The National Health and Nutrition Examination Survey (NHANES) is a population-based comprehensive study designed to assess the health and nutritional status of children and adults. NHANES health examinations are conducted by medical professionals at Mobile Examination Centers (MECs), focusing on collecting baseline physical examinations, biochemical tests, and other medical information. All data are publicly available through the official website.¹ The study protocol has been approved by the National Center for Health Statistics (NCHS) Research Ethics Review Committee, and all participants have provided informed consent.

This study analyzed NHANES data from 2011 to 2014. We excluded participants lacking NPS assessment data (serum albumin, TC, and complete blood count) (n = 194); those with extreme energy intake (men: >4,200 kcal/day or <800 kcal/day; women: > 3,500 kcal/day or < 500 kcal/day) (n = 57); and those with missing covariate information (n = 277). After these exclusions, a total of 2,595 participants were included in the final analysis (Figure 1). Ethical approval was granted by the NCHS Research Ethics Review Board (ERB), and all participants provided informed consent in accordance with ethical guidelines.

The NPS score was calculated according to the method established by Galizia et al., based on four biomarkers: serum albumin, total cholesterol (TC), NLR, and lymphocyte-to-monocyte ratio (LMR). Only total cholesterol was available in the NHANES dataset for this analysis; lipid subfractions were not considered. The optimal cutoff values for NLR and LMR were determined using MaxStat analysis. Each component scored one point if it met the following abnormal criteria: serum albumin <40 g/L, TC ≤ 180 mg/dL, NLR ≥ 2.96, or LMR ≤ 4.44; otherwise, it scored zero points. The total NPS was the sum of points from all four parameters, ranging from 0 to 4. Based on prior studies, participants were categorized into three groups: Group 0 (score = 0), Group 1 (score = 1–2), and Group 2 (score = 3–4).

Cognitive function was assessed in participants aged 60 years and older using three standardized neuropsychological tests. The Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) test evaluates immediate and delayed recall of verbal information through three consecutive learning trials and one delayed recall trial, with total scores ranging from 0 to 40. The Animal Fluency Test (CFDAST) measures semantic memory and verbal fluency by asking participants to name as many animals as possible within one minute; the score is the total number of correct responses (13). The Digit Symbol Substitution Test (CFDDS) assesses processing speed, visuomotor coordination, sustained attention, and working memory (14, 15). Participants were required to match symbols to numbers using a reference key within 120 s, and the total score was determined by the number of correctly completed items, up to a maximum of 133 points. Higher scores on all tests indicate better cognitive performance.

In the absence of universally accepted diagnostic criteria for these tests, we operationalized cognitive impairment as a score in the lowest quartile of our sample distribution for each test (16). It should be noted that this constitutes a relative, sample-specific threshold for poorer cognitive performance, not a clinical cutoff for diagnosing dementia or cognitive impairment.

Baseline covariates included demographics, lifestyle factors, and clinical conditions. Demographic variables comprised age (years), gender (male/female), education level (<high school, high school, >high school), race/ethnicity (Mexican American, other Hispanic, non-Hispanic White, non-Hispanic Black, other), and body mass index (BMI; <25.0, 25.0–29.9, ≥30.0 kg/m²). Smoking status was categorized as never (<100 cigarettes in lifetime), former (≥100 cigarettes but quit), or current (≥100 cigarettes and currently smoking). Alcohol consumption was classified as non-drinker, low-to-moderate (men <28 g/day, women <14 g/day), or heavy (men ≥28 g/day, women ≥14 g/day). Daily intake of caffeine, energy, total fat, and total sugar was averaged from two 24 h dietary recall interviews. Hypertension was defined as systolic blood pressure ≥140 mmHg, diastolic blood pressure ≥90 mmHg, self-reported diagnosis, or use of antihypertensive medication. Diabetes was defined as self-reported diagnosis, use of insulin or oral hypoglycemic agents, HbA1c ≥ 6.5%, or fasting plasma glucose ≥126 mg/dL.

All analyses were performed in R (version 4.4.1). Considering the complex sampling design of NHANES, all analyses applied survey weights to ensure the national representativeness of the results. For the combined 2011–2014 cycles, the original 2-year examination weights (WTMEC2YR) were halved to create adjusted weights.

Continuous variables are reported as weighted mean (standard error, SE), and categorical variables as unweighted count (weighted percentage). Between-group differences were assessed using analysis of variance (ANOVA) for continuous variables and the Rao–Scott adjusted chi-square test for categorical variables.

