The study data were derived from the Chinese Longitudinal Healthy Longevity Survey (CLHLS), an ongoing longitudinal cohort jointly established by Peking University and Duke University. To investigate the determinants of the health of the Chinese elderly, the CLHLS began in 1998 and was followed in 2000, 2002, 2005, 2008–2009, 2011–2012, 2014, and 2017–2018. The first eight waves of survey covered about half of the counties/cities in 23 China provinces, and the ongoing ninth wave of survey further expanded to 27 provinces. To compensate for missing samples due to death and loss to follow-up in the ongoing survey, each wave of surveys continually recruits new participants based on similar gender, age, and baseline characteristics (Zeng, 2012). CLHLS procedures were approved by the Ethics Committee of Peking University and Duke University (IRB00001052-13074). All participants provided written informed consent. A more detailed description of the design and procedures of CLHLS has been described elsewhere (Yi, 2008).

Considering the lack of height data in the first three waves of survey (1998, 2000, and 2002) and fewer than three follow-up visits in the last two waves (2014 and 2018); in the study, we included the participants who were enrolled in the three surveys conducted in 2005, 2008–2009, and 2011–2012 and followed up to 2018. As summarized in Supplementary Figure S1, a total of 26,477 participants were extracted from the three cohorts. However, 19,375 participants were excluded because of fewer than three visits during the follow-up, and 1,925 participants were excluded because they were diagnosed with dementia or cognitive impairment at the baseline. Among the remaining 5,177 participants, 8 participants aged <60 years at the baseline and 5 participants without assessment of cognitive impairment at the last visit were also excluded. Finally, a total of 5,164 participants were included in the study.

Cognitive function was evaluated using the Chinese version of the Mini-Mental State Examination (C-MMSE), which was a modified and validated cognitive scale for the China elderly (Zeng, 2012; Ming-yue et al., 2015; Gao et al., 2017). The C-MMSE scale consists of 6 dimensions with 24 items, including 5 items for orientation, 1 for naming, 3 for registration, 6 for attention and calculation, 3 for recall, and 6 for language. The total score of the C-MMSE ranges from 0 to 30, with lower scores indicating poorer cognitive function.

According to the definition of cognitive impairment (Ming-yue et al., 2015; An and Liu, 2016), we classified cognitive functions into four mutually exclusive groups: (1) 24 ≤ C-MMSE ≤30: no cognitive impairment; (2) 18 ≤ C-MMSE ≤23: mild cognitive impairment; (3) 10 ≤ C-MMSE ≤17: moderate cognitive impairment; and (4) 0 ≤ C-MMSE ≤9: severe cognitive impairment. Cognitive impairment (C-MMSE ≤23) was used as the main outcome variable.

In the follow-up duration, CVRFs including systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP), pulse pressure (PP), and body mass index (BMI), were continuously measured. After participants were required to rest for 5 min, blood pressure was measured on the right arm at heart level in a sitting position by trained internists using a mercury sphygmomanometer. The blood pressure was continuously measured twice, with at least one-minute interval between the two measurements. According to Korotkoff Phase I and V sound, SBP and DBP values were determined, respectively. MAP was further calculated based on the formula 1/3(SBP) + 2/3(DBP). PP was calculated as (SBP)-(DBP). Body weight (in kilograms) and height (in centimeters) were measured by trained interviewers. BMI (kg/m²) was computed by dividing body weight (kg) by the square of body height (m). Changes in CVRFs during the study were calculated as their values at the end of the follow-up minus those at baseline.

In the present study, we selected baseline covariates, such as sociodemographic characteristics, lifestyle, and medical examination, which may confound the relationships between CVRFs and cognitive impairment based on previous studies (Yi, 2008; Yuan et al., 2019; Zhang et al., 2019). Data on sociodemographic characteristics and lifestyle were obtained by questionnaire. The sociodemographic characteristics included gender (male/female), category of residence areas (city/town/rural), living pattern (with family member(s)/alone), education level, and self-reported economic status (very bad/bad/fair/good or very good). The lifestyle included smoking (yes/no), drinking (yes/no), regular exercise (current/past/never), self-reported sleep quality (very bad/bad/fair/good or very good), fresh fruit consumption (rarely or never/ occasionally/ quite often/ almost every day), and vegetable consumption (rarely or never/occasionally/quite often/almost every day). The medical examinations were from self-reported hospital diagnoses, which included hypertension (yes/no), diabetes (yes/no), heart disease (yes/no), cerebrovascular disease (yes/no), and cancer (yes/no).

