Plasma neurofilament light predicts mortality in patients with stroke

Abstract

Given the substantial heterogeneity of stroke brain injury, there is a clear need for a biomarker that determines the degree of neuroaxonal injury across diverse stroke types. We evaluated whether blood neurofilament light (NFL), a major component of neuronal axons, would fulfill this purpose for patients with acute cerebral infarction (ACI, N=227), aneurysmal subarachnoid hemorrhage (aSAH, N=58) or non-traumatic intracerebral hemorrhage (ICH, N=29). We additionally validated our findings in two independent cohorts of patients with ICH (N=96 and N=54) given the scarcity of blood biomarker studies for this deadliest stroke type. Compared to healthy individuals (N=79 and 48 for the discovery and validation cohorts, respectively), NFL was higher for all stroke types. NFL associated with radiographic markers of brain tissue damage. It correlated with the extent of early ischemic injury in patients with ACI, hemorrhage severity in patients with aSAH, and intracranial hemorrhage volume in patients with ICH. In all patients, NFL independently correlated with scores from the NIH Stroke Scale, the modified Rankin Scale, and the Mini-Mental State Examination at blood draw, which respectively assess neurological, functional, and cognitive status. Furthermore, higher NFL concentrations independently associated with 3- or 6-month functional disability and higher all-cause mortality. These data support NFL as a uniform method to estimate neuroaxonal injury and forecast mortality regardless of stroke mechanism. As a prognostic biomarker, blood NFL has the potential to assist with planning supportive and rehabilitation services, and improving the efficiency of acute therapeutic drug and device trials, and of recovery trials.

One Sentence Summary:

Plasma neurofilament light is a reliable, uniform marker to estimate neuroaxonal injury and forecast all-cause mortality across major stroke types.

Introduction

Stroke is a leading cause of death worldwide, but outcomes after a stroke range widely. Some individuals experience temporary, non-disabling symptoms, whereas others develop varying degrees of neurological, physical and/or cognitive impairment. Clinical scales provide insight into stroke severity, but they are often crude non-continuous measures of the extent of tissue injury. The National Institutes of Health Stroke Scale (NIHSS) correlates only modestly with volume of infarcted tissue (1), and this correlation is modulated by pre-stroke factors like the degree of leukoaraiosis (2). The intracerebral hemorrhage (ICH) score (3) and the World Federation of Neurosurgical Societies score (4) are useful outcome predictors of ICH and subarachnoid hemorrhage (SAH), respectively, but they are not well validated as neuroaxonal injury markers. There is thus a clear need for biomarkers that provide a continuous measure of the degree of neuroaxonal injury regardless of whether the stroke is ischemic or hemorrhagic. Such biomarkers have the potential to improve stroke trial efficiency and our ability to predict clinically meaningful outcomes like post-stroke cognitive impairment and response to rehabilitation.

Neurofilament proteins, major cytoskeletal components of axons, have emerged as promising candidate biomarkers of neuronal injury. They are abundantly and exclusively expressed in neurons, are released into extracellular fluids during axonal injury, and show prognostic utility for several neurological disorders (5, 6). Initial studies investigating neurofilament proteins as biomarkers for stroke focused largely on SAH, and measured neurofilament light (NFL) (7–9) or phosphorylated neurofilament heavy (pNFH) (10–13) in cerebrospinal fluid (CSF). These studies found elevated CSF neurofilament in patients with SAH compared to control individuals (7, 9–12), with higher concentrations generally predictive of worse functional outcomes (9, 12, 13). Nevertheless, to be most useful to patients, a stroke biomarker should be measurable in blood. With the development of neurofilament immunoassays with improved sensitivity, this is now feasible. Consequently, determining the prognostic utility of blood neurofilament proteins has become an important endeavor in the stroke field. To date, studies aimed at addressing this question have chiefly concentrated on acute cerebral infarction (ACI) (14–22), with far fewer investigations examining blood neurofilament in SAH (13, 23) or ICH (24, 25). In aggregate, findings from these studies are encouraging but nonetheless highlight the need for more intensive investigations on hemorrhagic stroke and on the interstudy discrepancies that must be addressed if blood neurofilament proteins are to be successfully translated to the clinic. For instance, associations of blood neurofilament with functional outcomes and, in the case of ACI, with infarct volumes, were observed in only some of the studies in which such correlations were examined (13–17, 19–26). Inconsistent findings among reports may be explained, at least in part, by sample size, the fact that neurofilament proteins in CSF (9, 13, 16) and blood (13, 15–18, 20, 21, 26) progressively rise in the days following the stroke, and the stringency of the statistical analyses. In addition to discrepant findings, the relationship between blood neurofilament proteins and post-stroke survival has yet to be rigorously evaluated.

Given the critical need for efficient and unbiased biomarkers for stroke, we undertook a comprehensive study to investigate the hypothesis that blood NFL could serve as a uniform means to predict the degree of neuroaxonal injury and all-cause mortality across all major types of ischemic and hemorrhagic stroke. Furthermore, since data on blood NFL in ICH are limited (25), we used two independent validation cohorts to confirm findings from our discovery ICH sample series, and evaluated the temporal profile of blood NFL following ICH.

Results

Post-stroke trends of plasma NFL over time

To test our hypothesis, we first measured NFL in plasma from 393 participants enrolled in the Mayo Clinic Florida Familial Cerebrovascular Diseases Registry between the dates of October 2010 to December 2018 (27). This series comprised plasma from 79 healthy individuals and 314 patients with stroke [227 with ACI, 58 with aneurysmal SAH (aSAH), and 29 with ICH] for whom plasma was collected within 20 days of stroke. Baseline characteristics, comorbidities, stroke information, and post-stroke outcomes are summarized in Table S1 for healthy controls and patients with stroke (overall and according to stroke type). Table S2 provides these same summaries for patients with ACI stratified by subtype using the TOAST classification system: large-artery atherosclerosis (LAA); small-vessel occlusion (SVO); cardioembolism (CE); stroke of undetermined etiology (UDE); and stroke of other determined etiology (ODE) (28).

Although subject-specific longitudinal data on NFL were not available in this cross-sectional series, we visualized post-stroke trends of NFL over time. As shown in Fig. 1A, we observed an increase in plasma NFL in the days following the stroke, which peaked approximately 9 days post-stroke. Indeed, plasma NFL in blood collected 9–20 days post-stroke was significantly higher than in blood collected 0–8 days post-stroke when adjusting for stroke group only (P<0.001) and when additionally adjusting for age at blood draw, sex, body mass index (BMI), current smoking, hypertension, diabetes and physical activity (P<0.001). Subsequent association analyses were thus stratified by time from stroke to blood draw (0–8 or 9–20 days) when sample size permitted and also adjusted for time from stroke to blood draw as a continuous variable in all analyses involving only individuals with stroke. When sample size allowed for such a multivariable analysis, we also adjusted for other potential confounding variables: age at blood draw, sex, BMI, current smoking, hypertension, diabetes and physical activity (29).

