Factors influencing serum neurofilament light chain levels in normal aging

Study cohort

The entire cohort consisted of 327 neurologically inconspicuous community-dwelling participants. Table 1 displays demographical and life-style related factors, renal and liver function, cerebrovascular risk factors and blood volume for the entire cohort and in age-stratified subsamples <60 years and ≥60 years.

Identification of factors influencing sNfL

When considering the entire cohort we found significant correlations of sNfL with age (rho=0.732, p<0.001), renal function (estimated glomerular filtration rate (eGFR) (rho=-0.523, p<0.001), creatinine (rho=0.152, p=0.006) and urea (rho=0.294, p<0.001)), liver function (glutamate pyruvate transaminase (GPT) (rho=-0.171, p=0.002)), cerebrovascular risk factors (systolic blood pressure (rho=0.254, p<0.001), HDL (rho=0.130, p=0.019) and Hba1c (rho=0.255, p<0.001)) and blood volume (rho=-0.240, p<0.001). Although, overall BMI was unrelated to sNfL in the total cohort, we found more indication of a correlation between sNfL and BMI in younger individuals (<60 years: rho=-0.283, p=0.005) compared participants aged 60 years and older (≥60 years: rho=-0.148, p=0.02).

Analyses on group differences revealed higher sNfL concentrations in hypertensive individuals compared to normo-tensive individuals (U=8455.0, p<0.001, sNfL-mean_hypertension=37.0 (SD=16.2) vs. sNfL-mean_{no-hypertension}=30.2 (SD=14.9)), individuals with hypercholesterolemia (U=4700.5, p<0.001, sNfL-mean_hyperchol=43.3 (SD=16.1) vs. sNfL-mean_no-hyperchol=33.3 (SD=15.6)) and diabetic individuals (U=2678.5, p=0.026, sNfL-mean_diabetic=42.8 (SD=18.7) vs. sNfL-mean_non-diabetic=34.3 (SD=15.7)).

The above-mentioned variables significantly associated with sNfL were used for further analyses.

Conditional importance of factors influencing sNfL

We then performed random forest regressions to determine the conditional importance of each factor influencing sNfL.

When considering the entire cohort, random forest regression revealed that among all examined factors age (mean=34.5, CI: 32.9, 36.3) had by far the highest conditional variable importance influencing sNfL, which was followed by eGFR (mean=2.9, CI: 2.4, 3.4) and creatinine (mean=1.9, CI: 1.5, 2.3) (Figure 1A).

Figure 1. Random forest regressions of determinants associated with sNfL concentration in the total sample (A), individuals ≥60 years (B) and <60 years (C).

Similar results were found in a subgroup analysis when analyzing individuals ≥60 years. In this subgroup age (mean=30.1, CI: 28.5, 31.7) had the highest conditional importance which was followed by creatinine (mean=6.2, CI: 5.4, 7.1), eGFR (mean=6.0, CI: 5.1, 7.0) and HDL (mean=3.7, CI: 3.0, 4.3) (Figure 1B).

In contrast, when analyzing in individuals <60 years we found that BMI (mean=3.0, CI: 2.7, 3.4), blood volume (mean=1.4, CI: 1.1, 1.7) and eGFR (mean=0.7, CI: 0.5, 0.9) had the highest conditional importance, whereas the impact of age (mean=0.6, CI: 0.4, 0.9) was lower (Figure 1C).

Significant association of identified factors influencing sNfL

Multiple regression analyses then served to determine statistically significant associations of factors with highest conditional importance from random forest analyses influencing sNfL.

Considering all individuals, multiple regression analysis revealed that age (β=0.513, 95% CI: 0.631, 0.919, p<0.001) was the most important factor influencing sNfL, which was followed by eGFR (β=-0.207; 95% CI: -0.355, -0.1; p<0.001), BMI (β=-0.165; 95% CI: -0.841, -0.294; p=0.001) and urea (β=0.113; 95% CI: 0.005, 0.322; p=0.019). Overall, this model explained 49.5% of the variance (Table 2).

