensitive) in detecting ischemic stroke. The ability to correlate blood-borne protein biomarkers with stroke phenotypes would aid in the development of such rapid tests.
Methods: ELISAs for >50 protein biomarkers were developed for use on a high-throughput robotic workstation. These assays were used to screen plasma samples from 214 healthy donors and 223 patients diagnosed with stroke, including 82 patients diagnosed with acute ischemic stroke. Marker assay values were first compared by univariate analysis, and then the top markers were subjected to multivariate analysis to derive a marker panel algorithm for the prediction of stroke.
Results: The top markers from this analysis were S-100b (a marker of astrocytic activation), B-type neurotrophic growth factor, von Willebrand factor, matrix metalloproteinase-9, and monocyte chemotactic protein-1. In a panel algorithm in which three or more marker values above their respective cutoffs were scored as positive, these five markers provided a sensitivity of 92% at 93% specificity for ischemic stroke samples taken within 6 h from symptom onset.
Conclusion: A marker panel approach to the diagnosis of stroke may provide a useful adjunct to CT scanning in the emergency setting.
The rapid diagnosis of stroke remains a major challenge for patient management and therapeutic intervention. Thrombolytic agents have been shown to improve outcome for stroke patients with evidence of salvageable tissue (1). However, the presence of intracranial hemorrhage (ICH), 1 in which case these agents are contraindicated, must first be ruled out. At most hospitals where the appropriate equipment is available, the diagnosis of ischemic stroke is made solely on clinical grounds after ruling out hemorrhagic lesions by computed tomography (CT), in some cases supplemented by diffusion- and perfusion-weighted magnetic resonance imaging (MRI). Once such a diagnosis is made, patients benefit from thrombolytic therapy administered within 6 h of stroke onset (2)(3). These observations suggest that the availability of a rapid diagnostic assay for acute stroke would be highly beneficial in assessing the care pathway for patients and their treatment options.
Ischemic stroke is a pathologic condition with acute onset that is caused by the infarction of a vessel supplying blood and nutrients to the brain. The immediate area of injury is called the core and contains brain cells that have died as a result of ischemia or physical damage. The penumbra is composed of tenuously perfused but still viable tissue adjacent to cells in the core. Cells within the penumbra have the potential to recover after reperfusion therapy. The ability to detect such an event by CT is related in part to the initial volume and rate of core expansion from the point of infarction. However, the volume of infarction would not necessarily be expected to correlate with clinical outcome because the location of the infarction is another contributing factor to persistent neurologic deficit. Therefore, small infarcts that are difficult to detect by CT could still lead to neurologic deficit, suggesting that alternative methods of diagnosis are needed.
During stroke, the blood–brain barrier (BBB) is compromised by endothelial cell death, and cytosolic contents released from injured brain tissues have the potential to cross the BBB. This suggests that the measurement of brain-derived proteins in plasma could be used to monitor stroke onset and severity. For example, rodent models have demonstrated the breakdown of the BBB after focal stroke (4). MRI scanning with gadolinium infusion has provided some evidence of the breakdown of the BBB in humans after cerebral infarction (5), but a time course for this process remains to be demonstrated.
Plasma concentrations of the astroglial protein S-100b, a cytosolic calcium-binding protein, have been found to correlate with the extent of tissue damage (infarct volume) and neurologic outcome (6)(7). The variations in S-100b concentrations were not significantly different within the first 6 h from stroke onset, indicating that this biomarker alone may not be sufficient for diagnosing acute stroke. In addition to markers of cellular activation after brain injury, secreted neurotrophic factors such as nerve growth factor and B-type neurotrophic growth factor (BNGF) have been studied in ischemic brain tissue as neuroprotective agents (8)(9), although studies correlating their plasma concentrations to stroke phenotypes have not been published to date.
Protein markers of nonbrain origin have also been investigated for their potential utility in the diagnosis of stroke. For example, serum concentrations of von Willebrand factor (vWF) and matrix metalloproteinase-9 (MMP-9) were shown to correlate with the onset of cerebral vasospasm after subarachnoid hemorrhage (SAH) (10). In addition, plasma concentrations of MMP-9 were found to predict hemorrhagic transformation in patients with acute ischemic stroke (11). Monocyte chemoattractant protein-1 (MCP-1) was found to be increased in the cerebrospinal fluid of stroke patients, although serum concentrations did not differ from those of control patients (12).
