Contributors to serum NfL levels in people without neurologic disease

Abstract

Objective.

To assess the effects of demographics, lifestyle factors, and comorbidities on serum neurofilament light chain (sNfL) levels in people without neurologic disease and establish demographic-specific reference ranges of sNfL.

Methods.

The National Health and Nutrition Examination Survey (NHANES) is a representative sample of the US population in which detailed information on demographic, lifestyle, routine laboratory tests and overall health status are systematically collected. From stored serum samples, we measured sNfL levels using a novel high-throughput immunoassay (Siemens Healthineers). We evaluated the predictive capacity of 52 demographic, lifestyle, comorbidity, anthropometric, or laboratory characteristics in explaining variability in sNfL levels. Predictive performance was assessed using cross-validated R² (R²_cv) and forward selection was used to obtain a set of best predictors of sNfL levels. Adjusted reference ranges were derived incorporating characteristics using generalized additive models for location, scale and shape.

Results.

We included 1706 NHANES participants (average age: 43.6±14.8y; 50.6% male, 35% non-white) without neurological disorders. In univariate models, age explained the most variability in sNfL (R²_cv=26.8%). Multivariable prediction models for sNfL contained 3 covariates (in order of their selection): age, creatinine, and glycosylated hemoglobin (HbA1c) (standardized β – age: 0.46, 95% CI: 0.43, 0.50; creatinine: 0.18, 95% CI: 0.13, 0.22; HbA1c: 0.09, 95% CI: 0.06, 0.11). Adjusted centile curves were derived incorporating identified predictors. We provide an interactive R Shiny application to translate our findings and allow other investigators to use the derived centile curves.

Interpretation.

Results will help to guide interpretation of sNfL levels as they relate to neurologic conditions.

Introduction

Neurofilaments are neuron-specific cytoskeletal proteins that are released into the extracellular space following neuroaxonal damage. Increased neurofilament light chain (NfL) levels have been found in the blood and cerebrospinal fluid (CSF) across a wide spectrum of neurological disorders.^1–3 For example, people with amyotrophic lateral sclerosis (ALS) or frontotemporal dementia (FTD) have an estimated 10-fold increase in CSF NfL levels relative to healthy controls;² CSF NfL levels in these individuals are also strongly correlated with serum levels, and both are associated with worse functional status.^4,5 Similarly, in multiple sclerosis (MS), serum NfL (sNfL) levels are associated with clinico-radiological measures of disease activity, are modulated by MS disease modifying therapies, and predict disability worsening and brain atrophy.^6–10 Thus, combined with its strong relevance to underlying pathological disease processes and the relative ease with which serum samples can be obtained, sNfL is emerging as a potential disease prognostic and monitoring tool in clinical care for several neurological disorders.

Accurate interpretation of sNfL levels in the context of neurologic disease relies heavily on comparison with expected values from individuals without disease, and, as a result, availability of resources that include large populations of ‘normal’ individuals. Ideally, these resources should sufficiently characterize and capture the normal range of population-level variability in sNfL so that appropriate conclusions can be drawn pertaining to a given sNfL level from a patient with neurologic diseases. Understanding of normal variation in sNfL is complicated; levels are strongly positively correlated with age, and may vary by certain anthropometric characteristics or results of various laboratory tests.^11–13 Whether other factors beyond age such as smoking, alcohol use, diabetes mellitus, renal function or hypertension, impact sNfL levels outside the context of a person’s primary neurologic condition has not been explored in depth in large scale studies. Thus, a critical next step required for sNfL to enter clinical practice as a monitoring or prognostic biomarker is the 1) rigorous evaluation of the influence of demographic, lifestyle or non-neurologic comorbidities on sNfL levels and 2) the derivation of adjusted reference ranges that account for these factors.

To address this need, we perform a comprehensive evaluation of over 50 potential contributors to variability in sNfL levels using a large, well-characterized and representative cohort of US adults without neurologic disease. We quantify the individual, comparative, and combined predictive capacity of each characteristic as it relates to sNfL levels and use this information to derive adjusted reference ranges that can be used to interpret sNfL more accurately at the individual level.

