The data were retrieved from the National Health and Nutrition Examination Survey (NHANES) database,¹ which employs a complex survey design and population-specific sample weights to assess the health or nutritional status of the non-institutionalized population in the United States. The assessment methodology includes a series of sampled household interviews and standardized physical examinations conducted at Mobile Examination Centers (MECs). We retrieved health examination results, laboratory samples, and interviewed personal information from cognitive function surveys conducted among adults aged 60 and over from 2011 to 2014. All participants signed informed consent forms, and the study was approved by the Research Ethics Review Board of the National Center for Health Statistics.

We excluded (1) adults aged under 60 years old. (2) individuals with TG higher than 400 mg/dL. (3) individuals with missing laboratory data on HDL-C, LDL-C, or TC; (4) individuals without cognitive data; (5) individuals with missing data on other co-variables, such as BMI, smoking status, diabetes, and hypertension. A total of 1,331 individuals were ultimately included in this study. The process of recruiting is shown in Figure 1.

Cognitive function was evaluated by four widely applied tests, which included the Consortium to Establish a Registry for Alzheimer’s Disease Word Learning (CERAD-WL), CERAD Delayed Recall (CERAD-DR), Animal Fluency Test (AFT), as well as Digit Symbol Substitution test (DSST) (26). CERAD, designed to identify Alzheimer’s disease, is a highly reliable and valid cognitive function evaluation tool (27). The CERAD-WL involves three consecutive learning trials and an additional delayed recall trial. Participants read 10 irrelevant words in the first three trials and immediately recall as many as possible in a disrupted order each trial. Each correct word scores 1, and the total score is 30 (26). The AFT assesses verbal fluency, where participants are asked to name as many animals as possible in 1 min. A score of less than 14 indicates a decrease in cognitive function (28). The DSST test is a global measure of brain health, evaluating processing speed, visual scanning, sustained attention, and short-term memory (29). In the NHANES study, the version of The Digit Symbol Substitution Test (DSST) used is from the Wechsler Adult Intelligence Scale (WAIS III), with a time limit of 120 s. Participants have 2 min to match symbols with numbers in 133 boxes. Each correct match scores one point. A score greater than 33 indicates healthy cognitive function.

Z-score is calculated by scores from the above four tests for a more reliable and comprehensive evaluation of cognitive function. Here is the formula Z = (x–m)/σ(x for the raw score, m for overall mean, and σ for the overall standard deviation) (30). Recognizing the substantial impact of age on cognitive function, age-stratified quartiles of Z-scores (≥80 years, ≥70 to 80 years, and ≥ 60 to 70 years,) were used to assess cognitive performance. Subsequently, participants were classified into four cognitive function tiers (1 to 4) based on weighted quartiles derived from Z-scores within these age groups. The bottom quartile indicated cognitive impairment, with ascending quartiles corresponding to normal cognitive function.

Sleep duration (14, 31) was self-reported by the question, “How much sleep do you get (hours)?.” Trouble sleeping was evaluated by the answer to “Have you ever told a doctor or other health professional that you have trouble sleeping?.” Sleep disorder was assessed by the response to “Have you ever been told by a doctor or other health professional that you have a sleep disorder?.” The responses to this question have been categorized into two distinct options: ‘yes’ or ‘no’.

In the NHANES laboratory, serum or plasma TC, HDL-C, and LDL-C were measured in participants who had fasted for more than 8.5 h but less than 24 h. TC and TGs were measured using an enzymatic assay method. HDL-C was measured using the heparin-manganese precipitation method or a direct immunoassay technique. Due to technological limitations, the precise numerical value of TG exceeding 400 mg/dL cannot be accurately displayed. When the TG value was 400 mg/dL or less, LDL-C was calculated by the following formula [LDL-C] = [TC]—[HDL-C]—[TG/5] according to the following Friedwald formula (32, 33). RC and TC/RC were the exposure variables. RC = TC—(LDL-C + HDL-C) (17). The TC/RC stands for TC divided by RC.

We collected potential confounding factors from standardized questionnaires. The data contained gender, age, race, education level, marital status, body mass index (BMI), alcohol use, smoking status, diabetes, and hypertension. Self-reported race was categorized into five groups: non-Hispanic White, non-Hispanic Black, Mexican American, other/multiracial race, and other/Hispanic race. Educational level was stratified into three categories: less than high school, high school or equivalent, and college or above. Marital status was divided into two groups: married and unmarried. BMI was categorized into four groups: underweight (<18.5), normal (18.5 to <25), overweight (25 to <30), and obese (30 or greater). Alcohol use was classified into two groups: no/unknown and yes. Smoking status was defined as current smoker, former smoker, and never smoker. Sleep duration was categorized as short (less than 7 h per night), normal (7–9 h per night), or long (greater than 9 h per night). Participants with diabetes were identified if they had any of the following conditions: diagnosed by a doctor or health professional, HbA1c ≥ 6.5%, fasting blood glucose ≥7.0 mmoL/L, random blood glucose ≥11.1 mmoL/L, 2-h OGTT blood glucose ≥11.1 mmoL/L, or use of diabetes medication or insulin. Hypertension was defined based on any of the following: diagnosed by a doctor or health professional, average systolic blood pressure ≥ 140 mmHg, or average diastolic blood pressure ≥ 90 mmHg.