First, categorical variables were presented as numbers (percentages), and continuous variables were expressed as median (interquartile range) because of nonnormal distribution. The differences in the baseline characteristics across different cognitive functions were compared using Chi-square and the Mann–Whitney U test.

Second, in the longitudinal development of CVRFs (SBP, DBP, MAP, PP, and BMI) during the follow-up, the latent growth mixture model (LGMM) was adopted to investigate heterogeneity and identify subgroups of participants who shared similar underlying trajectories of CVRFs. The functional form of the trajectories varied across a number of different orders of polynomials, and the best-fitting polynomial form can be specified for each trajectory separately. This property can depict differential age-related patterns of changes in CVRFs observed over the follow-up period in our samples. The trajectories of changes in CVRFs during the follow-up duration were modeled as follows. Because CVRFs were not normally distributed, the rank-order normalization procedure was first used to generate CVRFs with standard normal distributions. The first step was to determine the most appropriate number of trajectories based on the quadratic model. The number of trajectories was chosen based on better goodness of fit (2*ΔBIC >10, indicating better goodness-of-fit for n-class model than n-1 class model), internal reliability (mean posterior probability >0.65 for each latent class, reflecting an acceptable uncertainty of posterior classification) (Marioni et al., 2014; Norris et al., 2021), clinical plausibility, and interpretability (the size of the smallest class size ≥45 subjects). At last, the maximum 2*ΔBIC between the quadratic and cubic order terms was adopted to determine the shape of each trajectory group. Five trajectory models for SBP, DBP, MAP, PP, and BMI were generated based on adequate fit to data, classification accuracy, and clinical interpretability. All participants were assigned to their subgroups with the highest posterior probability for subsequent analyses.

Third, Cox regression models were used to evaluate the hazard ratios (HRs) of cognitive impairment across different trajectories of CVRFs. The age at which the participants first experienced cognitive impairment was used as the timescale for survival analyses. The participants who never experienced cognitive impairment were considered as censored observations and the censoring time was the age of the last assessment of cognitive function. Four adjusted Cox models were established. Model 1 was the basic model without covariates. Model 2 was established based on Model 1 after adjustments for the sociodemographic characteristics. Model 3 was established based on Model 2 with further adjustments for lifestyle. Model 4 was established based on Model 3 with further adjustments for the medical examination.

Fourth, we repeated Cox regression for participants with different sex, residence areas, and living pattern. The interactions were also tested by comparing models with and without a product term between trajectories of CVRFs and sex/residence areas/living patterns. Additionally, a series of sensitivity analyses were conducted to test the robustness of our findings. First, Cox regression was conducted with moderate/severe cognitive impairment (C-MMSE ≤ 17) as the outcome variable. Second, Cox regression was repeated with severe cognitive impairment (C-MMSE ≤ 9) as the unique outcome. Third, we excluded participants with a history of hypertension, diabetes, heart disease, cerebrovascular disease, and cancer at the baseline because major chronic diseases could influence both CVRF trajectories and cognitive impairment. Fourth, we imputed all missing covariates using multiple imputations to test the influence of missing variables. Fifth, we used restricted cubic splines to express the potentially non-linear relationship between changes in CVRFs and the risk of cognitive impairment. A two-sided p < 0.05 was considered statistically significant in all analyses. LGMM analyses were conducted using THE TRAJ procedure in Stata 12.0 (StataCorp LP, College Station, TX, USA). The restricted cubic splines were performed in R (version 4.1.2) with the package “rms.” All other statistical analyses were performed in SPSS 26.0 (IBM SPSS Inc., New York, NY, USA).