Fig. 1. Plasma NFL is elevated in patients with ischemic and hemorrhagic stroke from the discovery study.

(A) Plasma NFL trends over time in all patients with stroke combined (N=314).The mean plasma NFL concentration for controls (N=79) is shown by a solid horizontal line; upper and lower standard deviations are shown by horizontal dotted lines. (B) Comparison of plasma NFL between healthy controls (N=79) and patients with acute cerebral infarction (ACI, N=227), aneurysmal subarachnoid hemorrhage (aSAH, N=58) or intracerebral hemorrhage (ICH, N=29). Shown are NFL concentrations in all samples for a given stroke group, or NFL concentrations based on timing from stroke to blood draw (0–8 or 9–20 days). The black horizontal bars represent median NFL concentrations. ***P < 0.001 in multivariable analysis comparing plasma NFL between the indicated group and healthy controls after adjusting for age at blood draw, sex, BMI, current smoking, hypertension, diabetes and physical activity. ^###P < 0.001 in analysis comparing plasma NFL between the indicated group and patients with aSAH after adjusting for time from stroke to blood draw. See also Tables S3 and S5. NFL concentrations are shown on the base 10 logarithm scale.

Plasma NFL is elevated in patients with ACI, aSAH and ICH

In healthy individuals, plasma NFL concentrations strongly correlated with age (Spearman’s r=0.81, P<0.001, Fig. S1A). Compared to healthy controls, plasma NFL was significantly elevated in all patients with stroke combined and in patients with ACI, aSAH, or ICH, both in unadjusted analysis and when adjusting for the potential confounding variables listed above (all P < 0.001; Fig. 1B and table S3). Similarly, plasma NFL was significantly elevated in all studied ACI subtypes in comparison to controls (LAA, SVO, CE, UDE; all P<0.001, Fig. S2, Table S4). Differences in NFL between controls and patients with stroke were most evident for patients with ICH and for patients with a time from stroke to blood draw between 9 and 20 days (Fig. 1B and Table S3).

When comparing plasma NFL among stroke groups, we found that NFL was significantly higher in patients with ACI or ICH compared to patients with aSAH in analyses adjusting for length of time from blood draw (both P < 0.001; Fig. 1B and table S5). Results were similar in analysis stratified by length of time from stroke to blood draw (0–8 or 9–20 days), and when adjusting for potential confounding variables when possible (Table S5). No differences in plasma NFL were identified among ACI subtypes after correcting for multiple testing (fig. S2 and table S6).

Plasma NFL correlates with radiographic markers

To determine the relationship between plasma NFL and brain tissue damage, we evaluated associations between NFL and (i) infarct volume in patients with ACI determined using the Alberta Stroke Program Early Computerized Tomography (CT) Score (ASPECTS), (ii) severity of SAH in patients with aSAH determined using the modified Fisher grading scale, and (iii) the volume of intracranial hemorrhage in patients with ICH determined using the ABC/2 formula. A summary of ASPECTS, modified Fisher scores, and ABC/2 scores is provided in Table S7. In all three groups, higher plasma NFL was indicative of worse outcomes (Table 1). Specifically, when adjusting for time from stroke to blood draw, higher NFL associated with lower ASPECTS in patients with ACI (P<0.001), and this remained significant when additionally adjusting for potential confounding variables when possible (P<0.001). When further examining this association according to ACI subtype (and adjusting only for time from stroke to blood draw), we observed significant associations for patients with CE (P<0.001) or UDE (P=0.006) but not with LAA (P=0.41; Table S8). When adjusting for time from stroke to blood draw only, we additionally observed that higher NFL associated with higher modified Fisher scores in patients with aSAH (P=0.008), but not with ABC/2 scores in the smaller group of patients with ICH (P=0.063; Table 1).

Table 1:

Associations of NFL concentrations with ASPECTS (patients with ACI), modified Fisher grading scale scores (patients with aSAH), and ABC/2 scores (patients with non-traumatic ICH)

				Association with NFL concentrations
				Adjusting for time from stroke to blood draw		Adjusting for time from stroke to blood draw, age at blood draw, sex, BMI, current smoking, hypertension, diabetes, and physical activity1
Variable2	N	Patient group	Association measure	Estimate (95% CI)	P-value	Estimate (95% CI)	P-value
ASPECTS
Overall	225	ACI	Odds ratio	1.49 (1.28, 1.72)	<0.001	1.55 (1.32, 1.82)	<0.001
0–8 days from stroke to blood draw	195	ACI	Odds ratio	1.45 (1.23, 1.73)	<0.001	1.54 (1.28, 1.87)	<0.001
modified Fisher score=4
Overall	58	aSAH	Odds ratio	1.70 (1.20, 2.64)	0.008	N/A	N/A
ABC/2
Overall	29	ICH	β coefficient	0.39 (−0.02, 0.80)	0.063	N/A	N/A

CI=confidence interval. Associations of NFL concentrations with ASPECTS, modified Fisher score=4, and ABC/2 score were evaluated using proportional odds logistic regression models (ASPECTS), binary logistic regression models (modified Fisher score=4) and linear regression models (ABC/2 scores) that were adjusted for length of time from stroke to blood draw. For ASPECTS, the odds ratio is interpreted as the multiplicative increase in the odds of a lower ASPECTS for each doubling in NFL concentrations. For modified Fisher score=4, the odds ratio is interpreted as the multiplicative increase in the odds of a modified Fisher score equal to 4 for each doubling in NFL concentrations. For ABC/2 scores, the β coefficient is interpreted as the increase in mean ABC/2 score (on the square root scale) for each doubling in NFL concentrations. Information regarding ASPECTS was unavailable for 2 of the 227 patients with ACI. P-values <0.05 were considered as statistically significant for analysis of modified Fisher and ABC/2 scores. For ASPECTS, which was analyzed in four different overall groups (all patients with ACI, and patients with LAA, CE or UDE), p-values <0.0125 were considered as statistically significant after correcting for multiple testing (see also Table S6).

¹A rigorous multivariable analysis adjusting for variables other than time from stroke to blood draw was only performed for the large ACI groups.

²For analysis of ASPECTS, the 9–20 days from stroke to blood draw subgroup was not examined due to an insufficient sample size for analysis of this ordered categorical variable. For analysis of modified Fisher score=4, the 0–8 days and 9–20 days from stroke to blood draw subgroups were not examined due to the small number of patients with a modified Fisher score <4 in these subgroups. For analysis of ABC/2 scores, the 0–8 days and 9–20 days from stroke to blood draw subgroups were not examined due to small sample size.

