Use of high-sensitivity digital ELISA improves the diagnostic performance of circulating brain-specific proteins for detection of traumatic brain injury during triage.

Abstract

Background:

Historically, limited sensitivity associated with traditional immunoassay methods have prevented the use of brain-specific proteins as blood biomarkers of traumatic brain injury (TBI) during triage, as these proteins exhibit low concentrations in circulation. Digital ELISA is a newly-developed technique that is up to 1000 times more sensitive than conventional ELISA methods. The purpose of this study was to determine whether the use of digital ELISA over conventional ELISA methods improves the performance of brain-specific proteins as blood biomarkers of TBI during triage.

Methods:

Blood was sampled from TBI patients (n=13) at emergency department admission, as well as from neurologically normal controls (n=72). Serum levels of two brain-specific proteins, neurofilament light chain (NfL) and Tau, were measured via digital ELISA. Estimated conventional ELISA measures were generated by adjusting values according to the lower limits of detection achievable with commercially available conventional ELISA assays, and receiver characteristic analysis (ROC) was used to compare the diagnostic performance of digital ELISA measures to estimated conventional ELISA measures in terms of their ability to discriminate between TBI patients and controls.

Results:

Used in combination, Digital ELISA measures of NfL and Tau could discriminate between groups with 100% sensitivity and 91.7% specificity. Estimated conventional ELISA measures could only discriminate between groups with 7.7% sensitivity and 94.4% specificity. This difference in diagnostic performance was statistically significant according to a comparison of areas under ROC curves.

Conclusions:

The use of digital ELISA over conventional ELISA methods improves the diagnostic performance of circulating brain-specific proteins for detection of TBI during triage.

Introduction:

Over 2.5 million individuals sustain a traumatic brain injury annually in the United States, and close to 100,000 suffer permanent disability as a result [1]. Quick recognition of TBI during triage avoids clinical delay and ensures patients are referred to an appropriate care team which is trained to manage neurological injury; timely and appropriate referral can limit secondary complications and ultimately improve outcome. Unfortunately, the tools available to clinicians for recognition of TBI during early triage are limited. Symptom-based screening assessments currently available to clinicians in the prehospital and early in-hospital setting have been reported to be as low at 30% sensitive, especially in patients with closed head injuries [2,3]. Furthermore, even when these symptom-based tools are combined with CT imaging, mild to moderate TBI cases can be difficult to diagnose [4]. The identification of blood biomarkers associated with TBI could lead to the development of point-of-care molecular diagnostics which could be used to aid in the recognition of TBI during triage.

The proteomic composition of the brain is highly unique relative to other tissues; damage to neural tissue results in the release of brain-specific proteins into the extracellular environment and ultimately into peripheral circulation. Thus, the detection of such proteins in the blood could serve as a surrogate marker of TBI. However, the blood brain barrier limits the diffusion of brain-specific proteins into the blood [5], and they are often only present in circulation at low-picogram to sub-picogram concentrations; unfortunately, the lower limit of detection (LLOD) cutoffs associated with conventional ELISA techniques traditionally used to assay blood are not robust enough to detect them at early enough time-points in TBI progression to make them clinically informative during triage.

However, recent advances in proteomic techniques with greater sensitivity may allow for detection of these markers earlier in pathology. Digital ELISA is a recently developed immunoassay methodology which allows for femtogram-level detection of protein analytes in biofluids. First, antibody conjugated paramagnetic beads are used to capture single molecules of target protein, and protein-bead complexes are labeled with fluorophore conjugated detection antibody. Beads are then assessed for the presence or absence of target protein using a precision-fabricated microwell array capable of capturing one bead per well. This technique has been shown to be as much as 1,000 times more sensitive than traditional ELISA methods [6]. Thus, the extended detection limits associated with digital ELISA could allow for detection of brain specific proteins in the blood at early enough time points to aid in TBI recognition in the acute phase of care.

A handful of recent studies have used digital ELISA to measure brain specific proteins such as glial fibrillary acidic protein (GFAP), neurofilament light chain (NfL), tau, and ubiquitin c-terminal hydrolase L1 (UCHL1) in the blood of TBI patients during the acute phase of care with promising results [7–10], however whether there is diagnostic benefit to using digital ELISA over conventional ELISA methods in measurement of these proteins for TBI recognition has yet to be explicitly investigated. Thus, the purpose of this study was to determine whether the use of digital ELISA over conventional ELISA methods improves the performance of brain-specific proteins as blood biomarkers of TBI during triage.