Over the past several years, we have conducted a systematic investigation of >50 marker assays, using blood plasma samples taken from stroke patients and healthy donors. In a previous report, we described preliminary results with 26 marker assays and blood samples from 65 patients admitted to the emergency department with suspected stroke, including 44 patients who were diagnosed with acute stroke [persistent residual neurologic deficit at 24 h and meeting the clinical criteria for stroke (Lynch JR, White WD, Blessing R, Grocott HP, Newman MF, and Laskowitz DT, submitted for publication)]. Here we report results from a more extensive study using plasma samples from 214 healthy donors and 221 stroke patients (including 122 patients diagnosed with ischemic stroke). Our new findings demonstrate that five biomarkers, S-100b, BNGF, vWF, MMP-9, and MCP-1, are highly correlated with the diagnosis of stroke. Our results further demonstrate that the five biomarkers could be used collectively as a “marker panel” to diagnose hemorrhagic and ischemic stroke, and particularly acute ischemic stroke within the first 6 h of onset.
Materials and Methods
specimen collection
This study was approved by the Duke University Medical Center Institutional Review Board. All patients or their families provided informed written consent. The primary endpoint in this study was the presence of clinical stroke as documented by board-certified neurologists and defined by focal neurologic signs or symptoms felt to be of vascular origin that persisted for >24 h. Before discharge, CT and/or MRI imaging were used to confirm the diagnosis of ischemic strokes. After informed consent was obtained, blood samples from 274 patients admitted to the emergency department between November 21, 1999, and February 1, 2003, with suspected stroke were drawn into EDTA collection tubes and centrifuged for 10 min at 10 000g; the resulting supernatant was transferred to barcode-labeled cryovials and immediately frozen at −70 °C until shipment to Biosite® Incorporated for analysis. The specimens were shipped overnight on dry ice. On receipt, the specimens were logged in and immediately transferred to −70 °C storage. After several hundred specimens had been received, the cryovials were thawed, and aliquots were transferred to 384-well barcode-labeled microtiter plates by use of a Tecan Genesis workstation with barcode reading capability. Multiple copies of the specimen set were replicated in the 384-well microtiter plates in this fashion. The plates were sealed, frozen, and placed in −70 °C storage until the time of assay.
EDTA-plasma samples were likewise collected from healthy volunteers at Biosite. These samples were age-matched to the stroke samples that were received from Duke University Medical Center (above). In total, samples from 214 healthy donors were used in this study as a control population.
immunoassay reagents
MCP-1.
Two murine antibodies generated by Biosite, using phage display and recombinant protein expression as described previously (13), are used for this assay. The capture antibody is biotinylated, and the recorder antibody is conjugated to alkaline phosphatase. The antigen used to calibrate the assay is a recombinant protein also produced at Biosite.
MMP-9.
For this assay, a murine polyclonal antibody was generated by Biosite, using phage display and recombinant protein expression as described previously (13), is used for this assay. A portion of the antibody is biotinylated for use as the capture antibody, and another portion is conjugated to alkaline phosphatase for use as the recorder antibody. The antigen used to calibrate the assay is MMP-9 monomer, neutrophil granulocyte, purchased from Calbiochem (cat. no. 444231).
S-100b.
Two paired monoclonal antibodies purchased from Fitzgerald Laboratories (cat. no. 10-S17) were used for this assay. Clone M9940622 was biotinylated for use as the capture antibody, and clone M9940621 was conjugated to alkaline phosphatase for use as the recorder antibody. The antigen used to calibrate the assay was purchased from Advanced Immunochemical (cat. no HBS100-bb).
vWF.
Two paired monoclonal antibodies manufactured at Biosite were used for this assay. One was biotinylated for use as the capture antibody, and the other was conjugated to alkaline phosphatase for use as the recorder antibody. The antigen used to calibrate the assay was from Cortex Biochem (cat. no. CD4003U).
BNGF.
Two paired monoclonal antibodies were purchased from R&D Systems. Biotinylated goat-anti-BNGF (cat. no. BAF256) was used as the capture antibody, and alkaline phosphatase-conjugated monoclonal anti-BNGF (cat. no. MAB256) was used as the recorder antibody. The antigen used to calibrate the assay was also purchased from R&D Systems (cat. no. 256-GF).
immunoassay procedure
All immunoassays were forward immunometric (sandwich) assays performed in 384-well microtiter plates using a Tecan Genesis RSP 200/8 Workstation (TECAN). Each plasma sample was assayed in duplicate.