This study included 5,164 adults aged ≥60 (median age: 75 years). Among them, 51.8% were male, 35.0% resided in urban areas, and 85.9% lived with their families. During a median 8-year follow-up, the prevalence of cognitive impairment (C-MMSE ≤23) was 40.1% among the participants. There were significant differences in the baseline characteristics across different cognitive functions (Table 1). Compared with the participants without cognitive impairment, participants with cognitive impairment were more likely to be female, rural residents, living alone, less educated, less economically well-off, and eating fewer fruits and vegetables. At the beginning of the follow-up, most CVRFs (e.g., SBP, DBP, and MAP) did not differ in the different cognitive functions, while they differed significantly at the end of the follow-up (Table 2).

Meanwhile, moderate/severe cognitive impairments (C-MMSE ≤17) were developed in 21.4% of the participants, and severe cognitive impairment (C-MMSE ≤9) were observed in 11.6% of the participants (Supplementary Tables S1, S2). Supplementary Table S3 shows that patients excluded from the analysis had very different characteristics from those included, including age and the prevalence of heart disease and diabetes, among others. The included patients were younger, more frequently male, and more likely to be smokers but less likely to have diabetes, heart disease, and cerebrovascular disease.

In trajectory modeling, the number of trajectories was determined by using a quadratic model, and all fit indices are shown in Supplementary Tables S4–S8. The shape of each trajectory group was determined by comparing the goodness of fit (2*ΔBIC > 10) between quadratic and cubic order terms, as shown in Supplementary Tables S9–S13. SBP was taken as an example to illustrate the determination of the optimal model. In the first step, the indicator 2*ΔBIC > 10 indicated that the goodness of fit is better for the n-class model than the n−1 class model. Also, the proportion of subjects classified in each group with a posterior probability >0.65 and a minimum class sample size was greater than 45 in the four-class model. Considering all metrics, clinical plausibility, and interpretability, the four-class model was chosen as the optimal model for the SBP trajectory group. The optimal SBP trajectory model was determined after assessing the goodness of fit (maximum 2*ΔBIC) for the four-trajectory solution shapes (Supplementary Table S14). Based on the same indicators and principles, the optimal trajectory models for DBP, MAP, PP, and BMI were also determined (Supplementary Table S14). Supplementary Table S15 presents the parameter estimation results of modeling the trajectories of CVRFs.

Figure 1 shows the trends and trajectories of changes in CVRFs, including SBP, DBP, and BMI. During the follow-up visits, SBP trajectory was classified into four classes: (1) high stable SBP (n = 1,046, 20.26%) with consistently high SBP level at about 150 mmHg, (2) normal stable SBP (n = 3,863, 74.81%) with the stable SBP at about 130 mmHg, (3) lowered SBP (n = 204, 3.95%) with decreasing SBP from 149 mmHg to 137 mmHg, and (4) greatly elevated SBP (n = 51, 0.99%) with a large increase in SBP from 157 mmHg to 188 mmHg. The three-class DBP trajectory was obtained as follows: (1) normal stable DBP (n = 4,472, 86.60%) with a maintained normal stable DBP at about 81 mmHg, (2) low stable DBP (n = 394, 7.63%) with consistently low DBP around 74 mmHg, and (3) high stable DBP (n = 298, 5.77%) with consistently high DBP at about 92 mmHg. For the four-class BMI trajectory, 7.01% (n = 362) of the participants were progressively obese (from 26 kg/m² to 29 kg/m²), 55.62% (n = 2,872) were stably slim, 36.43% (n = 1,881) maintained a stable normal weight of about 23 kg/m², and only 0.95% (n = 49) had declined weight from 29 to 27 kg/m². MAP was classified as (1) low stable MAP (n = 483, 9.35%), (2) normal stable MAP (n = 4,196, 81.25%), and (3) high stable MAP (n = 485, 9.39%) (Supplementary Figure S2). PP included (1) elevated PP (n = 385, 7.46%), (2) normal stable PP (n = 4,668, 90.40%), and (3) lowered PP (n = 111, 2.15%) (Supplementary Figure S3).