Plasma NFL is associated with neurological, functional and cognitive status at blood draw

We next examined the relationship between plasma NFL and clinical measures at the time of blood draw. The NIHSS, the modified Rankin Scale (mRS), and the Mini-Mental State Examination (MMSE) were used to respectively assess neurological, functional, and cognitive status. These analyses were performed for all patients with stroke combined (ACI, aSAH, and ICH) and separately for patients with ACI, aSAH, or ICH. In all patients with stroke, higher NFL associated with a worse performance for all three tests. NFL positively correlated with NIHSS and mRS scores (all P<0.001, Table 2), and negatively correlated with MMSE scores (P<0.001, Table S9) when adjusting for time from stroke to blood draw, and also when additionally adjusting for potential confounding variables. Similar findings were observed when stratifying patients by time from stroke to blood draw (0–8 or 9–20 days), and when separately analyzing ACI and aSAH groups or ACI subtypes (Tables 2 and Tables S9–S11). In patients with ICH, NFL positively correlated with NIHSS (P=0.004, Table 2), and nominally negatively correlated with MMSE scores (P=0.016, Table S9). Due to the lack of variation in mRS scores in this small group, associations between NFL and mRS scores were not analyzed for patients with ICH.

Table 2:

Associations of NFL concentrations with NIHSS and mRS scores at blood draw

	Association between NFL concentrations and NIHSS at blood draw					Association between NFL concentrations and mRS score >3 at blood draw
	N	Adjusting for time from stroke to blood draw		Adjusting for time from stroke to blood draw, age at blood draw, sex, BMI, current smoking, hypertension, diabetes, and physical activity1		N	Adjusting for time from stroke to blood draw		Adjusting for time from stroke to blood draw, age at blood draw, sex, BMI, current smoking, hypertension, diabetes, and physical activity1
Disease group		β (95% CI)	P-value	β (95% CI)	P-value		OR (95% CI)	P-value	OR (95% CI)	P-value
All patients with stroke1
Overall	309	0.57 (0.49, 0.66)	<0.001	0.56 (0.47, 0.65)	<0.001	308	1.95 (1.65, 2.35)	<0.001	2.01 (1.66, 2.47)	<0.001
0–8 days	249	0.60 (0.50, 0.71)	<0.001	0.57 (0.46, 0.68)	<0.001	248	1.96 (1.60, 2.45)	<0.001	1.86 (1.49, 2.37)	<0.001
9–20 days	60	0.60 (0.42, 0.78)	<0.001	N/A	N/A	60	2.16 (1.46, 3.54)	<0.001	N/A	N/A
ACI
Overall	223	0.49 (0.40, 0.58)	<0.001	0.47 (0.37, 0.56)	<0.001	222	1.80 (1.49, 2.22)	<0.001	1.84 (1.50, 2.32)	<0.001
0–8 days	193	0.50 (0.39, 0.61)	<0.001	0.46 (0.34, 0.58)	<0.001	192	1.78 (1.43, 2.26)	<0.001	1.68 (1.32, 2.19)	<0.001
9–20 days	30	0.49 (0.27, 0.71)	<0.001	N/A	N/A	30	2.03 (1.28, 3.90)	0.010	N/A	N/A
aSAH
Overall	57	0.87 (0.61, 1.13)	<0.001			57	2.92 (1.79, 5.55)	<0.001	N/A	N/A
0–8 days	31	1.08 (0.74, 1.42)	<0.001	N/A	N/A	31	8.31 (2.75, 57.68)	0.003	N/A	N/A
9–20 days	26	0.76 (0.41, 1.12)	<0.001	N/A	N/A	26	2.42 (1.23, 6.13)	0.024	N/A	N/A
Non-traumatic ICH3
Overall	29	0.53 (0.18, 0.88)	0.004	N/A	N/A	N/A	N/A	N/A	N/A	N/A
0–8 days	25	0.31 (−0.02, 0.64)	0.064	N/A	N/A	N/A	N/A	N/A	N/A	N/A

β=regression coefficient; CI=confidence interval; OR=odds ratio. β values, 95% CIs, and p-values result from linear regression models. ORs, 95% CIs, and p-values result from binary logistic regression models. β values are interpreted as the change in mean NIHSS score (on the square root scale) for each doubling in NFL concentrations. ORs are interpreted as the multiplicative increase in the odds of a mRS score > 3 for each doubling in NFL concentrations. P values < 0.0071 (associations with NIHSS at blood draw) and <0.0083 (associations with mRS score >3 at blood draw) are considered statistically significant after correcting for multiple testing (see also Table S10 for association of NFL concentrations with NIHSS and mRS at blood draw by ACI subtype). Information was unavailable regarding NIHSS scores for five of the 314 patients. Information was unavailable regarding mRS scores for six of the 314 patients.

¹All models involving all patients with stroke were also adjusted for stroke type (ACI, aSAH, or ICH).

²A rigorous multivariable analysis adjusting for variables other than time from stroke to blood draw was only performed for the large ACI and all patients with stroke groups.

³Patients with ICH were not included in analysis of mRS at blood draw due to their lack of variability for this outcome.

Plasma NFL predicts 3-month functional outcome

mRS scores 3 months after stroke were available for 234 patients (74.5%), allowing us to evaluate associations between plasma NFL measured within 20 days of stroke and this 3-month functional outcome for 168 patients with ACI, 50 with aSAH, and 16 with ICH. In all patients combined, and in patients with ACI or aSAH, NFL was significantly associated with a 3-month mRS score >3 when adjusting for time from stroke to blood draw, and when also adjusting for potential confounding variables when possible (P<0.001, Fig. 2A, Table S12). These associations were consistent in analysis stratified by time from stroke to blood draw (0–8 or 9–20 days, Table S12). When assessing associations between NFL and 3-month mRS scores according to ACI subtype, higher NFL was associated with greater mRS scores in patients with CE (P<0.001) and UDE (P=0.004), but not with LAA (P=0.34) in analysis adjusting only for time from stroke to blood draw (Table S13).

Fig. 2. Plasma NFL predicts 3-month mRS and survival.

(A) Comparison of plasma NFL in patients with stroke with a 3-month mRS score of ≤3 or >3. Floating bars represent the maximum, mean and minimum NFL concentration for each group. *** P <0.001 using a binary logistic regression model adjusting for time from stroke to blood draw, age at blood draw, sex, BMI, current smoking, hypertension, diabetes and physical activity. NFL concentrations are shown on the base 10 logarithm scale. See also Table S12. (B-D) Survival after blood draw for all patients with stroke combined (B, N=314), and separately for patients with acute cerebral infarction (ACI, N=227, panel C) or with aneurysmal subarachnoid hemorrhage (aSAH, N=58, panel D). For ease of presentation, plasma NFL concentrations were divided into a two-level categorical variable based on sample medians. In (B): Low ≤79.1 pg/ml, High >79.1 pg/ml). In (C): Low ≤68.9 pg/ml, High >68.9 pg/ml). In (D): Low ≤68.6 pg/ml, High >68.6 pg/ml). See also Table S16.