Methods

Study population

Overview of overall study population

We include participants from the National Health and Nutrition Examination Survey (NHANES) 2013–2014 cycle. NHANES is a complex multistage representative sample of the United States (US) non-institutionalized population in which annual surveys and examinations are performed and is designed to observe the health and nutritional status of people living in the US. For each participant, a sampling weight is created to account for the complex sampling scheme, and weights are required by all in all subsequent analyses.

NHANES participants complete an in-person at home interview and a physical examination by a trained health care provider. Socio-demographic and lifestyle information including information on race, ethnicity, education status, income, employment, smoking, alcohol/drug use, diet, and physical activity are systematically collected. Other information on current health status, hospitalizations, common medical conditions (asthma, arthritis, diabetes, cancer, heart disease, stroke, lung disease, liver disease, thyroid disease, anemia), prescription medication usage (linked to corresponding ICD-10 codes), physical functioning, kidney conditions, and depression is also collected. Assessment of anthropometric measurements including body weight, height, and systolic/diastolic blood pressure is performed by a trained research technician within the mobile examination center (MEC). Participants also complete an oral glucose tolerance test. Eligible participants also undergo whole body dual-energy x-ray absorptiometry (DXA); total lean and fat mass are then quantified from these scans. Exclusion criteria for DXA scans include pregnancy, self-reported exposure to radiographic contrast use in the past 7 days, or self-reported weight ≥450 pounds (weight limit of the DXA table).

During the participant’s visit to the MEC, he or she provides a blood sample collected using a standardized protocol by a trained technician. Specimens are then 1) used to perform a series of common laboratory tests that include a complete blood count, standardized biochemistry panel, hemoglobin A1C, fasting glucose, insulin, a cholesterol panel, zinc levels, creatinine, vitamins D and B₁₂, among others and 2) stored for future research purposes. The biobanked specimens are aliquoted and frozen at −20°C until they are shipped weekly to the CDC’s central laboratory and where they are stored at −80°C. Quality control and quality assurance protocols for NHANES mobile centers are Clinical Laboratory Improvement Act (CLIA) compliant. The protocol for this study was approved by the National Center for Health Statistics Ethics Review Board; the Board determined that analyses of stored specimens conducted as a part of this study were in accordance with the informed consent information that was provided to the participants who yielded the specimens.

Eligible NHANES Participants

Our goal for this analysis is to identify contributors of variability in sNfL levels in a population without neurologic diseases or evidence of known neurologic dysfunction. Using the available ICD-10 codes linked to participants’ prescription medications, we exclude those with a history of cerebral infarction (I63), transient cerebral ischemic attack (G45.9), epilepsy or recurrent seizures (G40), Alzheimer’s disease (G30.9), multiple sclerosis (G35), mild cognitive impairment (G31.84), Parkinson’s disease (G20), degenerative disease of the CNS (G31.9), spastic hemiplegia (G81.1), trigeminal neuralgia (G50.0), or narcolepsy (G47.41). We additionally exclude participants with a history of self-reported stroke, self-reported difficulty walking, those who report use of a walking aid because of physical health problem, and pregnant women. In addition, because our characterization of neurologic status relies on prescription medications used to treat known conditions and may have missed individuals with undiagnosed or non-pharmacologically treated conditions, we also exclude individuals with sNfL levels exceeding an outlier limit for sNfL of >66pg/mL; this limit was derived using a generalized extreme Studentized deviate test.¹⁴ We present a flowchart detailing the exclusion criteria in Figure 1. Sensitivity analyses re-added individuals with sNfL levels exceeding the identified outlier limit to ensure consistency of our findings, as it is possible these individuals did not have overt neurological conditions contributing to their high values.

Figure 1.

Consort diagram describing inclusion abd exclusion of study participants.