The hazard ratios (HRs) of cognitive impairment by different CVRF trajectories using Cox regression are presented in Table 3. With the normal stable SBP group as a reference, a higher risk of cognitive impairment was observed in the lowered SBP group [adjusted HR (aHR) = 1.67; 95% CI: 1.24–2.25] in a univariate model. After adjusting for different covariates, including sociodemographic characteristics, lifestyle, and physical examination, sequentially from Model 2 to Model 4, the lowered SBP group had a higher risk of cognitive impairment (aHR = 1.59; 95% CI: 1.17–2.16) compared to the normal stable SBP group, and the low stable DBP group had a lower risk of cognitive impairment (aHR = 0.80; 95% CI: 0.67–0.95) compared to the normal stable DBP group in a univariate model. After additional adjustments for covariates, a lower risk of cognitive impairment was still observed in the low stable DBP group than in the normal stable DBP group (aHR = 0.80; 95% CI: 0.66–0.96). Compared with the group with stable normal weight, both the progressively obese group and the stable slim group had a higher risk for cognitive impairment in Model 4 (aHR = 1.28, 95%CI: 1.02–0.62 for the former and aHR = 1.13, 95%CI: 1.02 –1.25 for the latter). Figure 2 shows the adjusted cumulative hazard function curves of cognitive impairment by SBP, DBP, and BMI trajectories, consistent with the main results.

Compared to the participants with normal stable PP, the participants with lowered PP were more likely to suffer from cognitive impairment (aHR = 2.64; 95% CI: 1.66–4.19 in Model 4), while the participants with elevated PP had a lower risk of cognitive impairment (aHR = 0.76; 95% CI: 0.63–0.92 in Model 4). However, there was no statistically significant association between the MAP trajectory groups and the risk of cognitive impairment. The adjusted cumulative hazard of cognitive impairment by MAP and PP trajectories, also show consistent results (Supplementary Figures S4, S5).

According to the baseline characteristics, such as gender, residence areas, and living pattern, the associations between CVRF trajectories and cognitive impairment across different subgroups were broadly consistent with our main findings. Across subgroups, the positive association between the lowered SBP group and the risk for cognitive impairment remained significant (Supplementary Table 16), and the effect of living pattern on appreciable modification of SBP was significant (interaction value of p < 0.05). Low stable DBP was associated with a lower risk of cognitive impairment only in men (aHR = 0.63, 95%CI: 0.47–0.85), and only the effect of gender on appreciable modification of DBP was significant (Supplementary Table 17). The associations of trajectories of other CVRFs with cognitive impairment were unchanged after stratification by gender, residence areas, and living pattern, showing no statistically significant interactions (Supplementary Tables S18–S20).

When moderate/severe cognitive impairment (C-MMSE ≤ 17) and severe cognitive impairment (C-MMSE ≤ 9), were separately used as the outcome variables (Supplementary Tables S21, S22), the results were similar to the main findings. The results remained consistent in general after excluding subjects with a history of hypertension, diabetes, heart disease, cerebrovascular disease and cancer at the baseline (Supplementary Table S23). After imputing missing values of all covariates using multiple imputations, all results remained the same (Supplementary Table S23). The restricted cubic splines were further adopted to explore the relationships between changes in CVRFs during the life course and risk of cognitive impairment after adjustment for possible confounders (Supplementary Figures S6–S10). The more amount of reduction in SBP contributed to a higher hazard of cognitive impairment. The increments in SBP ranging from 5 to 42 mmHg, decreased the risk of cognitive impairment. However, the associations of changes in SBP with risk of cognitive impairment did not reach the predetermined level of statistical significance. More DBP lowering and ≥40 mmHg growth in DBP lead to an increasing hazard of cognitive impairment. When PP increased by 7–25 mmHg, the risk of cognitive impairment reduced, and when PP increased by over 25 mmHg, the hazard of cognitive impairment rose. The risk for cognitive impairment grew with the reduction in MAP during the study. The declines in BMI were associated with a higher likelihood of cognitive impairment.