For the two largest groups of all patients with stroke combined, and patients with ACI, we next examined whether these associations of increased NFL with greater 3-month mRS scores were independent of NIHSS scores at stroke or at blood draw. When additionally adjusting the aforementioned multivariable models for NIHSS scores at the time of stroke, associations between NFL and 3-month mRS scores remained significant for all patients with stroke combined and for patients with ACI (all P≤0.001, Table S14). However, when adjusting models for NIHSS scores at the time of blood draw, associations between NFL and mRS scores weakened for all patients with stroke (P=0.012) and for patients with ACI (P=0.041; Table S14).

We also examined the ability of NFL to independently predict an mRS score >3 by estimating the area under receiver operating characteristic curve (AUC) with and without NFL included in the multivariable logistic regression model. An AUC equal to 1.0 indicates perfect predictive ability for a given model, whereas an AUC equal to 0.5 represents predictive ability equal to chance. For all patients with stroke, adding NFL into the multivariable models that included NIHSS scores at stroke or at blood draw (in addition to the other aforementioned confounding variables) increased the AUC from 0.908 to 0.925 (when considering NIHSS at stroke) or from 0.940 to 0.947 (when considering NIHSS at blood draw) (Table S15). Results were similar for patients with ACI (Table S15).

Plasma NFL predicts survival

We next investigated whether NFL predicts survival in patients with stroke. With a median length of follow-up of 8.7 months (range: 0 days – 8.2 years), 112 (35.7%) patients died. In analyses adjusting only for time from stroke to blood draw, higher NFL associated with significantly poorer survival for all patients with stroke combined (HR=1.55, P<0.001), patients with ACI (HR=1.49, P<0.001), and patients with aSAH (HR=1.88, P<0.001) (Fig. 2B–D, Table S16). These results were consistent when also adjusting for age at blood draw, sex, BMI, current smoking, hypertension, diabetes and physical activity for all patients with stroke combined and for patients with ACI (Table S16). Similar results were seen when evaluating associations between NFL and survival according to ACI subtypes (Table S17).

For the two largest patient groups (all patients with stroke, and patients with ACI), we assessed whether associations of increased NFL with worse survival were independent of NIHSS scores. These associations indeed remained significant when additionally adjusting multivariable models for NIHSS scores at the time of stroke or at blood draw (P≤0.007, Table S18). We also examined the ability of NFL to independently predict survival in all patients with stroke by estimating the c-index with and without NFL included in the multivariable model, where a c-index of 1.0 for a given Cox regression model indicates perfect predictive ability, and a value of 0.5 indicates a predictive ability equal to chance. These analyses revealed that adding NFL into multivariable Cox regression models that also included the aforementioned potential confounding variables and either NIHSS scores at stroke or at blood draw respectively increased the c-index from 0.774 to 0.793 or from 0.810 to 0.816 (Table S19). Results were similar for patients with ACI alone (Table S19).

Plasma NFL is elevated and progressively increases following ICH in patients from the validation cohorts

Given that we found only one study investigating blood NFL as a prognostic biomarker for ICH (25), and given the lack of longitudinal assessments of blood NFL following ICH, we sought to validate and extend our findings above using two independent sample series from patients with ICH. The first corresponded to a cohort study of ICH from Yale University (Yale ICH study) that collected plasma from 96 individuals 72 hours following ICH. The second was a longitudinal cohort from the University of Pennsylvania (UPenn) composed of serum from 54 individuals collected 0 to 12, 12 to 36, and 60 to 84 hours after ICH. For comparison to these two cohorts, NFL was measured in plasma or serum from 48 and 46 healthy controls, respectively. Samples from these individuals, who were distinct from the control group used in our discovery study, were collected at Mayo Clinic Florida or at Massachusetts General Hospital and Washington University through their Dominantly Inherited ALS (DIALS) Network study. Baseline characteristics, comorbidities, stroke information and post-stroke outcomes are summarized in Table S20 for individuals in each ICH cohort and for control subjects.

As in our discovery study, plasma NFL concentrations strongly correlated with age in healthy controls (Spearman’s r=0.77, P<0.001, Fig. S1B), as did serum NFL (Spearman’s r=0.76, P<0.001, Fig. S1C). As anticipated, plasma and serum NFL concentrations from 46 matched controls were also strongly correlated (Spearman’s r=0.96, P<0.001, Fig. S1D). Compared to healthy controls, NFL was significantly elevated in patients from the Yale ICH study, and for all three stroke-to-blood draw time periods (0–12 hours, 12–36 hours and 60–84 hours) in the longitudinal UPenn cohort, both in unadjusted analysis and when adjusting for age at blood draw and sex (all P<0.001, Fig. 3 A, B, Tables S21 and S22).

Fig. 3. Blood NFL is elevated in patients with intracerebral hemorrhage in the validation groups.

(A) Comparison of plasma NFL between healthy controls (N=48) and patients with intracerebral hemorrhage (ICH) in the Yale ICH study series (N=96). (B) Comparison of serum NFL between healthy controls (N=46) and patients with ICH in the longitudinal UPenn series (N=54). ***P < 0.001 in unadjusted analysis comparing NFL between the indicated group and healthy controls. ^###P < 0.001 in analysis adjusting for age at blood draw and sex comparing NFL between the indicated group and healthy controls. ^†††P ≤ 0.002 when comparing the indicated group to the 60–84 hour group by paired Wilcoxon signed rank test. The black horizontal bars represent median NFL concentrations. NFL concentrations are shown on the base 10 logarithm scale. See also Tables S21, S22 and S23.

We evaluated the temporal profile of NFL in the longitudinal serum series through pair-wise comparisons of NFL among the three stroke-to-blood draw time periods. There was no increase in NFL from 0–12 hours to 12–36 hours post-stroke. However, NFL in serum collected 60–84 hours post-stroke was significantly elevated in comparison to concentrations at both earlier time periods (P≤0.002, Fig. 3, Table S23).

Plasma NFL correlates with hemorrhage volume in patients with ICH in the validation cohorts

In the Yale ICH study, we observed a strong positive association between NFL measured 72 hours post-ICH and hemorrhage volume, determined using the ABC/2 formula, both in unadjusted analysis and multivariable models adjusted for age, sex, hypertension, diabetes and BMI (P<0.001, Table 3). For the longitudinal UPenn series, no such association was seen when NFL was measured in plasma collected 0–12 hours following ICH in either unadjusted analysis or multivariable models adjusted for age, sex, hypertension and diabetes (Table 3). However, positive associations between hemorrhage volume and NFL were detected when NFL was measured 12–36 hours (P=0.009) or 60–84 hours (P=0.026) following ICH though findings did weaken for the smaller 60–84 hour group (N=29) in multivariable analysis (Table 3).

