This was a population-based longitudinal cohort study initiated in 2014 and followed up every two years. Two villages were selected using random cluster sampling method from suburbs near Xi’an in northwest China. These villages were chosen because their lifestyles and population compositions were similar to other rural areas of Xi’an. The research procedures have been described in detail elsewhere.11 Study inclusion criteria were as follow: (1) participants above 40 years old or who had lived in this area for more than 3 years in 2014 and before; (2) agreed to participate in this study and provided informed consent. The exclusion criteria were as follow: (1) participants had CogI at baseline in 2014; (2) participants refused to finish Mini-Mental State Examination (MMSE); (3) participants with other baseline information missed; (4) participants with other neurological disease that may affect cognitive function (e.g., epilepsy, central nervous system infection, essential tremor, Parkinson's disease, anxiety, depression, thyroid hypofunction, intracranial trauma or surgery); (5) participants who had fewer than three lipid measurements across four follow-up visits (Fig. 1).

Fig. 1.

Flow chart of participant screening. MCI, mild cognitive impairment. MMSE, Mini-Mental State Examination.

This study was approved by the Ethics Committee of The First Affiliated Hospital of Xi’an JiaoTong University, approval number [XJTU1AF2014LSK-038], March 1, 2014. Understanding and written informed consent was obtained from all participants or their next-on-kin and that the study conformed to World Medical Association Declaration of Helsinki published on the website of the Journal of American Medical Association. This study followed the Ethics Committee of The First Affiliated Hospital of Xi’an JiaoTong University’ protocols to access to the patient's data and that was done with the unique purpose of the scientific investigation and scientific disclosure.

In the baseline survey, all participants underwent standardized face-to-face interviews. Demographic information included age, gender, years of education, smoking (yes or no), drinking (yes or no), and lack of physical activity (yes or no). Physical examinations included height, weight, systolic blood pressure (SBP) and diastolic blood pressure (DBP). Medical histories included antihypertensive use (yes or no), antidiabetic use (yes or no), lipid-lowering agents use (yes or no), and other physical diseases. Additionally, all subjects received a neurological examination and a cognitive assessment. After every two years, all subjects accepted the same standardized face-to-face interview as in the baseline again. We tried to avoid the loss of follow-up through the combination of unified field investigation and household investigation (for villagers with inconvenient mobility). Only those who did not respond after contacting villagers themselves, their relatives and neighbors for 3 times were judged to be lost to follow-up (the reasons and the time of missing follow-up were recorded).

Global cognitive function was evaluated in a quiet room using the MMSE. The cutoff values were as follows: scores≤17 for illiteracy, scores≤20 for participants who only finished primary school, and scores≤24 for participants with junior high school or above level of education.12 Participants with a MMSE score below the cutoff value further underwent a neuropsychological battery to assess different cognitive domains as previously described, including Clock Drawing Task, Fuld Object Memory Evaluation, Rapid Verbal Retrieval, Block Design Test, Digit Span Test, and Trail Making Test. Additionally, we used the Activities of Daily Living Scale to assess basic and instrumental functional independence deterioration, and the 15-item Geriatric Depression Scale to evaluate the depression status.

The preliminary diagnosis was made based on the medical history and neuropsychological test results by a neurological expert panel. The diagnosis of mild cognitive impairment (MCI) was made according to the Petersen criteria.13 The diagnosis of dementia was according to the NIA-AA. Those with normal cognition at baseline and diagnosed with MCI or dementia during follow-up were defined as having new-onset CogI.

The collection method has been described in our previous article.14 Serum lipids, including TC, TG, HDL-c, LDL-c and fasting blood glucose (FBG) were detected by a biochemical automatic analyzer (C501, Roche, Sweden). An extraction kit (Tiangen Co. Beijing, China) was used to extract DNA from the frozen EDTA-anticoagulant blood according to the manufacturer's protocol. All PCR products were detected by Sanger sequencing (Sangon Company, Shanghai, China) to finally determine the apolipoprotein E (APOE) genotype.

The diseases and related conditions involved in this study were defined as follows: (1) Subject who have taken antihypertensive, antidiabetic or lipid-lowering agents at least two weeks before the investigation are defined as antihypertensive, antidiabetic or lipid-lowering user, respectively. (2) Continuous or cumulative smoking for 6 months or more defined as smoking. (3) Drinking was defined as alcohol consumption at least once a week in the past 1 year. (4) Subject who exercises less than 30min a day or three times a week over 1 year is defined as lack of physical activity. (5) APOE ɛ4 carriers were defined as having at least one allele of ɛ4 (APOE ɛ2/ɛ4, APOE ɛ3/ɛ4 and APOE ɛ4/ɛ4), while non-carriers referred to the ones without any ɛ4 allele (APOE ɛ2/ɛ2, APOE ɛ2/ɛ3 and APOE ɛ3/ɛ3).

