Investigating the dynamic association of multiple risk factors on cognitive impairment in older people: a retrospective cohort study in Baisha Town, Hepu County, China

Abstract

Background

With the acceleration of global aging, cognitive impairment has become a critical public health issue. However, evidence delineating the dynamic and multifactorial trajectories of its progression remains scarce. Early identification of modifiable lifestyle and clinical risk factors is therefore essential to enable timely and effective prevention and intervention.

Methods

This retrospective cohort study included 6,462 older adults from Baisha Town, Hepu County, Guangxi, China, who underwent health examinations between 2019 and 2024. Cognitive function was assessed using the Mini-Mental State Examination (MMSE). Correlation analysis, logistic regression, Cox proportional hazards models, restricted cubic splines (RCS), and linear mixed-effects models (LMMs) for longitudinal data were employed to examine the dynamic associations of multiple risk factors on cognitive impairment.

Results

Multivariable binary logistic regression showed that lower educational attainment, physical inactivity, poor dental status, higher total cholesterol (TC) and low-density lipoprotein cholesterol (LDL-C), abnormal high-density lipoprotein cholesterol (HDL-C), and elevated fasting blood glucose (FBG) were associated with cognitive impairment in older individuals. Cox proportional hazards models indicated that lower educational attainment, physical inactivity, poor dental status, abnormal HDL-C, and elevated FBG were associated with a higher hazard of incident cognitive impairment. Kaplan–Meier survival curves showed longer maintenance of normal cognition among participants with higher education and regular exercise, whereas the presence of dentures or missing teeth, abnormal HDL-C, and elevated FBG were associated with shorter durations of preserved cognition. LMMs demonstrated sustained or time-varying associations of educational attainment, exercise frequency, dental status, and metabolic factors with trajectories of cognitive function over time.

Conclusion

This study indicates that educational attainment, frequency of physical activity, oral health status, and metabolic factors are associated with temporal variation in the risk of cognitive impairment among older adults. Focusing on these modifiable factors may inform feasible, actionable screening and prevention strategies in primary care and community settings. Our findings also underscore the importance of dynamic monitoring of these indicators and suggest their potential value in slowing cognitive decline.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-07068-8.

Introduction

With accelerating population aging, the global burden of dementia continues to rise: by 2019, approximately 57.4 million people were living with dementia, and this number is projected to increase to 152.8 million by 2050, with more than two-thirds of patients residing in low- and middle-income countries [1, 2]. In China, the prevalence of dementia among people aged 60 and over is about 6.0% (approximately 15.07 million people), and that of mild cognitive impairment is about 15.5% [3]; meanwhile, China is experiencing one of the fastest population aging processes in the world — in 2019, the population aged 60 and above reached 254 million, and it is projected to reach about 402 million by 2040, accounting for approximately 28% of the total population [4]. The large population of older adults makes the prevention and control of cognitive impairment an urgent public health challenge. Therefore, community-based cohort studies that can reflect the dynamic associations of risk factors are significant for formulating screening and intervention strategies.

Cognitive impairment results from the interplay of multiple modifiable factors, and the underlying mechanisms remain uncertain and controversial. Research shows that hypertension may induce vascular cognitive impairment through cerebrovascular damage and may also accelerate the pathological changes of Alzheimer’s disease [5]; diabetes and poor glycemic control can significantly increase the risk of dementia, with the risk of developing Alzheimer’s disease increased by about 56% [6]. However, definitive conclusions about these associations have not yet been reached. The association between blood lipids and cognitive impairment is even more complex. Studies have found that higher total cholesterol (TC) levels in midlife may increase the risk of Alzheimer’s disease and vascular dementia, suggesting that the critical period for lipid management should be moved earlier [7]. High-density lipoprotein cholesterol (HDL-C) shows a nonlinear or U-shaped relationship with dementia risk, and other lipid markers such as low-density lipoprotein cholesterol (LDL-C) and triglycerides (TG) also exhibit heterogeneous effects on cognition, with their specific roles yet to be clarified [8–11].

Lifestyle also plays a role in cognitive impairment. Regular physical activity may improve cerebral blood flow, reduce inflammation and stress, and improve sleep and mood, thereby exerting a association effect on cognition; the studies have shown that those who exercise regularly have lower rates of cognitive decline or dementia, and although the effect size for an individual is limited, this finding has important public health significance for overall health promotion and disease prevention [12]. The effect of exercise may be modulated by genetic background (for example, carriers of the APOE ε4 allele can also partially benefit), suggesting that interventions need to consider gene–environment interactions [13]. Education, by increasing cognitive reserve and elevating baseline cognitive function, can delay the clinical manifestation of dementia, but does not necessarily alter the underlying pathological progression of the disease, and its specific mechanisms remain to be further elucidated [12, 14].