Table 3:

Association of NFL concentrations with ICH volume in patients with ICH from the validation cohorts

		Association between NFL and ICH volume
	N	Unadjusted analysis		Multivariable analysis1
Group		β (95% CI)	P-value	β (95% CI)	P-value
Yale ICH study series	96	0.62 (0.30, 0.94)	<0.001	0.60 (0.27, 0.93)	<0.001
UPenn ICH study series
0–12 hours from stroke to blood draw	36	0.13 (−0.42, 0.68)	0.65	0.12 (−0.53, 0.77)	0.72
12–36 hours from stroke to blood draw	40	0.63 (0.18, 1.07)	0.009	0.74 (0.21, 1.27)	0.010
60–84 hours from stroke to blood draw	29	0.71 (0.12, 1.31)	0.026	0.51 (−0.17, 1.19)	0.16

β=regression coefficient; CI=confidence interval; β values, 95% CIs, and p-values result from linear regression models. β values are interpreted as the change in mean ICH volume (on the square root scale) for each doubling in NFL concentration.

¹Multivariable models were adjusted for age, sex, hypertension, diabetes, and BMI for the Yale ICH study series, and for age, sex, hypertension, and diabetes for the longitudinal UPenn ICH series. P-values < 0.0125 are considered statistically significant after correcting for multiple testing.

Plasma NFL is associated with neurological status at initial evaluation or at blood draw in the validation cohorts

In the Yale ICH study, we observed a strong positive association of NFL with NIHSS at initial evaluation and at blood draw both in unadjusted analysis and when adjusting for age, sex, hypertension, diabetes and BMI (P<0.001, Table 4). Serum NFL similarly associated with NIHSS at initial patient evaluation in the longitudinal UPenn cohort in unadjusted analysis and when adjusting models for age, sex, hypertension and BMI, and this was most evident when NFL was measured 12–36 and 60–84 hours following ICH (Table 4).

Table 4:

Associations of NFL concentrations with NIHSS at initial evaluation and at blood draw in patients with ICH from the validation cohorts

		Association between NFL and NIHSS at initial evaluation				Association between NFL and NIHSS at blood draw
	N	Unadjusted analysis		Multivariable analysis1		Unadjusted analysis		Multivariable analysis1
Group		β (95% CI)	P-value	β (95% CI)	P-value	β (95% CI)	P-value	β (95% CI)	P-value
Yale ICH study series	96	0.51 (0.28, 0.73)	<0.001	0.52 (0.29, 0.75)	<0.001	0.55 (0.28, 0.82)	<0.001	0.58 (0.28, 0.89)	<0.001
UPenn ICH study series
0–12 hours from stroke to blood draw	36	0.34 (0.05, 0.63)	0.030	0.32 (−0.01, 0.65)	0.066	N/A	N/A	N/A	N/A
12–36 hours from stroke to blood draw	40	0.42 (0.20, 0.64)	<0.001	0.42 (0.18, 0.65)	0.001	N/A	N/A	N/A	N/A
60–84 hours from stroke to blood draw	29	0.57 (0.24, 0.91)	0.003	0.49 (0.12, 0.85)	0.016	N/A	N/A	N/A	N/A

β=regression coefficient; CI=confidence interval; β values, 95% CIs, and p-values result from linear regression models. β values are interpreted as the change in mean NIHSS (on the square root scale) for each doubling in NFL concentration.

¹Multivariable models were adjusted for age, sex, hypertension, diabetes, and BMI for the Yale ICH study series, and for age, sex, hypertension, and diabetes for the longitudinal UPenn ICH series. P-values < 0.0125 are considered statistically significant after correcting for multiple testing. Information was unavailable regarding NIHSS at initial evaluation (24 from the Yale ICH study series) and NIHSS at blood draw (41 from the Yale ICH study series).

Plasma NFL predicts 3-month and 6-month functional outcomes in patients with ICH

The functional status of patients with ICH in the Yale ICH study was evaluated 3 and 6 months post-stroke. In both unadjusted analysis and analysis adjusted for age, sex, hypertension, diabetes and BMI, we noted positive associations between plasma NFL and the occurrence of an mRS score >3, with findings being most pronounced for the 6-month time point (Table 5). Three-month functional status was also available for patients in the longitudinal UPenn cohort, and again positive associations were observed between NFL and mRS scores in unadjusted analysis and when adjusting for age and sex when possible. These associations were strongest when NFL was measured 12–36 or 60–84 hours following ICH (Table 5).

Table 5:

Associations of NFL concentrations with mRS at 3 months and 6 months after stroke in patients with ICH from the validation cohorts

		Association between NFL and mRS >3 at 3 months after stroke				Association between NFL and mRS >3 at 6 months after stroke
	N	Unadjusted analysis		Multivariable analysis1		Unadjusted analysis		Multivariable analysis1
Group		OR (95% CI)	P-value	OR (95% CI)	P-value	OR (95% CI)	P-value	OR (95% CI)	P-value
Yale ICH study series	96	1.45 (1.07, 2.02)	0.020	1.45 (1.04, 2.08)	0.034	2.23 (1.53, 3.46)	<0.001	2.24 (1.51, 3.59)	<0.001
UPenn ICH study series
0–12 hours from stroke to blood draw	36	1.96 (1.15, 4.14)	0.032	2.22 (1.23, 4.90)	0.020	N/A	N/A	N/A	N/A
12–36 hours from stroke to blood draw	40	3.32 (1.68, 9.14)	0.005	4.14 (1.84, 13.95)	0.005	N/A	N/A	N/A	N/A
60–84 hours from stroke to blood draw	29	4.23 (1.61, 19.22)	0.017	N/A	N/A	N/A	N/A	N/A	N/A

OR=odds ratio; CI=confidence interval; ORs, 95% CIs, and p-values result from binary logistic regression models. ORs are interpreted as the multiplicative increase in the odds of a mRS > 3 for each doubling in NFL concentration.

¹Multivariable models were adjusted for age, sex, hypertension, diabetes, and BMI for the Yale ICH study series, and for age and sex for the longitudinal ICH series (note that multivariable analysis was not possible for the 60–84 hours from stroke to blood draw group due to its small sample size). P-values < 0.0125 are considered statistically significant after correcting for multiple testing. Information was unavailable regarding mRS at 6 months (4 from the Yale ICH study series).

For both validation cohorts, we next examined whether associations of increased NFL with worse functional outcomes based on mRS scores were independent of NIHSS scores at initial evaluation. When additionally adjusting multivariable regression models for NIHSS scores, the previously observed associations between NFL and 3-month or 6-month mRS scores remained consistent (Table S24). These data indicate that associations of blood NFL with increased risk of mRS scores >3 were independent of NIHSS at the time of initial evaluation.

We also examined the ability of NFL to independently predict a mRS scores >3 by estimating the AUC with and without NFL added to the multivariable logistic regression models that included NIHSS scores at initial evaluation as well as other previously mentioned confounding variables. In the Yale ICH study, adding NFL into the multivariable model examining associations with 3-month mRS scores increased the AUC from 0.749 to 0.754 (Table S25). An even greater increase in the AUC was observed when NFL was included into the multivariable model examining associations with 6-month mRS scores (from 0.780 to 0.819; Table S25). In the longitudinal cohort, adding NFL to the multivariable models examining associations with three month mRS scores increased the AUC from 0.929 to 0.941 when NFL was measured 0–12 hours from stroke, and from 0.891 to 0.934 when NFL was measured 12–36 hours from stroke (Table S25).