Cognitive outcomes were assessed using three standardized tests: CERAD, CFDAST, and CFDDS. Given the exploratory nature of this first investigation of NPS in cognitive aging, we considered all three cognitive domains as equally important exploratory outcomes. Cognitive impairment was defined as scoring in the lowest quartile of each test, creating a binary outcome for logistic regression. To address multiple comparisons, we report unadjusted p values and indicate results that remain significant after a conservative Bonferroni correction for three tests (adjusted α = 0.0167).

A multivariate weighted logistic regression model was used to examine the association between NPS and cognitive function, constructing three progressively adjusted models: a crude model without covariate adjustments; Model 1 adjusted for age, gender, and race/ethnicity; Model 2 further adjusted for education level, smoking status, alcohol consumption, caffeine intake, BMI category, hypertension, diabetes, and energy intake. Results are presented as odds ratios (OR) with 95% confidence interval (CI). Trend analysis was performed by modeling NPS as a continuous variable.

To evaluate potential effect modification, interaction terms between NPS and each stratification variable (gender and age groups [60–75 vs. >75 years]) were included in the models; corresponding p values are reported.

Sensitivity analyses were conducted to assess robustness: (1) treating cognitive test scores as continuous outcomes in survey-weighted linear regression models, and (2) redefining cognitive impairment using a stricter cutoff (lowest decile) in logistic regression models.

A two-sided p value < 0.05 defined statistical significance. Analyses employed the following packages: survey, tidyverse, broom, and gt.

To validate findings from the NHANES population, patients from the Department of Neurology at Huzhou Central Hospital were included. This included 87 patients with clinically diagnosed Alzheimer’s dementia (AD) and 102 patients with vascular dementia (VaD). Patients with severe hepatic/renal dysfunction, critical infections, or other major neurological disorders were excluded. Fasting blood samples were obtained to calculate the NPS. This study was approved by the Ethics Committee of Huzhou Central Hospital, and the requirement for informed consent was waived. The Mann–Whitney U test was used to compare NPS values between the two groups, and the common language effect size (CLES) was calculated to assess the magnitude of the effect. Data visualization was performed using the ggplot2 package in R to generate boxplots illustrating the distribution of NPS scores in both groups.

This study included a final sample of 2,595 participants from the NHANES 2011–2014 dataset. Table 1 compares baseline characteristics by NPS group, with categorical variables presented as weighted percentages and continuous variables as weighted means (standard errors).

Among the participants, 341 (13.1%) scored NPS 0, 1,776 (68.4%) scored NPS 1–2, and 478 (18.4%) scored NPS 3–4. Compared with the NPS 0 group, participants in the NPS 3–4 group were older (71.44 vs. 67.45 years), had a higher proportion of males (56.9% vs. 20.9%), and included a greater representation of Non-Hispanic White participants (81.0% vs. 71.4%). This group also exhibited a significantly higher prevalence of hypertension (75.5% vs. 55.0%) and diabetes (37.0% vs. 16.4%). Inflammatory biomarkers showed notable differences: the NLR was significantly higher (3.96 vs. 1.58), while the lymphocyte-to-monocyte ratio was markedly lower (2.48 vs. 5.85). Regarding cognitive function, the NPS 3–4 group had significantly lower scores on the CERAD (23.79 vs. 27.11), the CFDAST (15.83 vs. 19.13), and the CFDDS (46.26 vs. 55.90), with all differences statistically significant (p < 0.001).

Table 2 illustrates the relationship between NPS groupings and cognitive dysfunction. After comprehensive adjustment for potential confounders—including age, gender, race, education level, smoking status, alcohol consumption, caffeine intake, BMI category, hypertension, diabetes, and energy intake—a domain-specific association was observed: higher NPS scores (3–4 points) were significantly associated with cognitive dysfunction assessed by the CFDAST (executive function/language fluency: OR 2.03, 95% CI 1.18–3.51, p = 0.016). This association remained statistically significant even after applying a conservative Bonferroni correction for three tests (adjusted α = 0.0167). In contrast, no significant associations were found between NPS groupings and cognitive dysfunction assessed by the CERAD test (verbal memory: OR 1.43, 95% CI 0.76–2.69, p = 0.230) or the CFDDS test (processing speed: OR 1.78, 95% CI 0.95–3.34, p = 0.069).