In recent years, oral health has been identified as a potential modifiable factor for cognitive impairment. Mechanistically, poor oral status can damage brain health through chronic oral inflammation releasing proinflammatory mediators, dysbiosis of the oral microbiome affecting systemic and brain inflammation and metabolism, and reduced masticatory function leading to impaired nutrition and dietary diversity [15–17]. For example, periodontal pathogens (such as Porphyromonas gingivalis) have been reported in the brain tissue of Alzheimer’s disease patients, and periodontitis may be associated with β-amyloid deposition and brain atrophy; periodontal treatment can reduce systemic inflammation and potentially slow the progression of cognitive decline [18, 19]. It should be emphasized that there may be a bidirectional relationship between oral health and cognitive impairment: on one hand, oral inflammation and functional impairments can promote neurodegenerative changes through inflammatory responses, microbial imbalance, and nutritional deficiencies; on the other hand, cognitive decline may weaken an individual’s ability to maintain oral hygiene and access dental care, thereby exacerbating oral problems [20]. This bidirectionality increases the risk of reverse causation and time-varying confounding, so there is an urgent need to employ longitudinal or dynamic modeling for further study. In addition, oral frailty is considered a complex syndrome spanning “structure–function–behavior,” and its individual indicators are related to cognitive outcomes in later life. A systematic review including 63 studies (n = 56,520,662) covering 11 indicators showed that reduced number of remaining teeth, periodontitis, and chewing/swallowing difficulties were significantly associated with cognitive impairment, mild cognitive impairment, and dementia [21]; among these, the evidence for “worsening oral health” and “chewing/swallowing/saliva disorders” was most consistent, and decline in oral motor skills was mainly linked to cognitive impairment. This suggests that modifiable oral indicators (such as number of remaining teeth and periodontal status) may be priority targets for monitoring and intervention; focusing on single indicators is more actionable than the general concept of “oral frailty”.

Therefore, multiple modifiable risk factors may have complex dynamic relationships with cognitive impairment in older adults. In China, especially, urban–rural differences lead to significant disparities in the prevalence and risk factors of cognitive impairment [22]. Currently, most related research has focused on populations in Europe and the United States, and systematic, large-scale population evidence in China concerning rural areas, lifestyle, metabolic characteristics, and their interactions with genetic background remains insufficient [23, 24]. This study is based on a retrospective cohort of 6,462 adults aged 65 years and older residing in Baisha Town, Hepu County, Guangxi Province, China, from 2019 to 2024. It comprehensively explores the relationships between multiple modifiable factors (including blood pressure, blood lipids, blood glucose control, physical activity, oral health, and education level) and cognitive impairment in older adults. We hope that these findings can fill the gap in domestic baseline data and provide a scientific foundation for the future development of targeted intervention strategies, public health policies, and personalized health management measures for cognitive impairment in older adults in community settings.

Materials and methods

Study design

The aim of this study is to examine the longitudinal associations between multiple modifiable risk factors and cognitive impairment using a retrospective cohort design within a community-based population in southern China. The data for this study were drawn from a public health surveillance programme in Baisha Town, Hepu County, Guangxi, which monitors the health status and lifestyle of older residents to inform health management and chronic disease prevention. We assembled and analysed all surveillance records collected from 1 January 2019 through 30 November 2024. Throughout this six-year period, assessments and data collection were carried out under consistent, standardised protocols, with no changes to the evaluation instruments or measurement methods. Because these data were originally gathered as part of routine health surveillance, the present work constitutes a secondary analysis of routinely collected monitoring data. This was an observational study with no interventions; clinical trial registration: not applicable.

Participants and eligibility

Participants were permanent residents of Baisha Town aged ≥ 65 years. Inclusion criteria were: (1) at least one complete record containing educational level, exercise frequency, dental status, relevant metabolic factors, and the Mini-Mental State Examination (MMSE) score; (2) longitudinal analysis requires at least two years of data (i.e., two or more time points); and (3) provision of written informed consent. Exclusion criteria were: (1) severe systemic illness (e.g., end-stage renal disease, malignancy); (2) a prior diagnosis of dementia or other neurological disease; (3) hearing impairment that severely hindered data collection (defined as the inability to adhere to the study protocol or comprehend essential instructions despite the use of hearing aids and other assistive methods), while excluding mild to moderate hearing loss that could be managed with such aids to complete assessments and gather relevant data; and (4) major data missingness or inability to determine key variables. In total, 6,462 individuals were included, yielding 10,737 observations, with an overall data completeness of 95.1%. The patient selection process is illustrated in Fig. 1.