The R software version 4.3.1 was utilized for the analysis of NHANES data, taking into account the intricate sampling design and corresponding sampling weights. The sampling weights for the combined four-year cycles were derived by dividing the individual sampling weights by 2. The data were presented as the mean ± standard deviation for continuous variables, median with interquartile range for skewed variables, and sample counts with weighted percentages for categorical variables. To fit the regression models, TC/RC and continuous RC values were subjected to natural log-transformation. Categorical RC and TC/RC were divided into four subgroups using quartiles as cutoff points.

To evaluate the association between RC, TC/RC, and cognitive performance, we conducted weighted logistic regression analyses. Sequential models were adjusted for potential risk factors: Model 1 without covariate adjustments, Model 2 adjusted for age, gender, and race, Model 3 further including education, marital status, and BMI in addition to the variables in Model 2, and Model 4 incorporating all co-variables such as alcohol use, smoking status, diabetes, and hypertension, along with the adjustments made in Model 3.

In order to assess the robustness of our analytical results, we conducted sensitivity analyses using logistic regression based on unweighted raw data. Additionally, to assess nonlinearity and confirm the dose–response relationship between residual cholesterol levels and cognitive function, a restricted cubic spline (RCS) regression analysis was performed, accounting for all variables in Model 4. Subgroup analysis was conducted to validate the robustness of the results in populations with specific covariate variables. A p-value of less than 0.05 was deemed statistically significant.

Table 1 outlines the sociodemographic and baseline characteristics. 1,331 individuals aged 60 years or older were included. The median age was 68, with females constituting 56% of the cohort. Cognitive impairment was identified in 498 subjects. Populations with different RC levels had similar prevalence rates in terms of age, sex, education level, marital status, trouble sleeping, sleep duration, smoking status, and alcohol-related conditions. However, there were significant differences in the prevalence of race, BMI, sleep disorder, diabetes, hypertension, Z-score, and cognitive function. The prevalence rates of cognitive impairment across the quartiles of RC levels were 18, 23, 33, and 27%, respectively.

Significant differences were observed in age, race, education level, marital status, smoking status, diabetes, sleep duration, HDL-C, RC, LDL-C, TC, and TC/RC between the normal and cognitive impairment groups (Supplementary Table S1). Participants with cognitive impairment exhibited lower levels of TC, LDL-C, HDL-C, and TC/RC (all p < 0.05) and higher levels of RC (p = 0.048). In the cognitive normal group, individuals with prolonged sleep (≥ 9 h) accounted for 13%, surpassing the 6.9% in the cognitive impairment group.

In Table 2, we delineated the relationships between RC, TC/RC and cognitive performance. Participants were stratified based on the interquartile range of RC levels. The lowest quartile (Q1) was the referential baseline. A statistically significant positive association was identified between RC and cognitive performance at the third quartile (Q3) in all analytical models examined. Specifically, for Model 1, the odds ratio (OR) with 95% confidence interval (CI) was calculated as 2.2 (1.43, 3.37), with a p-value of less than 0.001. Similarly, Model 2 exhibited an OR of 2.31 (1.45, 3.69), p = 0.001; Model 3 showed an OR of 2.21 (1.45, 3.37), p = 0.001; and Model 4 demonstrated an OR of 2.11 (1.37, 3.25), p = 0.003. Furthermore, a positive association between TC/RC and cognitive performance was also identified at the Q3 level in Model 4, with an OR of 0.56 (0.33, 0.94) and a p-value of 0.032. Additionally, a non-linear correlation between RC and cognitive performance was observed within restricted cubic spline (RCS) regression model limitations (P-nonlinear = 0.0439) as depicted in Figure 2.

Similarly, sensitivity analyses using unweighted logistic analyses showed that individuals with the highest quartile of the RC had a lower cognitive decline than those with the lowest quartile of the RC, according to the different models, Model 1 (OR = 1.28, 95% CI: 0.93–1.75, p < 0.12), Model 2 (OR = 1.52, 95% CI:1.08–2.14, p < 0.017), and Model 3 (OR = 1.47, 95% CI: 1.01–2.14, p < 0.046) (Table 3). The RC and cognitive performance appear to consistently correlate negatively, according to these findings.

After adjusting for the full model in these subgroups, a significant interaction emerged between RC levels and cognitive impairment specifically among individuals experiencing trouble sleeping (p = 0.035). Americans with higher RC levels who also struggle with sleep difficulties may face an increased risk of cognitive impairment, with an odds ratio of 2.33 (95% CI, 1.18–4.59). Conversely, no such interactions were identified in other subgroups (p > 0.05) in Figure 3. After adjusting for confounding factors, the impact of RC levels on cognitive performance quartiles was consistent across various subgroups, including gender, age, race, education level, marital status, BMI, sleep duration, alcohol use, smoking, diabetes, and hypertension (p ≥ 0.05). Exceptions were noted in individuals with trouble sleeping, no/unknown alcohol use, and no hypertension, where the p-values <0.05.

The study presented several notable strengths, including a substantial sample size and the utilization of Z-scores as a metric for cognitive function evaluation. However, there are some limitations. Firstly, some older adults potentially experiencing cognitive impairment might have been excluded if they were unable to complete cognitive function assessments. Additionally, this research, grounded in the U.S. population and database, necessitates a broader inclusion of diverse regions and further clinical trials to validate the role of RC in cognitive function more conclusively. Since most of the data is self-reported, an information bias may have been induced. What’s more, given the cross-sectional design and lack of follow-up, it is not impossible that a reverse causality bias has been introduced.