Assessment of sNfL levels

We measured sNfL using a novel, high throughput acridinium ester immunoassay developed by Siemens Healthineers on the Atellica Platform. A full description of the assay’s development and validation has been reported elsewhere.¹⁵ Briefly, the light reagent uses a monoclonal anti-NfL antibody labeled with a proprietary acridinium ester (AE) for chemiluminescent detection.¹⁵ Specifically, the NfL Assay is immunometric — using solid-phase magnetic bead capture with one antibody, and direct AE chemiluminometric detection with the other. Siemens Healthineers has a licensing and supply agreement to use Quanterix’s proprietary NfL antibodies, which were acquired originally from Uman Diagnostics. The accumulated light signal relays the NfL concentration in the sample, and calibrators across the range of the assay were used to develop the standard reagent curve. The assay has a range of 3.9 to 500 pg/mL and excellent reproducibility; inter and intra-assay variability estimated from serum samples from 100 people with MS and 100 healthy controls in which the assay was run on separate days and duplicates in the same run was 15.0% and 12.1%, respectively. sNfL levels measured using this novel assay correlate strongly with sNfL levels measured using Single Molecule Array (Simoa; Quanterix).¹⁵ In addition, our previous work found that in a population of 2143 people with MS in which levels were ascertained using both methods, the observed correlation was 0.89.¹⁶ Passing-Bablok regression models also suggest excellent agreement between the two methods and the estimated intra-class correlation coefficient (ICC) was also high (ICC: 0.87; 95% CI: 0.86, 0.88). Importantly, both serum and CSF NfL values estimated using this novel assay were associated with radiological and clinical disease activity measures in people with MS and ALS.^15,16

Statistical analyses

Initial descriptive analyses compared the distribution of sNfL levels across key demographic, clinical, lifestyle, laboratory and comorbidity characteristics.

Contributors to variability in sNfL levels

We next generated a list of 52 potential predictors from the comprehensive information available in NHANES incorporating demographic, lifestyle, anthropometric, laboratory, non-neurologic comorbidity characteristics. We calculated the estimated glomerular filtration rate (eGFR) using the CKD-EPI 2021 equations.¹⁷ We also created total comorbidity burden score as the sum of physical non-neurologic comorbidities affecting an individual that includes: asthma, celiac disease, gout, rheumatoid arthritis, heart failure, coronary artery disease, angina, liver disease, kidney failure, thyroid disorders, hyperlipidemia, diabetes, chronic obstructive pulmonary disease (COPD), diabetes, or cancer. A full list of predictors considered is available in Supplemental Table 1. We did not consider additional demographic predictors beyond age and sex; other factors including race, ethnicity, indicators of socio-economic status were not included as potential predictors, as these factors are largely social rather than biological constructs and could lead to misinterpretation of sNfL values. Prior to fitting all models, we imputed missing values using non-parametric missing value imputation for mixed-type data (e.g., categorical and continuous variables) based on a random forest.¹⁸ For all potential predictors that we considered, rates of missingness were generally <5%. All models also incorporated survey weights to account for the complex sampling scheme used in NHANES.

Overall, our main goals for this analysis were to 1) rank non-neurological predictors in terms of their predictive capacity for sNfL levels and 2) identify the subset of these predictors that would best explain variability in sNfL levels. For our first set of analyses and using 5-fold cross-validation repeated for 100 splits, we evaluated the univariate predictive capacity of each characteristic by calculating R², mean squared predictive error (MSPE) and mean absolute prediction error (MabE) for each predictor. We log-transformed sNfL levels. For continuous variables, each variable was considered as a linear term and using a natural spline with 3 knots. We tested whether the spline term improved model fit using a Wald-test for survey data; spline terms were selected over the linear term if it improved model fit in at least 3 of the 5 folds. For the second set of analyses, we then developed a multivariable prediction model using a stepwise procedure in which we added predictors sequentially to the model by adding the predictor which resulted in the smallest prediction error for the associated model. We applied two sets of stopping criteria in which we stopped adding predictors once the improvement in cross-validated R² did not exceed 0.01 or 1% (primary analyses) and 0.005 or 0.5% (secondary analyses). We selected these thresholds in order to maximize the likelihood that the resultant model would be parsimonious and therefore more readily translatable in helping to improve interpretation of sNfL levels in a clinical care setting. In sensitivity analyses we also considered stopping criteria of 2%. In complementary analyses and to describe directionality of the associations, we also calculated the relative effect size (and 95% CI) of each continuous predictor on sNfL levels by calculating a standardized parameter estimate (i.e., a standardized beta coefficient) in which we multiplied each parameter estimate by the ratio of the population standard deviation (SD) for the individual predictor to the population SD in sNfL. For binary covariates, we estimated the effect size using the standardized mean difference (i.e., Cohen’s d). We calculated standardized coefficients for all univariate models as well as our multivariable model.