Methods:

Experimental Design:

Peripheral blood samples were collected from a group of TBI patients at hospital admission (n=13), as well as from a group of neurologically normal control subjects with various other medical conditions (n=72). Serum concentrations of two brain-specific axonal proteins, NfL and tau, were assayed with high sensitivity digital ELISA. In order to compare the diagnostic performance of digital ELISA to conventional ELISA, the mean LLOD associated with commercially available conventional ELISA assays for each biomarker were determined via a systematic search, and estimated conventional ELISA measures were generated by replacing all digital ELISA values below the mean LLOD with zero-values. ROC analysis was used to compare the diagnostic performance of the original digital ELISA values to estimated conventional ELISA values in terms of their ability to discriminate between TBI patients and controls.

Subjects:

TBI patients were recruited in the emergency department at Ruby Memorial Hospital (Morgantown, WV). TBI patients received a clinical diagnosis of TBI based on clinical evaluation and a review of lab and radiographic findings by an experienced neurologist. Severity of TBI was determined by Glasgow Coma Score (GCS), as assessed by a trained clinician [11]. Patients under 18 years of age, previously hospitalized within 30 days, or presenting more than 24 hours post-injury were excluded. Control subjects were free of neurological symptoms, and included patients with a variety of chronic medical conditions including cardiovascular disease and HIV infection who were recruited as participants in various other clinical investigations at Case Western Reserve University (Cleveland, OH) and University Hospitals (Cleveland, OH). All procedures were approved by the institutional review boards of University Hospitals (IRB protocol# 20181112) and West Virginia University (IRB protocol# 1410450461R001). Written informed consent was obtained from all subjects or their authorized representatives prior to any study procedures.

Blood collection and processing:

Blood for serum specimens was drawn via venipuncture and collected in gel barrier serum separator vacutainers (Becton Dickenson, Franklin Lakes, NJ). Blood was allowed to clot for a minimum of 30 minutes, and was then spun at 2,500*g for 10 minutes to separate serum. Resultant serum was aliquoted and stored at −80°C until analysis.

Digital ELISA:

Absolute concentrations of NfL and Tau in serum specimens were measured via the HD-1 Single Molecule Array (SiMoA) platform (Quanterix Corporation, Billerica, MA) as described previously [12]. All measurements were performed by a technician blinded to clinical diagnosis. Total NfL and total Tau were measured via preconfigured assays (Catalog# 103186 and 101995) using manufacture recommended protocols. Standard curves were generated via triplicate measurements of recombinant proteins, and all specimens were assayed in duplicate.

Comparison to conventional ELISA:

Similar to the approach previously used by Akama et al. [13], a systematic search was used to aggregate the manufacture-stated LLOD values associated with commercial conventional ELISA kits available for each of the two brain-specific proteins. The web search was performed via Google search engine using the search strings provided in Supplemental Table 1. Assay specifications were recorded manually, and 20% of data were verified via a second extractor; the average interrater reliability was 98.7% across all recorded variables.

For each subject, estimated conventional ELISA measures for both brain-specific proteins were generated. To calculate these estimated measures, the concentrations of specimens with digital ELISA values below the mean manufacture-stated LLOD associated with available conventional ELISA assays were replaced with zero values. ROC analysis was used to compare the sensitivity and specificity of these estimated conventional ELISA measures to those of the original digital ELISA measures.

Statistics:

Statistics were performed via R version 3.4 [14]. Fisher’s exact test was used for comparison of dichotomous variables, while T-test was used for comparison of continuous variables. Strength of correlational relationships were assessed via Spearman’s rho. The diagnostic ability of binary classifiers was assessed via receiver operator characteristic (ROC) using the ‘pROC’ package. In the case of classification using a single variable, ROC curves were generated directly using the roc() function [15]. In the case of classification using multiple variables, classification was performed using k nearest neighbors (kNN) via the knn.cv() function of the ‘class’ package [16], and the resultant prediction probabilities were used for receiver operator characteristic (ROC) curve generation as described previously [17–19]. Statistical comparison of area under ROC curves was performed using roc.test() function according the non-parametric method described by DeLong et al. [20]. The null hypothesis was rejected when p<0.05. The parameters of all statistical tests performed are outlined in detail within the figure legends.