In the analysis of comparing the characteristics between normal cognition group and CogI group, ANOVA, Student's t-test, and Mann–Whitney U-test were used for continuous variables (expressed as mean±SD or median (interquartile range) (M [IQR])). Chi-square test was used for categorical variables (expressed as the number (percentage)).

Group-based trajectory model (GBTM) was used to identify trajectories of serum lipids from 2014 to 2020. GBTM is an exploratory technique that isolates developmental trajectories within a population, aiming to identify subgroups following distinct trajectories over time. We performed GBTM that assigned individuals with similar features of serum lipids trajectories into distinct, exclusive subgroups, using the “traj” command implemented in Stata software. According to the distribution characteristics of serum lipids, censored normal distribution is sensible for modeling the data in GBTM. Models with two or more subgroups were conducted after accounting for the varying trajectory shapes of linear, quadratic and/or cubic terms. Data-based parameters and principles were applied to decide the final subgroups according to Nagin,15 including (i) A minimum Bayesian Information Criterion (BIC) value (Close to 0), (ii) average of the posterior probabilities by class (>0.7), (iii) odds of correct classification based on the posterior probabilities of group membership>5, (iv) minimizing overlap in confidence intervals (CIs) and capturing the distinctive features of the data as parsimonious as possible, (v) class size≥5% of the study population. To facilitate interpretability, we labeled to the trajectories based on their modeled graphic patterns. Considering the non-normal distribution of TG, we applied the natural lg transformation to our data set. Other lipid measurements (including TC, HDL-c and LDL-c) conformed to the normal distribution, and the GBTM method could be directly used for trajectory fitting.

First, when analyzing the relationship between TC and CogI, the incidence rate of CogI among the three groups of low, medium, and high TC level were compared. Secondly, CogI was taken as the dependent variable, low TC level and high TC level were taken as the independent variables, and age, gender, years of education, antihypertensive drug use, lipid-lowering drug use, SBP, DBP, and APOE ɛ4 status (these measures were considered pertinent covariates because they were associated with dementia16) were adjusted in the logistic regression models. Through this step, we analyzed the risk of developing CogI after 6 years in the low TC and the high TC level (compared to the medium level, the reference group). Similar analysis steps were used between TG, HDL-c, LDL-c trajectories and CogI. 167 of 1605 participants at baseline had missing value for a CogI diagnosis (10.4%) due to loss to follow-up in 2020 that were dealt with using Last Observation Carried Forward as predictors. The remaining variables (age, gender, years of education, antihypertensive drug use, lipid-lowering drug use, SBP, DBP, and APOE ɛ4 status) had no missing values. In addition, sensitivity analysis was conducted in 1096 participants who had participated in all 4 follow-ups. The analysis method was the same as described above. Confounding bias was avoided by incorporating confounding factors into the multiple logistic regression model to correct and increased sensitivity analysis.

Multicollinearity statistics (tolerance and variance inflation factors) were calculated for each logistic regression model in this study to avoid multicollinearity. All statistical analysis were performed with IBM SPSS Statistics version 20.0 (IBM. Armonk, New York, USA) and Stata 15.0 (Stata Corp, College Station, Texas, USA). We calculated Odds Ratio (OR) and their 95% confidence intervals (CIs) as a measure of effect size.

The cohort established in 2014 with 2158 villagers who had reached the age of 40. Of these, 2011 individuals enrolled to this cohort. According to the exclusion criteria, 1605 participants with normal cognitive function were enrolled at baseline (visit 1). The follow-ups were conducted in 2016, 2018 and 2020. Except for the visit 1, during the subsequent 3 visits (visit2–visit4), the number of people who visited at each time of study was 1480, 1276 and 1103 respectively. According to our follow-up records, 117 of the 1605 people died during the follow-up years. Other reasons for not participating in the follow-up included relocation of villagers due to demolition, migrant work, migration to other cities, and reluctance to participate in the study. 337 participants who had fewer than three lipid measurements were excluded. Finally, a total of 1268 participants (79.0%) were included in the analysis (Fig. 1). The medium age was 54 [48–62] and 58.6% were women at baseline.

Among the 1268 participants, 33 (2.6%) were diagnosed as CogI according to the criteria, including 20 MCI and 13 dementia during the 6-years follow-up. As shown in Table 1, CogI group was older (65 [57–68.5] vs 54 [47–62]), had higher SBP level (145.36±19.306 vs 131.48±17.785), less educated (3.64±3.507 vs 6.63±3.267) and had lower MMSE scores (22 [18.5–25.5] vs 27 [25–29]), while other characteristics had no difference among the two groups (Table 1).