Derivation of adjusted sNfL reference ranges for US adults without neurologic disease

Next, using the results from our initial multivariable model, we derived adjusted reference ranges for sNfL levels using generalized additive models for location, scale and shape (GAMLSS). Briefly, GAMLSS are an extension of generalized linear and generalized additive mixed models that instead of assuming and an underlying exponential family for the outcome (i.e., sNfL for our purposes), a more flexible approach is adopted as a general distribution family is assumed only. This approach has been previously applied to derive age-adjusted reference ranges of sNfL levels in smaller studies. Other large, international groups (e.g., the Global Lung Function Initiative)^19–21 have employed similar modeling strategies to derive demographic-adjusted respiratory function outcomes that are widely applied. We selected the Box-Cox Cole and Green distribution family as this family minimized the Generalized Akaike information criterion. We then used this model to calculate centile curves (for the 95^th and 97.5^th percentile) and corresponding Z scores that account for the selected covariates. As above, analyses incorporated survey weights to account for the complex sampling procedure.^19–21

Data availability

All NHANES data is managed by the Centers for Disease Control and are publicly available.

Results

Descriptive analyses

Participant characteristics are provided in Table 1 by quartile of sNfL levels. The average age of the study sample was 43.6 years (SD: 14.8) and participants were 50.6% male, and 35.4% non-white. On average, 57.3% of participants had at least one non-neurologic comorbidity. As expected, there was a strong positive and potentially non-linear association between age and higher sNfL levels (Figure 2).

Table 1.

Characteristics of NHANES participants (n=1706) by sNfL levels for selected characteristics.