Trend tests indicated significant dose–response relationships for CFDAST (P for trend = 0.007) and CFDDS (P for trend = 0.025, though borderline) after full adjustment, suggesting a graded increase in risk with higher NPS scores. In contrast, no such trend was observed for CERAD (P for trend = 0.247). For CFDAST, the highest NPS group (3–4 points) exhibited significantly elevated cognitive impairment risk across all models (Crude Model: OR 2.41, 95% CI 1.46–3.96, p = 0.001; Model 1: OR 2.19, 95% CI 1.32–3.64, p = 0.004; Model 2: OR 2.03, 95% CI 1.18–3.51, p = 0.016). Notably, the dose–response trend for CFDAST also remained significant after Bonferroni correction (p = 0.007 < 0.0167). Similarly, for CFDDS, the highest NPS group showed significantly increased odds in the crude model (OR 2.31, 95% CI 1.35–3.95, p = 0.003) and Model 1 (OR 1.93, 95% CI 1.10–3.38, p = 0.024), although the association became marginally non-significant in the fully adjusted model (OR 1.78, 95% CI 0.95–3.34, p = 0.069).

The dose–response relationship between NPS as a continuous variable and cognitive impairment risk across all three tests is visually presented in Figure 2.

These results indicate that higher NPS scores are associated with an increased risk of executive function-related cognitive impairment, even after rigorous adjustment for multiple confounding variables. The domain-specific pattern suggests that NPS may be more strongly related to frontal-executive circuits than to memory or processing speed networks.

To validate the rationale for using the composite NPS over individual biomarkers, we compared the model fit (AIC) and discrimination (AUC) of the NPS risk stratification system (Groups 0, 1–2, 3–4) against models based on continuous individual components (NLR, LMR, albumin, and TC) for executive function (CFDAST). As shown in Supplementary Table S1, the NPS Group model demonstrated the superior predictive performance among all tested models, achieving the lowest AIC (1,970.55) and the highest AUC (0.7323). Notably, the composite NPS model exhibited a significant improvement in model fit (ΔAIC >10) compared to the second-best model (TC, AIC = 1,980.99) and the inflammatory marker NLR alone (AIC = 1,983.58). These results suggest that the NPS classification effectively integrates inflammatory and metabolic risks, providing a more robust screening tool than relying on any single continuous biomarker.

To assess the robustness of our primary findings, we conducted two sensitivity analyses. First, when cognitive test scores were analyzed as continuous outcomes in fully adjusted linear regression models, each 1-point increase in NPS (range 0–4) was associated with significantly lower scores on all three cognitive tests: CERAD (β = −0.42, 95% CI: −0.83 to −0.02, p = 0.042), CFDAST (β = −0.75, 95% CI: −0.99 to −0.51, p < 0.001), and CFDDS (β = −1.29, 95% CI: −1.93 to −0.65, p = 0.001) (Supplementary Table S2). Second, when applying a more stringent cutoff (lowest decile, 10%) to define cognitive impairment, the direction and magnitude of associations were similar to the primary analysis (Supplementary Table S3). For CFDAST, the odds ratio for the highest NPS group (3–4) versus the lowest (0) was 2.11 (95% CI: 0.85–5.18, p = 0.096) in the fully adjusted model, compared to 2.03 (1.18–3.51, p = 0.016) in the primary analysis, with the dose–response trend remaining significant (P for trend = 0.046). These sensitivity analyses confirm that the inverse association between higher NPS scores and poorer cognitive performance is robust across different analytical approaches.

To explore potential variations in different population subgroups, we conducted stratified analyses by age and gender (Supplementary Table S4). Although point estimates varied slightly across subgroups, formal interaction tests did not reach statistical significance (all p for interaction >0.05). This suggests that the association between NPS and executive function impairment is generally consistent across age groups and genders, and no definitive effect modification was observed in this population.

In the expanded exploratory clinical validation cohort (n = 189), NPS scores were highly similar between the AD group and the VaD group. The mean NPS score was 3.11 ± 1.00 in the AD group and 3.03 ± 0.97 in the VaD group, with both groups exhibiting a median of 3. The Mann–Whitney U test revealed no statistically significant difference between the two groups (p = 0.432). The common language effect size (CLES) was 0.469, indicating that the probability of a randomly selected individual from the AD group having a higher NPS score than one from the VaD group was 46.9% (Table 3).

The distribution of NPS scores in the two groups was visually presented through box plots (Figure 3). As shown in the figure, the data distribution ranges highly overlapped, further confirming the statistical result that there was no significant difference in NPS between the two groups. These findings suggest a shared inflammatory-metabolic burden across these dementia subtypes.