The biotinylated antibody was added to neutravidin-coated 384-well black plates (Pierce Chemical Co.) and incubated at room temperature for 1 h. The plate was washed, and then plasma samples (10 μL) were aliquoted into the wells. After incubation for 1.25 h, the plate was washed again, and the alkaline phosphatase-conjugated antibody was added. After an additional incubation for 1.25 h, the plate was washed a third time, and AtttoPhos® substrate (Promega) was added to measure the amount of alkaline phosphatase-conjugated antibody bound in each well. The plates were read by a fluorometer with an excitation wavelength of 430 nm and an emission wavelength of 570 nm. Each well was read six times at 114-s intervals, and a rate of fluorescence generation was calculated. With the exception of vWF, calibrators were prepared gravimetrically in pooled plasma from healthy donors. One tube in each set of calibrators included neutralizing antibody for correction of endogenous antigen present in the plasma pool. Calibrators for vWF were prepared in an assay buffer containing 10 mmol/L Tris-HCl (pH 8.0), 150 mmol/L NaCl, 1 mmol/L MgCl2, 0.1 mmol/L ZnCl2, 10 mL/L polyvinyl alcohol (Mr 9000–10 000), 10 g/L bovine serum albumin, and 1 g/L NaN3. Calibration curves were eight points run in duplicate in columns 1 and 2 and repeated in columns 23 and 24 of the assay plate. The calibration curve was calculated using a four-parameter logistic fit.
data analysis
Panel algorithm.
Assay values for each biomarker were examined by univariate analysis to calculate optimized cutoffs for the stratification of diseased vs nondiseased individuals. For the purposes of this calculation, diseased individuals were classified as having either an ischemic or a hemorrhagic stroke diagnosis, and assay values for samples taken at all time points from symptom onset were used. Patients identified with hemorrhagic stroke included those who were diagnosed with either an ICH or a SAH. The cutoff values were obtained by iterative permutation–response calculations using all five markers in a panel algorithm wherein three of five markers above their respective cutoffs scored as positive for stroke. The permutation–response calculations were conducted using a macro written in Microsoft Excel. Briefly, an initial cutoff value was chosen arbitrarily for each marker (typically within the third or fourth quartile of the healthy donor values), and the diagnostic sensitivity of the resulting panel response (see above) was calculated at 93% specificity. Each cutoff value was then changed iteratively by 10% increments, and a new sensitivity was determined after each iteration until a maximum sensitivity was achieved. The resulting cutoffs were then used to assess the potential diagnostic utility of the panel algorithm when applied to samples taken within specified time windows from symptom onset. In this case, patients diagnosed with either ischemic or hemorrhagic stroke were examined separately with the algorithm.
Multivariable regression.
Assay values were next subjected to a conventional five-variable logistic regression analysis designed to test the strength of the overall association of the five selected markers with the stroke outcome. Instead of scalar cutoffs, this analysis used each marker assay value in a multiparameter function to optimize predicted likelihood of stroke. As in the previous analysis (above), the multivariate logistic regression analysis defined diseased individuals as patients diagnosed with either an ischemic or hemorrhagic stroke. However, the multivariate analysis used only patient samples taken within 12 h from symptom onset. After checking model fit, the resulting model was then applied to samples at various time windows from symptom onset for inspection of sensitivity and specificity.
Results
Of the 274 patients enrolled in this study, 223 were diagnosed with stroke, including 82 patients with ischemic stroke. The patients diagnosed with stroke comprised ∼60% women and ∼60% African Americans (∼38% white). Approximately 30% of these patients had a previous history of myocardial infarction. The subcategories of clinical diagnosis for these patients are summarized in Table 1⇓ . For the purposes of our analysis, hemorrhagic stroke patients were identified as those diagnosed with either an ICH or a SAH. Patients diagnosed with closed head injuries (CLH) and transient ischemic attacks (TIAs) were excluded from our panel analysis because our first objective was to determine whether a marker panel could diagnose the presence of acute stroke, i.e., persistent neurologic deficit at 24 h from symptom onset.
Summary of clinical diagnoses for patients enrolled.