In individuals of above 60 years regression analysis revealed that age (β=0.394; 95% CI: 0.868, 1.632; p<0.001), creatinine (β=0.376; 95% CI: 13.912, 35.886; p<0.001), blood volume (β=-0.198; 95% CI: -6.797, -1.048; p=0.008) and HDL (β=0.149; 95% CI: 0.01, 0.215; p=0.013) were significantly associated with sNfL. This model explained 37.9% of the variance. BMI (p=0.246) and urea (p=0.435) were unrelated to sNfL in this analysis (for summary see Figure 2).

Figure 2. Summary of the most important variables affecting the sNfL concentration.

In individuals below 60 years BMI (β=-0.298; 95% CI: -0.719, -0.144; p=0.013) was the only determinant significantly associated with sNfL. In this model, blood volume (p=0.600), eGFR (p=0.126) and age (p=0.222) were no longer significantly associated with sNFL. Overall, the model explained 17.0% of the variance.

Ethics

The study protocol was approved by the ethics committee of the Medical University of Graz, Austria, and written informed consent was obtained from all participants.

Participants

In this study we analyzed neurologically inconspicuous individuals who participated in the Austrian Stroke Prevention Family Study (ASPS-Fam), an extension of the Austrian Stroke Prevention Study (ASPS) [15, 16]. The ASPS is a prospective single-centre community-based study established in 1991 on the cerebral effects of vascular risk factors in the normal aged population of the city of Graz, Austria. For the ASPS-Fam a total of 419 individuals were included. Between 2006 and 2013 study participants underwent thorough clinical history taking and neurological examination, magnetic resonance imaging (MRI), laboratory evaluation and cognitive testing accompanied by a comprehensive vascular risk factor assessment. Inclusion criteria were no history of previous stroke or dementia and a normal neurologic examination. None of the subjects had suffered a traumatic brain injury. From the 419 individuals included in the ASPS-Fam, blood sampling was available in 371 subjects. Another 44 subjects were excluded due to diagnosis or suspicion of dementia (Mini Mental Status Examination (MMSE) ≤24, n=11), visible brain infarcts on Magnetic Resonance Imaging (MRI) (n=19), a history of stroke (n=9), other diseases (chronic myeloid leukemia) (n=1), missing laboratory data (n=4) or a combination of the above (n=3), leaving 327 subjects to examine confounding factors of sNfL alteration in a normal aging population.

Demographic, life-style and clinical variables

Overall, the effect of 24 determinants potentially affecting sNfL was assessed. We included demographical factors (age, sex) and life-style related factors (waist-hip-ratio (WHR), body mass index (BMI), alcohol consumption). Laboratory variables of renal and liver function included the eGFR, creatinine, urea, uric acid, and GPT, glutamic oxaloacetic transaminase (GOT), gamma-glutamyl transferase (GGT), alkaline phosphatase (AP). Estimated GFR was calculated as suggested by the National Kidney Foundation estimating GFR [17] from serum creatinine, age, sex and race using the formula eGFR = 141 x min(SCr/κ, 1)α x max(SCr /κ, 1)-1.209 x 0.993Age x 1.018 [if female] x 1.159 [if Black] with SCr being standardized serum creatinine (mg/dL), κ = 0.7 for females and 0.9 for males, α = -0.329 for females and -0.411 for males, min indicates the minimum of SCr/κ or 1, max indicates the maximum of SCr/κ or 1 and age in years. Further, cerebrovascular risk factors (hypertension, hypercholesterolemia, diabetes, mean systolic and diastolic blood pressure, cholesterol, low- (HDL) and high-density lipid protein (LDL), triglycerides, glycated hemoglobin (HBA1C)) and blood volume were included. Blood volume was calculated separately for males and females based on weight and height using the formula by Nadler: blood volume = 0.3669 * height³ + 0.03219 * weight + 0.6041 for males and blood volume = 0.3561 * height³ + 0.03308 * weight + 0.1833 for females as in [18].