We estimated the best-fitting number of trajectories based on a minimum BIC and satisfied class size ≥5% of the population. Trajectory shapes of TC, lgTG, HDL-c and LDL-c are selected as 333, 11, 33 and 333 respectively in total population (see Supplementary Material, Supplementary Table 1–4, marked in red font).

According to the best-fitting number of subgroups, trajectory of TC can be divided into three groups. Low level (41.3%, 524): TC remained at a low level, with no tendency to fluctuate. Medium level (47.8%, 606): TC remained at a medium level, with a tendency to increase. High level (10.9%, 138): TC remained at a high level, with an upward trend (Fig. 2a).

Fig. 2.

Trajectory of serum lipids over time in 1268 participants. Trajectories of TC, lgTG, HDL-c and LDL-c were fitted using the GBTM respectively in picture (2a–d). TC, total cholesterol; lgTG, lg triglyceride; HDL-c, high-density lipoprotein; LDL-c, low-density lipoprotein.

According to the best-fitting number of subgroups, trajectory of lgTG can be divided into two groups. Low level (58.2%, 738): It remained at a persistently lower level. High level (41.8%, 530): It remained at a persistently higher level (Fig. 2b).

According to the best-fitting number of subgroups, trajectory of HDL-c can be divided into two groups. Both low level (70.2%, 890) and high level (29.8%, 378) displayed a tendency rising at the beginning and declining in late, but the high trajectory was always higher than the low trajectory, and the two trajectory lines had no trend of intersecting (Fig. 2c).

According to the best-fitting number of subgroups, trajectory of LDL-c can be divided into three groups. Low level (37.5%, 475): It declined in the first two years and showed a slight downward trend after a rebound in the second year. Medium level (53.2%, 675): It rebounded after a decline in the first two years and had maintained a steady trend since its recovery. High level (9.3%, 118): After a rapid decline in the second year, it continued to rise and had a continuous upward trend (Fig. 2d).

Univariate analysis showed that the incidence rate of CogI was different among the low TC, the medium TC, and the high TC group (the incidence rate of CogI: low TC level vs. medium TC level vs. high TC level, 3.4% vs. 1.5% vs. 4.3%, Fig. 3a). Logistic regression model adjusted for confounders showed that compared with the medium TC level, the risk of developing CogI was increased for both low TC level (OR=2.679 [95% CI, 1.140–6.295]) and high TC level (OR=2.828 [95% CI, 0.927–8.624]) (Table 2).

Fig. 3.

Comparison the differences of CogI incidence rate between different lipid trajectory groups in 1268 participants. Comparison the differences of CogI incidence rate between different trajectories of serum lipid groups in 1268 participants. TC, total cholesterol; lgTG, lg triglyceride; HDL-c, high-density lipoprotein; LDL-c, low-density lipoprotein. CogI, cognitive impairment. (a–d) χ2-test.

Univariate analysis showed that no difference was found in low lgTG and high lgTG level group (the incidence rate of CogI: low lgTG level vs. high lgTG level, 2.4% vs. 2.8%, Fig. 3b). Logistic regression model showed that compared with low lgTG level, the high lgTG level was not associated with the risk of developing CogI (Table 2).

Univariate analysis showed that no difference was found in low HDL-c and high HDL-c level group (the incidence rate of CogI: low HDL-c level vs. high HDL-c level, 2.8% vs. 2.1%, Fig. 3c). Logistic regression model showed that compared with low HDL-c level, the high HDL-c level was not associated with the risk of developing CogI (Table 2).

Univariate analysis showed that no difference was found among the low, the medium, and the high LDL-c level group (the incidence rate of CogI: low LDL-c level vs. medium LDL-c level vs. high LDL-c level, 3.6% vs. 1.8% vs. 3.4%, Fig. 3d). Logistic regression model showed that compared with low LDL-c level, the medium and the high LDL-c level was not associated with the risk of developing CogI (Table 2).

To confirm the robustness of the results, sensitivity analysis was conducted in 1096 participants who had participated in all 4 follow-ups and the GBTM was recon-ducted to classify the different serum lipid trajectory groups (see Supplementary Material, Supplementary Tables 5–8, Supplementary Fig. 1) according to the criteria mentioned in method. The results were the same as the previous main analysis after adjusting for the same covariates mentioned above (Table 3). As shown in Table 3, Compared to medium TC level, the risk of developing CogI was increased for both low TC level (OR=2.846 [95% CI, 1.198–6.670]) and high TC level (OR=3.041 [95% CI, 0.982–9.421]). Other lipid trajectory levels were not associated with the risk of developing CogI.