The first challenge in defining a marker panel for stroke was to identify proteins in plasma that could be correlated to patient phenotypes by ELISA. To address this, we developed fluorescent immunoassays to >50 different proteins, including brain-specific markers (attributable to cellular injury and apoptosis), inflammatory mediators, markers of coagulation and hemostasis, and acute-phase response markers. The initial ranking of marker assay values was determined by univariate analysis of control vs diseased populations, using a subset of patient samples comprising 157 healthy donors and 161 individuals diagnosed with acute stroke, excluding patients diagnosed with TIAs and CLH and patients presenting postresuscitation. A preliminary analysis of 26 of these proteins is described elsewhere, along with detailed descriptions of the various biochemical categories from which these proteins were selected (Lynch JR, White WD, Blessing R, Grocott HP, Newman MF, and Laskowitz DT, submitted for publication). For the present study, five protein markers were selected that showed moderate (P <0.1) or better individual contributions with stroke diagnosis for univariate comparisons of nondiseased vs ischemic + hemorrhagic stroke individuals: S-100b, BNGF, vWF, MMP-9, and MCP-1. As illustrated in Fig. 1⇓ , MMP-9 and vWF showed good univariate discrimination of nondiseased vs diseased individuals (P <0.0001). However, we realized that these two markers alone could be of limited clinical value for stroke diagnosis because they are not uniquely of brain origin and their plasma concentrations could potentially be increased in disease phenotypes other than stroke. We therefore expanded our analysis to include three additional markers for evaluation as a five-marker panel.
Box- and-whisker plots of biomarker concentrations in patient samples taken within 12 h from symptom onset.
TIA, patients with a discharge diagnosis of TIA (symptoms resolved within 24 h from onset); Isch., patients diagnosed with ischemic stroke; Hemor., patients diagnosed with hemorrhagic stroke (excluding CLH); Controls, samples taken from healthy donors.
The assay values for S-100b, BNGF, vWF, MMP-9, and MCP-1 were first examined in a panel algorithm that defined three of the markers above their respective cutoff values as a positive diagnosis for stroke. Optimized cutoffs were determined as described above, using all of the available stroke patient samples and defining diseased individuals as having either an ischemic or a hemorrhagic stroke diagnosis. The resulting algorithm (including optimized cutoffs) was then applied to samples taken within various time windows from symptom onset, and the results are summarized in Table 2⇓ . The algorithm provided optimum sensitivity for prediction of ischemic stroke within 0–9 h from symptom onset (93% sensitivity at 93% specificity; n = 75 samples), with slightly decreasing sensitivity at later sampling times (86.2% sensitivity at 93% specificity for 0–48 h from symptom onset; n = 195 samples). Furthermore, the algorithm provided 80–89% sensitivity at 93% specificity for prediction of hemorrhagic stroke based on samples taken after 6 h from symptom onset (n = 450 total samples).
Prediction of stroke by a five-marker panel algorithm with optimized marker cutoffs.1
To provide a separate demonstration of diagnostic utility for the five-marker panel, we developed a multivariable logistic regression model, again as described above. The continuous values of the five markers showed an extremely good fit and predictive value, with a P value <0.0001 (χ2 = 216.1; df = 5), and an area under the ROC curve of 99%. In this model, 54 samples from stroke patients taken <12 h after symptom onset were compared with samples from 214 healthy controls. The model achieved a goodness-of-fit P value of 0.947 (where P <0.05 indicates poor fit), and showed overall specificity of 96.7% with a sensitivity of 90.7%. Estimated odds ratios for the separate markers are shown in Table 3⇓ . Odds ratios are expressed in terms of 1 SD of the marker, except for MMP-9, where the unit for the odds ratio is the interquartile range. This means that the odds ratios represent the changes in the odds of stroke when the marker values increase by 1 SD (or 1 interquartile range for MMP-9).
Multivariable logistic regression analysis: confidence intervals for adjusted odds ratios.
The sensitivity and specificity of the multivariate logistic regression model as applied to separate subgroups of stroke type and sample time from onset are shown in Table 4⇓ . As described above, the model was developed using stroke samples taken within 0–12 h from symptom onset and then applied to samples taken in various time windows from symptom onset. As expected, the model provides good prediction of stroke diagnosis for samples taken within 0–6 h from symptom onset (e.g., 98.1% sensitivity at 93.1% specificity for ischemic stroke diagnosis). Furthermore, the model provides good prediction of stroke diagnosis for samples taken after 12 h from onset, although these samples were not used for model development.