Characteristics	Overall	Quartile of sNfL (pg/mL)
		Q1	Q2	Q3	Q4
sNfL, pg/mL, mean (SD)	13.9 (9.1)	5.8 (1.3)	9.6 (1.1)	14.0 (1.6)	26.6 (8.8)
Age, years, mean (SD)	43.6 (14.9)	33.6 (9.8)	40.3 (12.5)	47.8 (14.2)	53.3 (14.4)
Female sex, %*	49.4%	56.6%	48.2%	47.6%	45.2%
Race/ethnicity, %*
White	64.6%	53.2%	58.9%	73.0%	74.0%
Black	11.6%	13.2%	15.4%	7.7%	10.1%
Asian	5.7%	6.0%	6.7%	6.2%	4.1%
Mexican American	9.8%	17.3%	9.5%	5.6%	6.5%
Other Hispanic/Latino	6.0%	7.1%	6.7%	6.5%	3.7%
Other races	2.2%	3.2%	2.8%	1.0%	1.6%
Poverty index, mean (SD)	2.8 (1.7)	2.5 (1.6)	2.7 (1.6)	3.1 (1.8)	3.0 (1.6)
College education or more, %	32.7%	29.8%	32.6%	39.1%	29.6%
Have health insurance, %	78.6%	70.8%	77.2%	81.2%	85.6%
Alcoholic drinks/day, mean (SD)	2.1 (2.4)	2.3 (2.5)	2.3 (2.5)	2.0 (2.2)	2.0 (2.4)
Ever used crack, cocaine, heroin, methamphetamine, %	15.1%	11.2%	16.7%	18.2%	14.4%
Cigarettes per day, mean (SD)	2.1 (5.7)	1.6 (4.6)	2.2 (5.9)	2.4 (5.9)	2.3 (6.4)
Vigorous physical activity, min/week, mean (SD)	21.5 (44.4)	20.1 (36.6)	27.2 (47.9)	21.2 (45.0)	17.8 (47.2)
Health care visits in past year, mean (SD)	4.4 (3.9)	3.8 (3.6)	4.4 (4.2)	4.4 (3.7)	4.8 (3.9)
Hospital stay in past year, %	8.6%	9.9%	9.1%	7.2%	8.2%
Weight, kg, mean (SD)	83.0 (21.7)	85.2 (23.8)	81.9 (19.9)	81.6 (21.1)	83.3 (21.5)
Height, cm, mean (SD)	169.3 (9.8)	168.7 (9.2)	169.9 (10.3)	170.1 (9.6)	168.6 (10.2)
BMI, kg/m2, mean (SD)	28.9 (6.9)	29.9 (7.5)	28.3 (6.5)	28.1 (6.5)	29.2 (7.0)
Total comorbidity burden*, mean (SD)	1.0 (1.2)	0.6 (0.8)	0.9 (1.1)	1.1 (1.2)	1.5 (1.4)
Liver disease, %	1.5%	1.9%	1.2%	2.0%	1.0%
Hypertension, %	27.5%	14.6%	26.1%	31.4%	38.7%
Hyperlipidemia, %	30.6%	19.8%	23.8%	35.9%	43.5%
Any history of head trauma, %	9.6%	3.4%	8.1%	13.6%	13.5%
Cancer, %	8.0%	2.6%	7.6%	9.0%	12.9%
Gout, %	2.3%	0.2%	2.5%	2.3%	4.2%
Rheumatoid arthritis, celiac disease, %	2.2%	0.9%	2.3%	2.1%	3.5%
Diabetes, %	8.6%	3.2%	4.6%	11.0%	15.9%
Kidney failure, %	1.2%	0.8%	0.1%	1.4%	2.6%
Any cardiovascular disease, %	2.7%	0.6%	1.2%	2.2%	6.8%
Chronic obstructive pulmonary disease, %	1.5%	0.2%	0.3%	1.3%	4.3%

^*Percentages incorporate sampling weights; Displayed values are estimated percentages of the US non-institutionalized population. Total comorbidity burden is the sum of physical non-neurologic comorbidities affecting an individual (asthma, celiac disease, gout, rheumatoid arthritis, heart failure, coronary artery disease, angina, liver disease, kidney failure, thyroid disorders, hyperlipidemia, diabetes, COPD, diabetes, or cancer.

Figure 2.

Association between age and sNfL levels in 1706 eligible NHANES participants. The fitted blue line denotes a smoothed curve fit estimated using local regression. Scatterplot incorporates survey weights which are depicted by the relative size of the data point; larger datapoints denote individuals with larger weights.

Assessment of contributors to variability in sNfL levels

Figure 3A depicts the ranking of a given predictor according to R² in univariate regression models in which each sNfL prediction model was fit with one covariate at a time. Age was the strongest predictor of sNfL levels (R² = 26.8%) followed by eGFR and total comorbidity burden, explaining 18.9% and 15.2% of the variability in sNfL levels, respectively. Beyond age, when considering the effect size of individual predictors, other non-neurologic comorbidities tended to have strong associations with higher sNfL levels (e.g., COPD, any CVD, diabetes, kidney disease; Figure 3B). Increasing eGFR, estradiol, and albumin had the largest significant negative effect on sNfL levels.

Figure 3.

Univariate associations between candidate predictors and sNfL levels. A. Cross-validated R² between sNfL levels versus a given predictor that are obtained from separated univariate models. Colors denote the classification of predictor type. B. Effect size estimates (95% CI) between a given predictor and sNfL levels. To compare relative associations across predictors, parameter estimates for each continuous predictor are standardized by the ratio of the population standard deviations (SD) of the predictor in question divided by the population SD for sNfL. For binary predictors, the presented effect size denotes the standardized mean difference.