Definition of categorical variables

Alcohol consumption was defined as daily uptake exceeding half a liter beer, a glass of wine or 1cl of high-proof alcohol. Hypertension was defined as use of antihypertensive therapy or systolic blood pressure of 140mmHg or above or diastolic blood pressure of 90mmHg or above [19]. Hypercholesterolemia was considered if the subject was on therapy. Diabetes was considered if the subject either was on therapy, blood sugar level exceeded 126 mg/dl or a positive medical history [19].

NfL measurement

Serum samples of all participants were collected by venipuncture and processed on the same day of examination between 2006 and 2013 within 2h. After venipuncture, blood tubes remained at room temperature for 30 min and were then centrifuged at room temperature for 10 min at 2000 x g. Serum was then aliquoted in polypropylene tubes and stored at -70° C until analysis of NfL.

All serum samples were analyzed at the University Hospital Basel, Switzerland as previously described [5]. In brief, serum NfL levels were determined by single molecule array (Simoa) assay using the capture monoclonal antibody (mAB) 47:3 (initial dilution 0.3 mg/mL; Art. No. 27016) and the biotinylated detector mAB 2:1 (0.1 μg/mL; Art. No. 27018) [20]. The mean coefficients of variation (CVs) of duplicate determinations for concentration were 8.5% (9.5 pg mL^-1, sample 1), 5.4% (23.2 pg mL^-1, sample 2) and 7.8% (98.5 pg mL^-1, sample 3). Interassay CVs for serum were 7.8% (sample 1), 8.3% (sample 2) and 4.9% (sample 3) [5].

Statistical analyses

Statistical analysis was performed using SPSS (version 26, SPSS Inc., Chicago, IL, USA) and the statistical software R (version 4.0.2; R: a language and environment for statistical computing, Vienna, Austria). A two-sided p-value <0.05 was considered to be statistically significant. Explorative data analysis with extreme scores on the upper and lower ends of the score distributions were used for the identification of outliers. A score was considered an outlier if it exceeded 1.5 x interquartile range (IQR) of a variable.

The entire cohort was also stratified by age in the age groups <60 years (age range: 38.5-59.9, n=99) and 60 years and above (age range: 60.1-84.3, n=228) equivalent to [5]. This stratification was chosen due to the non-linear relationship between sNfL and age showing increasing sNfL concentrations above the age of 60 potentially indicating a subclinical pathological process.

Spearman correlations were used to identify determinants significantly correlated with sNfL in either the entire cohort or the stratified samples. With respect to categorical variables, Mann-Whitney-U-Tests were used to identify differences in the sNfL level. Significant variables were used in further analyses. Variables with non-significant association with sNfL were excluded from further analyses.

To determine the conditional importance of factors influencing sNfL, we calculated random forest regressions. Random forest assesses the explanatory power of each individual variable while at the same time accounting for all other variables. Using the R package ‘party’ (version 1.3-5 [21]) 1001 inference trees with unbiased variable selection were calculated using 5 randomly selected variables for each split (unbiased resampling scheme). From these trees conditional permutation importance was calculated, using the “mean decrease in accuracy” principle as importance measure for each variable together with a 95% confidence interval from 400 repetitions. This approach was chosen (1) to account for the intercorrelation between predictor variables (multicollinearity), (2) to uncover variables that are non-linearly related to sNfL and, ultimately, (3) to identify the variables with highest conditional variable importance. Random forest regressions were calculated in the entire cohort and in the age-stratified samples.

Determinants for the sNfL level having the lower limit of the confidence interval >0 were then used in multiple linear regressions including bootstrapping with replacement (10.000 samples, bias corrected and accelerated (BCa-method), 95% confidence interval) to determine significant associations with sNfL. In case of correlation of variables, one was excluded from the multiple regression analysis due to multicollinearity (variance inflation factor, VIF>5). Bootstrapping was chosen for deriving more robust estimates of the standard error and confidence intervals. Lower and upper 95% confidence intervals from the bootstrapped coefficients were estimated to determine significance. If the 95% confidence interval (CI) of the indirect effect does not contain 0, a significant effect is probable.

Data availability

Data sets generated and/or analyzed during the study are available from the corresponding author upon reasonable request.