Discussion

Stroke severity can vary drastically from temporarily symptomatic and non-disabling to disabling and ultimately fatal. This heterogeneity underscores the need for robust measures of clinical outcomes. Reliable prognosis not only is important to the patient, family and caregivers, it informs treatment and rehabilitation decisions, and can improve multiple aspects of clinical trial design. Despite encouraging data supporting blood NFL as an indicator of post-stroke neuronal injury, it has yet to be translated to the clinic as a stroke biomarker. This is partly due to the lack of independent validation studies among multiple centers, and insufficient rigorous investigations across stroke types. Towards overcoming these barriers, we evaluated the prognostic utility of blood NFL in ischemic and hemorrhagic stroke.

In our discovery study, we noted elevated plasma NFL in patients with ACI, aSAH or ICH compared to controls with differences most evident for patients with ICH, which has a high morbidity and mortality rate (30). Blood NFL was also higher in patients with ICH in our two validation cohorts compared to healthy controls. For patients with ischemic stroke, a rise in blood NFL has been observed within 24 hours of stroke or admission to hospital (14, 16, 17), and longitudinal studies consistently show a progressive increase in its concentration, which is estimated to peak between ~7 to 21 days but can remain elevated for several months (16–18, 20). Longitudinal measures of blood NFL in patients with aSAH and ICH have not been reported, but analysis of our cross-sectional discovery cohorts suggests that NFL progressively rises after aSAH and ICH. This was confirmed in our longitudinal cohort of patients with ICH showing that serum NFL was higher 60–84 hours post-stroke in comparison to concentrations measured within 36 hours of ictus. The gradual increase in NFL following stroke may reflect axonal degeneration secondary to the primary insult. This may partly explain why the differences in NFL that we observed between healthy controls and patients with stroke were especially evident when blood was drawn 9 to 20 days (vs. 0 to 8 days) after the stroke in our discovery cohort or 60–84 hours (vs. 0–36 hours) after ICH in our longitudinal validation cohort.

To test the prognostic utility of NFL across stroke types, we investigated the relationship between blood NFL and widely utilized outcome measures. We found that NFL associated with radiographic indicators of brain tissue damage. In patients with ACI, NFL associated with ASPECTS, a 10-point topographic grading system used to estimate ischemic changes on CT scans (31). These data are in agreement with a study showing that serum NFL correlates with infarct volume when measured 3 or 7 days following ACI (17). However, no association between infarct volume and serum NFL measured within 24 hours of admission was observed (14, 17, 22). We also noted that plasma NFL associated with radiologic severity in patients with aSAH. The latter was assessed using the modified Fisher grading scale, which grades SAH secondary to intracranial aneurysm rupture (32). We found that high plasma NFL increased the odds of a worse score. In contrast to our finding, it was reported that NFL measured in plasma collected from patients at admission to hospital for aSAH does not correlate with Fisher scores (23). We also observed that higher NFL associated with intracranial hemorrhage volume in both ICH validation cohorts. In the longitudinal cohort, this association was apparent for NFL measured 12–36 and 60–84 hours following ICH, but not during the earliest time period (0–12 hours). As a whole, these findings strongly support blood NFL as a predictor of radiologic outcomes, and highlight that the timing of NFL measurements must be considered to maximize their predictive value.

We additionally examined associations between NFL and indicators of neurological, functional and cognitive status at the time of blood draw or at first evaluation. In our Mayo Clinic sample series, higher NFL associated with worse outcomes based on the NIHSS, mRS and MMSE at blood draw in all patients with stroke combined, and separately for patients with ACI or aSAH. For patients with ICH in the discovery cohort, plasma NFL correlated with NIHSS scores, and nominally correlated with MMSE scores. Similarly, NFL correlated with NIHSS scores at time of blood draw and/or at initial evaluation in our two validation cohorts of patients with ICH. Prior studies evaluating the relationship between blood NFL and clinical scales at the time of blood draw are sparse but Traenka et al. reported that serum NFL correlates with NIHSS scores at the time of blood sampling in patients with cervical artery dissection (26). Although associations between blood neurofilament and MMSE scores in patients with stroke have not, to our knowledge, been examined, blood NFL does correlate with MMSE scores in other neurodegenerative disease (33, 34).

The ability of NFL to predict the clinical status of patients at blood draw is expected to inform their prognosis. Our data demonstrating the relationship between NFL and future outcomes, namely 3- and 6-month mRS scores, support this notion. Our findings that blood NFL in patients with ischemic stroke correlate with 3-month mRS are consistent with those of prior studies. The Dichgans group found that NFL in serum collected 7 days after admission to hospital for ACI independently correlated with 3-month mRS scores (17). Similarly, Pedersen et al. observed that serum NFL measured within 14 days of ACI or 3 months after ictus associated with 3-month mRS. However, these associations were lost when analyses were limited to NFL measured between 6–8 days or within 24 hours of stroke (20). Because only 24 patients had blood sampled within 24 hours, the analysis of early NFL measures with 3-month mRS scores was likely underpowered in this study (20). Nevertheless, discrepant findings exist among studies evaluating early NFL concentrations and functional outcomes of ischemic stroke. Whereas De Marchis et al. failed to detect associations of blood NFL measured 24 hours post-stroke and 3-month mRS scores (14), Uphaus et al. reported that serum NFL measured 24 hours after admission did predict unfavorable 3 months mRS scores (22).

As observed in patients with ischemic stroke, we found blood NFL to associate with 3-month mRS scores in patients with aSAH, and with 3- and 6-month mRS scores in patients with ICH. In our longitudinal study of patients with ICH, 3-month mRS scores correlated with NFL measured 12 to 36 hours, but not 0 to 12 hours, after stroke. In a previous study, plasma NFL measured in patients with aSAH at admission reportedly correlated with 1-month mRS scores despite NFL concentrations not being higher than in control individuals (23). In patients with ICH, neither plasma NFL measured at admission or 24 hours after ictus associated with 1-month mRS scores (25).

Finally, given that stroke is a leading cause of death worldwide, and given the absence of a blood biomarker for all-cause mortality, we evaluated whether plasma NFL would fulfill this purpose. Both in all patients with stroke combined, and in patients with ACI or aSAH alone, NFL correlated with survival after blood draw.

Strengths of our study include: 1) the evaluation of three types of stroke, as well as ACI subtypes, in patients recruited concurrently from the same center and assessed using the same assay thus allowing direct comparisons among groups; 2) the inclusion of two independent validation cohorts to confirm our findings in ICH; 3) the longitudinal assessment of serum NFL in patients with ICH; 4) the use of a highly sensitive and well-validated assay to quantify NFL (35); and 5) the rigor of our statistical analyses.