Results:

Clinical and demographic characteristics:

TBI patients and control patients were well matched in terms of age. There was a slightly higher proportion of female subjects in the control group, however this difference was not statistically significant. The TBI group included a mix of isolated and non-isolated head injuries, most of which a result of motor vehicle accidents or falls. Furthermore, the TBI group exhibited an average GCS of 10.8±2.4, and a majority of cases were mild or moderate severity (Table 1).

Table 1.

Clinical and Demographic characteristics.

	Control (n=72)	TBI (n=13)	p-value:
aAge mean±SD:	50.4±12.8	52.8±26.0	0.743
bMale n(%):	22 (30.6)	6 (46.2)	0.339
Minutes from onset mean±SD:	-	659±413	-
Injury severity:
cMild n(%):	-	5 (38.4)	-
dModerate n(%):	-	6 (46.2)	-
eSevere n(%):	-	2 (15.4)	-
Cause of injury:
Motor vehicle accident n(%):	-	4 (30.8)	-
Sports injury n(%):	-	1 (7.7)	-
Assault n(%):	-	2 (15.4)	-
Fall n(%):	-	5 (38.4)	-
Other accident n(%):	-	1 (7.7)	-

^aMeans statistically compared using two-sample two-tailed t-test

^bProportions statistically compared using Fisher’s exact test

^cGlasgow coma scale of 13–15

^dGlasgow coma scale of 9–12

^dGlasgow coma scale of ≤8

^*statistically significant.

Circulating levels of brain specific proteins measured with digital ELISA:

Serum concentrations of NfL and tau measured with digital ELISA were above the manufacture-stated lower limit of detection for all but one sample. Concentrations of both NfL and tau were significantly higher in serum sampled from TBI patients than in serum sampled from controls (Figure 1A, 1B). Furthermore, circulating levels of both proteins were positively associated with injury severity as assessed by GCS, however neither association was statistically significant (Figure 1C, 1D).

Figure 1. Circulating concentrations of brain-specific proteins in TBI patients and controls measured by digital ELISA.

(A-B) Serum concentrations of NfL and Tau in TBI patients and controls measured by digital ELISA. Means were compared via two-sample two-tailed t-test. Tests were performed using log2 values. (C-D) Correlations between GCS and serum concentrations of NfL and Tau measured by digital ELISA in the TBI group. Strength of correlations were tested using Spearman’s rho. *statistically significant.

Diagnostic performance of brain specific proteins measured with digital ELISA:

Serum levels of NfL measured with digital ELISA were able to discriminate between TBI patients and controls with 92.3% (0.95 CI: 64.0–99.8%) sensitivity and 88.9% (0.95 CI: 79.3–95.1%) specificity (Figure 2A), while serum levels of Tau were able to discriminate between TBI patients and controls with 84.6% (0.95 CI: 54.6–98.1%) sensitivity and 90.3% (0.95 CI: 81.0–96.0%) specificity (Figure 2C). Used in combination, the coordinate serum levels of both markers were able to collectively discriminate between TBI patients and controls with 100% (0.95 CI: 75.3–100%) sensitivity and 91.7% (0.95 CI: 82.7–96.9%) specificity (Figure 4C). While the combination of both markers exhibited slightly higher diagnostic accuracy relative to either marker used in isolation, this difference was not significantly different upon comparison of areas under ROC curves (Figure 2D).

Figure 2. Diagnostic performance of digital ELISA measures of brain specific-proteins.

(A-C) ROC curves depicting the ability of digital ELISA measures of serum NfL, Tau, and their combination to discriminate between TBI patients and controls. To assess combined diagnostic performance, ROC curves were generated via k-nn prediction probabilities. The diagnostic cut-off with the highest combined sensitivity and specificity is indicated. (D) Comparison of the areas under ROC curves for digital ELISA measures of serum NfL, Tau, and their combination. Areas were statistically compared using the DeLong method. 95% confidence intervals associated with all diagnostic statistics are indicated. AUC, area under curve; Pr, probability; *statistically significant.

Figure 4. Diagnostic comparison of digital ELISA measures and estimated conventional ELISA measures.

Comparison of the areas under ROC curves between digital ELISA measures and estimated conventional ELISA measures of serum NfL, Tau, and their combination, with respect to their ability to discriminate between TBI patients and controls. Areas were statistically compared using the DeLong method. 95% confidence intervals associated with all diagnostic statistics are indicated. AUC, area under curve.