Multivariable regression analysis: sensitivity and specificity of model prediction using 0–12 h samples applied to all times.1
Discussion
The objective of this study was to assess the potential diagnostic utility of blood-borne protein biomarkers in predicting acute stroke, i.e., stroke within the first 6 h of symptom onset. The availability of a rapid diagnostic test for stroke based on blood biomarkers would be a substantial adjunct to CT, which is the only diagnostic tool commonly available in an emergency department setting. Furthermore, a simple blood test for acute stroke could be of benefit in hospitals where CT is not yet available.
Diagnostic tests based on blood markers are commonly used for cardiogenic disorders, for example, B-type natriuretic peptide for the diagnosis of heart failure and troponin I and troponin I complexes for the diagnosis of myocardial infarction. Although troponin I and the complexes are highly specific for heart tissue, its appearance in the blood is delayed relative to myoglobin, a less specific but highly sensitive marker of infarcted heart tissue. Thus, a marker panel using myoglobin, troponin I and troponin I complexes, and creatine kinase isoenzyme MB has shown increased sensitivity compared with troponin I and complexes alone (14). We decided to use a similar approach to the diagnosis of stroke by assaying proteins of brain specific origin together with other nonspecific markers. From the brain-specific markers that were analyzed, including cytosolic proteins and secreted neurotrophic factors, S-100b and BNGF showed predictive diagnostic value by univariate analysis (P <0.1). However, a preliminary analysis of immunoassay data obtained for these markers indicated that additional markers would be needed for early diagnosis, particularly for acute stroke within the first 6 h from symptom onset. Therefore, marker assays were developed for other proteins that could be associated physiologically with acute stroke. These included inflammatory mediators, markers of coagulation and hemostasis, and acute-phase response markers.
In this study, we demonstrate that a marker panel comprising S-100b, BNGF, vWF, MMP-9, and MCP-1 can be used for the prediction of acute stroke. The five-marker panel provided increased diagnostic sensitivity compared with any of the markers treated individually. A panel algorithm using optimized cutoff values for all five markers, and assigning positive scores to samples showing three or more markers above their respective cutoffs, showed good sensitivity and specificity for predicting acute stroke (Table 2⇑ ). The multivariable logistic regression model confirmed the predictive power of this set of markers in a classic biostatistics setting.
The predictive value of the panel algorithm compared favorably with clinical CT scanning for diagnosis of ischemic stroke; the latter approach was found to be ∼33% sensitive (n = 36 patients; data not shown). Although the results from CT will likely vary among institutions, with some reporting higher sensitivities, it is not widely held to be a sensitive method of diagnosing acute stroke (15). Considering that the panel algorithm showed 91.7% sensitivity at 93% specificity for predicting ischemic stroke using samples taken within 6 h from symptom onset, it is evident that a diagnostic test using these markers could be a useful adjunct to CT scanning.
Although the results of the present study are encouraging, additional studies are needed to establish the validity of this approach. An extension of the present work to include additional samples from multiple centers is ongoing, including patients who presented to the emergency department with a nonstroke discharge diagnosis. Such patient samples are expected to enhance our control population for marker panel analysis.
Acknowledgments
We thank Robert Blessing, Beth Perry, and Donna Phinney for coordinating and processing patient samples at Duke University Medical Center and William D. White at Duke University Medical Center for assistance with statistical analyses. We also thank Andrea Brooks, Kelline Rodems, Cheryl Carina, Brian Janaky, Ziad Saleem, and Fan Dong for assistance with immunoassay development and high-throughput marker assays.
Footnotes
1 ISCH, ischemic stroke.
1 Three or more marker values above their respective cutoffs score positive for stroke.
2 At a specificity of 93%.
1 Units of marker for which odds ratio is expressed. Units are equal to 1 SD of marker except for MMP-9, for which units equal interquartile range.
1 The sample numbers in this table are different from those in Table 2⇑ because all marker values must be present for a sample to be included in this multivariable regression analysis.
↵1 Nonstandard abbreviations: ICH, intracranial hemorrhage; CT, computed tomography; MRI, magnetic resonance imaging; BBB, blood–brain barrier; BNGF, B-type neurotrophic growth factor; vWF, von Willebrand factor; MMP-9, matrix metalloproteinase-9; SAH, subarachnoid hemorrhage; MCP-1, monocyte chemoattractant protein-1; CLH, closed head injury; and TIA, transient ischemic attack.
- © 2003 The American Association for Clinical Chemistry