We also considered a multi-predictor model and used forward selection, adding one variable at a time by maximizing the cross-validated R². Figure 4 displays the R² at each stage of the forward selection procedure; full results including the MSPE and MabE are provided in Supplemental Table 1. After applying our pre-specified stopping criterion of observing a less than 1% increase in the cross-validated R², the final sNfL prediction model contained (in order of their selection) age and creatinine. When applying a stopping criterion of 0.5% increase in the cross-validated R² in secondary analyses, the sNfL prediction model contained (in order of their selection) age, creatinine and hbA1c. We obtained relatively similar models when using a more stringent cut-off of 2% (age and creatinine). For the individual covariates selected, age, creatinine, and HbA1c were positively associated with sNfL levels (standardized β – age: 0.46, 95% CI: 0.43, 0.50; creatinine: 0.18, 95% CI: 0.13, 0.22; HbA1c: 0.09, 95% CI: 0.06, 0.11). BMI or other measures of body composition did not explain >1% of the variability in sNfL levels, and the association between BMI and sNfL was most notable for underweight individuals in NHANES. Relative to normal weight individuals, underweight participants had a 27.42% (95% CI: 9.79%, 47.89%) increase in sNfL levels relative to normal weight individuals. It should be noted that only 28 individuals in NHANES were classified as underweight, and a BMI cut-point of 18.5 represents approximately the 1^st percentile of the distribution of BMI in NHANES. Results were generally similar if we refit models including individuals with sNfL levels exceeding the identified outlier limit (Supplemental Figure 1).

Figure 4.

A. Cross-validated R² at each stage of the forward selection. The dotted vertical denotes time at which stopping criteria was met (e.g., the next additional predictor did not increase the R² by at least 1% [blue; dot-dashed]). Colors denote the classification of the predictor type. B. Standardized predictor estimates for covariates from final multivariable model derived from selection procedure that including age, creatinine, and HbA1c

Derivation of adjusted reference ranges of sNfL

Using the results of our initial set of analysis, we fit a GAMLSS model incorporating age, creatinine and HbA1c, and derived corresponding Z scores. Whereas in Figure 2, there was a strong association between age and sNfL, Figure 5a displays the relationship between age and the derived age, creatinine, and HbA1c-adjusted Z scores in which no notable association is observed when using the Z scores. In Figure 5b and 5c, we present estimated centile curves corresponding to the 95^th centiles across continuous levels of age in which varying values of the other sNfL-relevant parameters considered are plotted. For example, in 95^th centile curves, the distribution of expected sNfL scores across ages is higher in those with kidney disease versus normal individuals (Figure 5b) and higher in diabetics versus normal individuals (Figure 5c). As proof of concept, we also estimated adjusted sNfL Z scores for participants originally excluded for possibly having existing neurologic disease (n=324); 10.2% of these individuals were identified as having elevated sNfL (i.e., sNfL > 97.5^th percentile or Z scores > 1.96). Average Z scores in this population were also on average 0.46 SD higher (95% CI: 0.26, 0.67; p=0.0006) when compared to Z scores from the general population without neurologic disease.

Figure 5.

Results of GAMLSS model for sNfl that includes age, HbA1c, and creatinine. Values displayed denoted the 95th percentile across ages for sample combinations of identified relevant contributors to sNfL. A. Age, creatinine, and HbA1c-adjusted Z scores across ages demonstrating no association between age and adjusted sNfl Z scores (in contrast to Figure 2) The dotted line denotes Z scores of 1.64. B. Centle curves across ages comparing individuals with kidney disease (e.g., eGFR < 60; mg/dL) versus those without kidney disease (e.g., eGFR≥60; median creatinline = 0.85mg/dL). C. Centile curves across ages comparing diabetics (e.g., HbA1c>6.4%) versus non-diabetics (e.g., HbA1c<5.4%).