Comparison of digital ELISA and conventional ELISA:

Our systematic search identified 44 commercially available conventional ELISA assays for NfL, and 33 for tau, which met the inclusion criteria for analysis. The mean LLOD for conventional ELISA assays targeting NfL was 36.4 pg/mL, while the mean LLOD for conventional ELISA kits targeting Tau was 41.9 pg/mL (Table 2). Based on our measures with digital ELISA, only 6 out of 85 total samples (7.1%) exhibited serum NfL concentrations above the mean LLOD for commercially available conventional ELISA assays. Furthermore, only 1 out of 85 total samples (1.2%) exhibited serum tau concentrations above the mean LLOD for commercially available conventional ELISA assays (Table 3).

Table 2.

Characteristics of commercially-available conventional ELISA assays.

	NfL (n=44):	Tau (n=33):
Immunoassay format:
Sandwich n (%):	34 (77.3)	32 (97.0)
Competitive n (%):	8 (18.2)	0 (0.0)
Unknown n (%):	2 (4.5)	1 (3.0)
Mode of detection:
Colormetric n (%):	35 (79.5)	25 (75.8)
Chemiluminesent n (%):	5 (11.4)	3 (9.1)
Flourescent n (%):	0 (0.0)	1 (3.0)
Unknown n (%):	4 (9.1)	4 (12.1)
LLOD mean (range):	36.4 (12.4–156.0) pg/mL	41.9 (6.3–500.0) pg/mL

LLOD, lower limit of detection.

Table 3.

Number of samples above LLOD for digital ELISA versus conventional ELISA.

	Digital ELISA (actual):	Conventional ELISA (estimated):	ap-value:
NfL:
Total samples above LLOD proportion (%):	85/85 (100)	6/85 (7.1)	<0.001*
Control samples above LLOD proportion (%):	72/72 (100)	2/72 (2.8)	<0.001*
TBI samples above LLOD proportion (%):	13/13 (100)	4/13 (30.8)	<0.001*
Tau:
Total samples above LLOD proportion (%):	84/85 (98.8)	1/85 (1.2)	<0.001*
Control samples above LLOD proportion (%):	71/72 (98.6)	0/72 (0.0)	<0.001*
TBI samples above LLOD proportion (%):	13/13 (100)	1 (7.7)	<0.001*

^aProportions statistically compared using Fisher’s exact test; LLOD, lower limit of detection

^*statistically significant.

In terms of diagnostic performance, estimated conventional ELISA measures for NfL generated based on the mean LLOD for commercially available conventional ELISA assays could only discriminate between TBI patients and controls with 30.7% (0.95 CI: 9.1–61.4%) sensitivity and 100% (0.95 CI: 95.0–100%) specificity (Figure 3A). Estimated conventional ELISA measures for Tau could only discriminate between TBI patients and controls with 7.7% (0.95 CI: 0.2–36.0%) sensitivity and 100% (0.95 CI: 95.0–100%) specificity (Figure 3B). Used in combination, estimated conventional ELISA measures for both markers were able to collectively discriminate between TBI patients and controls with 7.7% (0.95 CI: 0.2–36.0%) sensitivity and 94.4% (0.95 CI: 86.4–98.5%) specificity (Figure 4C). Based on statistical comparison of areas under ROC curves, estimated conventional ELISA measures exhibited significantly lower levels of diagnostic accuracy in comparison to the original digital ELISA measures in the case of all markers (Figure 4A-4C), suggesting that the extended LLOD afforded by digital ELISA allows for improved diagnostic performance relative to conventional ELISA techniques.

Figure 3. Diagnostic performance of estimated conventional ELISA measures of brain specific-proteins.

(A-C) ROC curves depicting the ability of estimated conventional ELISA measures of serum NfL, Tau, and their combination to discriminate between TBI patients and controls. To assess combined diagnostic performance, ROC curves were generated via k-nn prediction probabilities. The diagnostic cut-off with the highest combined sensitivity and specificity is indicated. (D) Comparison of the areas under ROC curves for estimated conventional ELISA measures of serum NfL, Tau, and their combination. Areas were statistically compared using the DeLong method. 95% confidence intervals associated with all diagnostic statistics are indicated. AUC, area under curve; Pr, probability; *statistically significant.

Discussion:

Historically, limited lower detection ranges associated with traditional immunoassay techniques have hindered the use of brain-specific proteins as blood biomarkers of TBI in the acute phase of care, as these proteins are often only present in circulation at low concentrations. In this study, we aimed to determine whether the use of digital ELISA over conventional ELISA techniques improves the performance of brain-specific proteins as blood biomarkers of TBI during triage. In our analysis, digital ELISA measures of both NfL and tau exhibited dramatically higher levels of diagnostic accuracy relative to estimated conventional ELISA measures, and digital ELISA measures of NfL in particular showed high enough levels of discriminatory performance to suggest future clinical utility.