Our study also has limitations. First, as this was a retrospective study, our discovery sample series was comprised of plasma drawn from patients at various times from the day of stroke up to 20 days post-stroke. While one advantage to this design is that it allowed us to explore when, after a stroke, NFL rises above normal, the heterogeneity in the timing of blood sampling has potential to influence associations of interest. To address this possibility, our association analyses were both stratified by time from stroke to blood draw, and adjusted for time from stroke to blood draw. Second, the sizes of the discovery aSAH and ICH cohorts were relatively small, which decreased the power of some analyses and/or our ability to examine certain associations. The fact that our findings in patients with ACI were congruent with those of other studies (17, 18, 20), and that our findings in patients with ICH were validated across three independent groups, nonetheless gives us confidence in the soundness of our results.

Findings from the present study have important implications. Although our data, in aggregate with those of others, suggest less utility for NFL measured within 24 hours of stroke in guiding acute patient care decisions, they do provide compelling evidence supporting blood NFL as a promising prognostic biomarker for patients with ACI, aSAH and ICH. At present, age and stroke severity are among the strongest predictors of outcome. Stroke severity is commonly estimated based on the degree of neurologic impairment determined by scales like the NIHSS (36–38). Functional, cognitive, and radiographic scales also provide insights into stroke severity. Yet, like the NIHSS, these scales are often non-continuous measures of the extent of tissue injury, and some scales are specific to a stroke type. Furthermore, many clinical circumstances compromise their validity as a marker of new injury specific for an acute stroke. For instance, anesthetic and paralytic agents as well as medications given for pain, agitation or depression can influence the accuracy of the NIHSS and other scales. These circumstances would not affect measures of blood NFL.

In addition to its prognostic value, blood NFL can positively impact multiple facets of ischemic and hemorrhagic stroke research. Subacute NFL measures would be of great value to clinical trials seeking to reduce stroke complications, or to reduce final neuroaxonal injury. They also have potential to substitute for repeated head imaging. Although head imaging is generally tolerated, it can expose patients to ionizing radiation (39) and the risks of acute adverse reactions to contrast (40) and claustrophobia (41), as well as add to the cost of care (42). Not only would NFL measures be more cost-effective than protocol-specified brain magnetic resonance imaging (MRI), they could be applied to all patients, even those who are not suitable MRI candidates (due to claustrophobia, ferromagnetic devices or medical instability). For example, NFL would be useful for clinical trials comparing pre-perfusion/recanalization strategies for treating acute ischemic stroke, which often rely on final infarct volume as a surrogate endpoint (43). Moreover, post-operative NFL concentrations could be used as a biomarker in lieu of brain imaging to detect symptomatic ischemic stroke and silent brain infarcts.

Our data highlight areas in need of additional investigation if they are to be applied to the clinic. Among the key steps in translating our findings from bench to bedside are delineating the earliest optimal window at which blood NFL concentrations correlate with long-term functional outcomes, and establishing normative NFL values by age in healthy individuals. Furthermore, including blood NFL as a prognostic biomarker in randomized clinical trials focusing on recovery would be particularly informative. Should blood NFL correlate with meaningful outcomes (disability, function, cognition) in this setting, its incorporation into the care of patients in the clinic would be justified.

Overall, based on our data indicating that plasma NFL provides a uniform method to estimate neuroaxonal injury and forecast outcomes, including all-cause mortality, across stroke mechanisms, it has the potential to assist in planning supportive and rehabilitation services, and with improving the efficiency of trials of acute therapeutic and recovery-enhancing interventions.

Materials and Methods

Study design

The goal of this study was to investigate blood NFL as a predictor of neuroaxonal injury and all-cause mortality across major types of ischemic and hemorrhagic stroke. NFL in plasma or serum was measured using the NF-Light digital immunoassay from Quanterix (see Supplemental Materials and Methods). Our primary analyses were to determine whether blood NFL is elevated post-stroke, and whether NFL associates with: 1) radiographic markers; 2) neurological, functional and cognitive status at blood draw/initial evaluation or 3-months post-stroke; and 3) all-cause mortality. We first completed a retrospective study involving 393 individuals enrolled in the Mayo Clinic Florida Familial Cerebrovascular Diseases Registry (314 patients with ACI, aSAH or ICH and 79 control individuals) (27). We included all patients (no randomization) for whom plasma was available and collected within 20 days from stroke. Our study also included two independent sample series from patients with ICH: 1) plasma collected from 96 patients 72 hours following ICH through the Yale ICH study; and 2) serum collected longitudinally from 54 patients at UPenn. NFL was measured in a blinded manner. Sample sizes for all series were based on what as available when the study was initiated and not on sample size calculations. Written informed consent was obtained from all participants or their legally authorized surrogate if they were unable to give written consent, and biological samples were obtained with ethics committee approval. Additional information on our samples series is provided in Supplemental Materials and Methods

Statistical Analysis – Mayo Clinic discovery series

All statistical tests for the discovery and validation series were two-sided and performed using R Statistical Software (version 3.6.2; R Foundation for Statistical Computing, Vienna, Austria).

NFL was compared according to time from stroke to blood draw in patients with stroke (0–8 days vs. 9–20 days) using linear regression models that were adjusted for stroke group only, and also additionally adjusted for age at blood draw, sex, BMI, current smoking, hypertension, diabetes and physical activity. Due to its skewed distribution, NFL concentrations were examined on the base 2 logarithm scale in this and all subsequently described regression analysis.

In addition to performing statistical analyses irrespective of length of time from stroke to blood draw, and adjusting regression models involving solely patients with stroke for this time length (as a continuous variable), we also performed analyses (sample size permitting) in the separate patient subgroups with a length of time from stroke to blood draw of 0 to 8 days or 9 to 20 days (while still adjusting for time from stroke to blood draw).

The association between NFL and age in controls was examined using Spearman’s test of correlation. Comparisons of NFL concentrations between controls and patients with stroke (overall and according to stroke type and ACI subtype) were made using linear regression models. β coefficients and 95% confidence intervals (CIs) were estimated and are interpreted as the difference in mean NFL concentration (on the base 2 logarithm scale) between the two groups of interest. Models were initially adjusted for length of time from stroke to blood draw as a linear term and a quadratic term to allow for a non-linear relationship between this variable and NFL concentrations. Subsequently, linear regression models were also adjusted for a pre-defined group of potential confounding variables, which included age at blood draw, sex, BMI, current smoking, hypertension, diabetes and physical activity, when the sample size allowed for such a multivariable analysis (29). In order to adjust for multiple testing, we utilized a Bonferroni correction separately for each group of similar tests, after which p-values <0.0063 were considered as significant when making comparisons of NFL among controls and patients with stroke (8 overall pair-wise comparisons), and p-values <0.0056 were considered as significant when comparing NFL among different stroke subtypes (9 overall pair-wise comparisons).