Application development

To help facilitate translation of the results of our study for and to help other investigators apply the adjusted centile curves to their own data, we developed an R Shiny application available at [https://brbdai-kathryn-fitzgerald.shinyapps.io/sNfL_dashboard/]. The application includes three primary functions. First, for a single patient, data on sNfL and relevant clinical characteristics (age, creatinine, and HbA1c) can be entered. The application will then estimate the expected percent difference in sNfL between the new patient and a reference participant from our study accounting for these values. The second functionality allows one to compare expected percent differences in sNfL levels across a range of relevant characteristics. For both options, the 95% CI for the relative percent difference in sNfL levels is estimated using bootstrapping. The application’s final feature allows one to upload data on sNfL and relevant characteristics for a set of new individuals and will calculate covariate adjusted Z-scores and the corresponding upper limit of normal sNfL levels (which we set as the 95^th or 97.5^th percentile) for a given set of characteristics. Also, in order to maximize flexibility, we allow for Z scores to be calculated incorporating age, age and creatinine, or age, creatinine, and HbA1c.

Discussion

In this study, we examined contributors to variability in sNfL levels in a large, representative cohort of US adults. We considered over 50 potential predictors and performed detailed analyses in which we assessed individual, relative, and combined predictive performance of each characteristic as it relates to sNfL levels. Age, creatinine, and HbA1c were strongly positively predictive of elevated sNfL levels. These results strongly support the utility of considering additional characteristics beyond age in the interpretation of sNfL levels. Lastly, we applied these findings to derive adjusted reference ranges and provide a tool that can help translate our results and aid future research endeavors focusing on sNfL in the context of neurologic diseases.

An accumulating body of evidence has established a strong positive and potentially non-linear association between age and both blood and CSF NfL levels. One cross sectional study of 359 individuals found a log-linear association between age and CSF NfL in which a 1-year greater in age was associated with an approximately 3% increase in CSF NfL.²² Similar findings have been noted for blood NfL levels, with a 2.2% annual increase in sNfL levels noted in a cohort of 254 healthy individuals; this finding is similar to those described here, in which we note a 2.1% increase in sNfL associated with a 1-year increase in age. In addition, the strong relationship between age, sNfL, and overall brain morphology has been explored in a study of 335 healthy participants living in Austria who completed a detailed clinical examination, cognitive testing and 3T brain MRI.¹¹ Results implicate subclinical brain damage in individuals with higher sNfL levels; sNfL levels were linked with brain volume changes in cross-sectional and longitudinal analyses. The study also noted an increase in variability of sNfL levels among individuals older than 60 years, which the authors hypothesized could be related to subclinical comorbid pathology.

While prior studies of NfL have largely focused on understanding the association between aging and NfL levels, our findings highlight the importance of also considering markers of renal function and metabolic health. Both serum creatinine and HbA1c were strongly associated with increasing sNfL levels, even after accounting for age. Observations of a positive association between markers of renal function and sNfL in people without neurologic disease is consistent with prior studies in which blood levels of NfL were strongly correlated with creatinine and eGFR.^12,13,23 Another large cohort of elderly Swiss adults with atrial fibrillation found that even after accounting for age and MRI characteristics, eGFR explained a non-trivial proportion of the variability in sNfL.²⁴ Combined with our findings, these collective results suggest that kidney function may be critical for the clearance of NfL and should be taken into consideration in interpreting NfL measures in the broader context. Furthermore, some studies suggest eGFR may underestimate true GFR in people with markedly low muscle mass, which is common in many degenerative diseases and causes reduction in creatinine.²⁵ For HbA1c, the prior study of healthy Austrian adults evaluated possible contributions of vascular risk factors, but differences in the underlying prevalence of metabolic conditions may have limited their capacity to observe a strong effect. For example, the prevalence of diabetes in the Austrian cohort of individuals aged on average 65 years was approximately 8%, whereas in the US estimates of prevalence of diabetes in individuals aged >60 years is upwards of 20%.^11,26 Other studies based the US found associations between HbA1c and sNfL in a cohort of diabetics²⁷ as well as another study finding an association between higher plasma NfL and diabetes in a population of community dwelling older adults focusing on Alzheimer’s disease.²⁸ The specific mechanisms by which HbA1c is associated with higher sNfL levels is not entirely clear, but we speculate that this may relate to microvascular disease complications including diabetic neuropathy, retinopathy and cerebrovascular disease, leading to NfL release. Lastly, while we note a potentially inverse association between BMI and sNfL, which is consistent with observations from large European cohorts of healthy individuals,^12,23,30 BMI was not a primary predictor of variability in sNfL beyond underweight individuals in NHANES. The underlying distribution of BMI may differ considerably from the average American participant in NHANES. The average BMI in NHANES (28.8 kg/m²; 68.8% overweight or obese) is quite a bit higher than the average BMI observed in European countries.^23,29 The prevalence of weight-related comorbidities like diabetes (and related biomarkers like HbA1c) is also higher in US populations, which together with the different distributions of BMI could contribute to this discrepancy. Longitudinally, one study also noted a faster rate of plasma NfL increases for individuals with high BMI and suggested this result was mediated by increasing cardiovascular risk factors.³¹