Numerious prior studies have evaluated the diagnostic potential of circulating brain-specific proteins including S100 calcium-binding protein B (S100B), neuron specific enolase (NSE), glial fibrillary acidic protein (GFAP), ubiquitin C-terminal hydrolase L1 (UCHL1), tau, and NfL for TBI recognition [21]. However, a significant majority of these prior studies have used conventional ELISA techniques for biomarker quantification, and in many cases have reported large numbers of samples under LLOD, effectively limiting diagnostic potential. To our knowledge, this investigation is the first to compare the diagnostic performance of digital ELISA to conventional ELISA in measure circulating levels of brain-specific proteins for TBI diagnosis during triage. Serum concentrations of NfL and tau were above the LLOD for digital ELISA in nearly all samples, however, our analysis suggested that only a small percentage of samples contained analyte levels which would be detectable by conventional ELISA techniques. The diagnostic impact of this difference in sensitivity between the two platforms was apparent, as digital ELISA measures exhibited dramatically higher levels of accuracy for TBI diagnosis relative to estimated conventional ELISA measures, suggesting that the extended lower detection range of digital ELISA offers a distinct diagnostic advantage. Based on these results, future investigations exploring the diagnostic characteristics of blood-borne brain specific proteins in the acute phase of TBI should utilize digital ELISA methods where possible. Furthermore, these findings also suggest that future diagnostic meta-analyses investigating brain-specific proteins in TBI should be careful to account for differences between studies in the use of digital versus conventional ELISA techniques.

Albeit in a limited sample size, our results suggest that the extended LLOD associated with digital ELISA could allow for levels of diagnostic performance which could have clinical utility for TBI recognition during triage, especially in the case of NfL. Digital ELISA measures of NfL were able to discriminate between TBI patients and controls with over 90% sensitivity; these results are consistent with three recently-published emergency department TBI studies which measured NfL in the peripheral blood with digital ELISA and reported similar levels of diagnostic performance [7,9,10]. The symptom-based assessments currently used by emergency medical technicians, paramedics, triage nurses, and emergency department physicians to detect TBI have been shown to be as low at 30% sensitive, especially in the case of mild closed head injuries. Thus, digital ELISA NfL measures could add diagnostic value to the current tools used to triage TBI clinically. However, it is important to note that digital ELISA assays currently require extensive sample preparation and large equipment, thus limiting their immediate use for the type of stat blood testing required during TBI triage. However this limitation will likely be overcome, as promising new rapid point-of-care digital ELISA technologies are in development which could enable use in emergency medicine settings in the near future [22].

While our results are exciting, it must be stated that this study is not without limitations. Perhaps most notably is the fact that we did not directly measure biomarker levels with conventional ELISA in our comparison with digital ELISA, and instead generated estimated measures based on average manufacture-stated LLOD values. However, because there is such a large number of conventional ELISA assays available for each marker, it would be imprudent to assume that if we directly tested one or even a small handful of these assays, that their performance characteristics would generalize in terms of the full complement of assays on the market. Thus, it could be argued that while this approach used estimated measures, it provides a more comprehensive and generalizable results. It is also important to note that we only measured two commonly investigated brain-specific proteins, and there are numerous others being considered as potential biomarkers of TBI. While we would expect the use of digital ELISA to measure these other markers to be similarly diagnostically advantageous, the extent of such advantages would have to be directly investigated. Lastly, the TBI sample size employed in this investigation was relatively small. While this sample size was adequate to support our primary conclusion that digital ELISA provides improved diagnostic performance over conventional ELISA, in isolation, it is too small to draw definitive conclusions regarding the true clinical ability of these biomarkers to diagnose TBI. However, our results do provide important independent evidence which supports the conclusions made in other recent investigations exploring the diagnostic performance of the same markers [7,9,10].

Collectively, our results provide compelling evidence that the use of digital ELISA over conventional ELISA methods significantly improves the diagnostic performance of circulating brain-specific proteins for detection of TBI during triage. Therefore, we recommend that future investigations exploring the diagnostic characteristics of blood-borne brain specific proteins in the acute phase of TBI utilize digital ELISA methods where possible. Furthermore, our findings also suggest that future diagnostic meta-analyses investigating these markers in TBI should be careful to account for differences between studies in the use of digital versus conventional ELISA techniques.