Associations of NFL with radiographic markers (ASPECTS, modified Fisher score, and ABC/2) were examined using regression models appropriate for the given radiographic measure. In patients with ACI (overall and separately for ACI subtypes), the association between NFL and ASPECTS was assessed using proportional odds logistic regression models that were adjusted for time from stroke to blood draw, where ASPECTS between 0 and 6 were collapsed into one category due to their rare frequencies. Odds ratios (ORs) and 95% CIs were estimated and are interpreted as the multiplicative increase in the odds of a lower ASPECTS for each doubling in NFL concentration. Additional multivariable adjustment for the aforementioned pre-defined confounding variables was also examined in the larger overall ACI group only. In patients with aSAH, the association between NFL and modified Fisher score was examined using a binary logistic regression model that was adjusted for length of time from stroke to blood draw, where modified Fisher score was dichotomized as <4 or 4; an OR and 95% CI were estimated and correspond to each doubling in NFL concentration. Finally, in patients with ICH, the association between NFL and ABC/2 scores was evaluated using a linear regression model that was adjusted for length of time between stroke and blood draw, where ABC/2 scores were examined on the square root scale owing to their skewed distribution. A β coefficient and 95% CI were estimated and are interpreted as the change in the mean ABC/2 score (on the square root scale) for each doubling in NFL concentration. P-values <0.05 were considered as statistically significant for tests of association of NFL with the separate ABC/2 and modified Fisher score outcomes, which were each examined in only one patient group. As associations between NFL and ASPECTS were examined in four different groups of patients with ACI, p-values <0.0125 were considered as statistically significant in those analyses after applying a Bonferroni correction for multiple testing.

In patients with stroke (overall and according to stroke type and ACI subtype), associations of NFL with measures obtained at blood draw (NIHSS, mRS and MMSE) and post-blood draw outcomes (3-month mRS and survival) were evaluated using regression models that are appropriate for the nature of the given outcome measure. Specifically, associations between NFL and NIHSS were examined using linear regression models, where NIHSS was examined on the square root scale due to its skewed distribution. β coefficients and 95% CIs were estimated and are interpreted as the change in mean NIHSS score (on the square root scale) for each doubling of NFL concentration. Associations of NFL with mRS scores were evaluated using binary logistic regression models, where mRS was dichotomized as ≤3 or >3. ORs and 95% CIs were estimated. Associations between NFL and MMSE scores were examined using binary logistic regression models, where MMSE scores were dichotomized based on the sample median (<median, ≥median) in the given patient group being examined owing to its bimodal distribution. ORs and 95% CIs were estimated. Finally, associations between NFL concentration and survival were evaluated using Cox proportional hazards regression models. The baseline time point for survival was the date of blood draw, and censoring occurred on the date of last follow-up. Hazard ratios (HRs) and 95% CIs were estimated. All HR and OR estimates correspond to a doubling in NFL concentration. All regression models (Cox, linear, logistic) were initially adjusted for length of time from stroke to blood draw (as a linear term only). Subsequently, for the two largest groups of patients with ACI and all patients with stroke, Cox, linear, and logistic regression models were also adjusted for age at blood draw, sex, BMI, current smoking, hypertension, diabetes, and physical activity. We utilized a Bonferroni correction to account for multiple testing separately for each outcome measure, after which p-values <0.0071 (or <0.0083 for tests involving mRS and survival) were considered as statistically significant (6 or 7 overall tests of association were performed for each outcome).

We further evaluated the ability of NFL to independently predict 3-month mRS and survival in two ways. First, the aforementioned logistic regression (for 3-month mRS) and Cox regression (for survival) models were additionally adjusted for NIHSS (separately for NIHSS scores at stroke and NIHSS scores at blood draw). Second, we assessed the predictive ability of the aforementioned multivariable logistic and Cox regression models (including NIHSS scores) with and without NFL in the model. Specifically, when examining 3-month mRS (which was dichotomized as ≤3 or >3), AUC was estimated both with and without NFL in the multivariable model. When evaluating survival, we estimated c-index with and without NFL in the multivariable model.

Statistical Analysis – validation series

Similar to the Mayo Clinic discovery series, NFL concentrations were examined on the base 2 logarithm scale in all analysis of the validation series due to their skewed distribution. All analyses were performed separately for the Yale ICH study and the three separate stroke to blood draw time periods (0–12 hours, 12–36 hours, and 60–84 hours) for the longitudinal ICH series.

The association between NFL and age in controls was examined using Spearman’s test of correlation. Comparisons of NFL concentrations between controls and patients with ICH were made using linear regression models. β coefficients and 95% CIs were estimated and are interpreted as the difference in mean NFL concentrations (on the base 2 logarithm scale) between the two groups of interest. Unadjusted models were first examined, and these were followed by multivariable models that were adjusted for age at blood draw and sex. We applied a Bonferroni correction to adjust for the four different comparisons vs. controls that were made, after which p-values <0.0125 were considered as significant.

For the longitudinal ICH series, comparisons of NFL concentrations between the 0–12 hour, 12–36 hour, and 60–84 hour stroke to blood draw time periods were made using paired Wilcoxon signed rank tests. Only patients with ICH with an NFL measurement for both of the time periods involved in a given pair-wise comparison were included for analysis of that comparison. P-values <0.0167 were considered as statistically significant after utilizing a Bonferroni correction to adjust for the three pair-wise comparisons between groups that were made.

Associations of NFL with ICH volume and NIHSS at initial evaluation and blood draw were examined in patients using unadjusted and multivariable linear regression models. Multivariable models were adjusted for age, sex, hypertension, diabetes, and BMI for the Yale ICH study series, and for age, sex, hypertension, and diabetes for the longitudinal ICH series from UPenn. ICH volume and NIHSS were examined on the square root scale owing to its skewed distribution. β coefficients and 95% CIs were estimated and are interpreted as the change in the mean outcome measure (on the square root scale) for each doubling in NFL concentration. Associations of NFL with mRS scores at 3 and 6 months after stroke were evaluated using unadjusted and multivariable binary logistic regression models, where mRS was dichotomized as ≤3 or >3. ORs and 95% CIs were estimated and correspond to a doubling in NFL concentration. Multivariable models were adjusted for age, sex, hypertension, diabetes, and BMI for the Yale ICH study series, and for age and sex in the longitudinal ICH series (adjustment for additional variables in the ICH series was not possible for this binary outcome where ability to adjust for covariates is more limited). P-values <0.0125 were considered as statistically significant after applying a Bonferroni correction for multiple testing separately for each outcome measure, as association tests were performed for four separate patient groups.

We performed additional examination of the ability of NFL to independently predict 3-month and 6-month mRS in two ways, similar to the analysis of the Mayo Clinic discovery series. First, the aforementioned multivariable binary logistic regression models were additionally adjusted for NIHSS at initial evaluation (NIHSS at blood draw was not examined here due to missing data in the Yale ICH study series and lack of any data in the longitudinal ICH series). Next, we assessed the predictive ability (as measured by AUC) of the aforementioned multivariable logistic regression models (including NIHSS at initial evaluation) with and without NFL in the model.

Data and materials availability:

All requests for data or materials should be made to LP and JFM. They will be made available through appropriate administrative channels (material transfer agreements).