Our study has several notable limitations. Namely, the physical exam included in the NHANES assessment does not include any type of neurologic exam performed by a neurologist, and we characterized neurologic status for several diseases using prescription medications used to treat known conditions or by self-report for some diagnoses. This approach may have missed participants with undiagnosed or non-pharmacologically treated conditions; we attempted to mitigate this bias by also excluding 1) outlying values with a standardized statistical approach and 2) those with ambulatory disability due to a physical health problem. Furthermore, these results also support the notion that within person changes in sNfL levels may be likely to provide the more insightful information in interpreting sNfL levels rather than a single measurement. Other studies have noted sNfL to be a dynamic biomarker that may vary over time. By design, in this study, we could not evaluate the association between changes in patient characteristics, laboratory values or comorbidity prevalence as they relate to changes in sNfL levels.

Strengths of our study include its large size and depth of predictors considered, ranging from demographics, anthropometric and lifestyle characteristics to comorbidity. NHANES is also designed to be a representative sample of the non-institutionalized US individuals; previous studies of sNfL in the context of neurologic diseases often included cohorts of healthy individuals who were recruited from tertiary care centers and may not be representative of the average American. We also assess the predictive performance of a set of widely available and commonly ordered laboratory tests to characterize their potential relationship with sNfL in an unbiased setting (i.e., the tests were available for all cohort members and were not derived from routine clinical care in which interpretation may be vulnerable to indication biases). We also performed a comprehensive set of analyses incorporating both the impact of individual predictors such as age, as in previous analyses, but also the impact of other important predictors in a multivariable context.

Ultimately, the development and validation of blood-based biomarkers for neurological disorders will represent a major advance in the management and care of patients with these disorders. Our findings represent an important step in refining the understanding of how to account for age and other features that impact sNfL levels so as to move towards more certainty in interpreting test results as they relate to the condition of interest. Combined with ongoing work in target populations to better elucidate test characteristics and the clinical importance of changes in levels, these results will ensure that, when deployed in the clinical setting, sNfL levels can be more confidently interpreted as representing a measure of the intended underlying disease process.

Summary for Social Media.

1. Twitter handles: @katefitzg; @ESotirchos, @Calabresi

2. What is the current knowledge on this topic?

   Serum neurofilament light chain (sNfL) is an emerging prognostic and monitoring biomarker relevant to several neurological disorders. sNfL levels are strongly correlated with age; however, limited research has evaluated whether demographic, clinical, or lifestyle factors beyond age impact sNfL levels.

3. What question did this study address?

   We evaluated >50 potential contributors to variability in sNfL levels using a large, well-characterized and representative cohort of US adults without neurologic disease. We evaluate their individual and combined contribution to sNfL variability and derive adjusted reference ranges that allow for a more accurate interpretation of an individual’s sNfL level.

4. What does this study add to our knowledge?

   Age, creatinine, and HbA1c were strongly associated with higher sNfL levels.

5. How might this potentially impact on the practice of neurology?

   Results support the utility of considering additional characteristics beyond age in the interpretation of sNfL levels. We also derive adjusted reference ranges and provide a resource to 1) help in the translation of our results and 2) aid future studies evaluating sNfL in neurologic diseases.