Fig. 1.

Flow chart for selection of the study cohort

Data collection and quality control

Data collection was coordinated by the Baisha Town Central Health Center and covered demographics (age, sex, education), lifestyle factors (smoking, alcohol use, exercise frequency), oral health (dental caries, tooth loss, dentures), and metabolic measures including height, weight, blood pressure, fasting blood glucose (FBG), and serum lipids. Lipid assays comprised TC, TG, LDL-C, and HDL-C, measured enzymatically using a Hitachi 7600 automated analyzer; all participants provided venous blood samples after an 8–10 h fast.

All oral examinations were conducted on-site by uniformly trained licensed dentists from primary healthcare centers, following a standardized protocol. Examinations were performed in a private, well-lit room using disposable tongue depressors and mouth mirrors, with visual inspection supplemented by gentle probing; radiographic imaging was not used. The number of remaining teeth, the presence of active carious lesions, and denture use were recorded, and these indicators were updated at each follow-up health check.

Cognitive function was assessed using MMSE [25–28], administered in Mandarin by uniformly trained clinicians at the community health-check sites in accordance with a standardised protocol. Specifically, before the survey we provided standardized clinical training to the physicians. During this training, two evaluators independently administered the MMSE and performed oral examinations on a subset of participants, and when we compared their findings we observed a high degree of consistency. When dialect barriers arose, local healthcare staff provided isosemantic clarifications without altering item intent. Assessments were conducted in a quiet, private room; responses were recorded on site and subsequently verified through double data entry by the research team. Participants unable to complete the assessment had the reason documented and were excluded from analysis. In the course of the study, we implemented a series of supportive measures for participants with hearing impairment to enhance their engagement and ensure accurate data collection. These measures included making sure assessments were done in a quiet environment, having evaluators face participants and speak more slowly and clearly (with increased volume as needed). When necessary, we also employed visual aids—such as written text or pictures—to help explain test instructions. Additionally, depending on each participant’s needs, we considered providing a calibrated behind-the-ear hearing aid (Banglijian model C-108) for additional assistance. The MMSE ranges from 0 to 30, with higher scores indicating better cognition. We classified MMSE < 27 (i.e., ≤ 26) as cognitive impairment. Quality control was performed by personnel with standardized training, using double data entry and logical consistency checks.

Variable definitions

Dyslipidemia thresholds followed the 2016 Chinese Guideline for the Prevention and Treatment of Adult Dyslipidemia [29]: TC > 5.2 mmol/L, TG > 1.81 mmol/L, and LDL-C > 3.1 mmol/L. In line with prior literature [30] and clinical interpretability, HDL-C was categorized into five groups: very low (< 1.00 mmol/L), low (1.00–<1.49 mmol/L), reference (1.50–<1.99 mmol/L), high (2.00–<2.49 mmol/L), and very high (≥ 2.50 mmol/L). Other variables included body mass index (BMI, kg/m²) categorized as underweight (< 18.5), normal (18.5–23.9), overweight (24.0–27.9), and obesity (≥ 28.0); hypertension defined as systolic blood pressure ≥ 140 mmHg and/or diastolic blood pressure ≥ 90 mmHg [31]; and normal FBG was defined as 3.9 to < 6.1 mmol/L, and elevated FBG as ≥ 6.1 mmol/L.

Statistical analysis

Analyses were performed using IBM SPSS Statistics 26.0 and R 4.2.1.

Missing data

We evaluated the missingness mechanism with Little’s test for missing completely at random (MCAR). If P > 0.05, complete-case analysis was used for the primary analyses. Sensitivity analyses employed multiple imputation by chained equations (MICE) to assess robustness.

In addition, to evaluate whether excluding participants with severe hearing impairment (defined as hearing loss that continued to impede assessment and data collection even with hearing aids or other assistance) influenced the results, we compared the baseline characteristics of those excluded versus those included. We quantified any imbalance using standardized mean differences (SMD), with an SMD < 0.10 considered negligible.

Cross-sectional analyses

Group differences between participants with and without cognitive impairment were compared using the χ² test or the Mann–Whitney U test. Pearson and Spearman correlations were used to examine associations among variables. Multivariable binary logistic regression then identified risk factors associated with the outcome.

Longitudinal analyses

We fitted Cox proportional hazards models to estimate the influence of risk factors on incident cognitive impairment, defining the event as the first occurrence of an MMSE score falling below the threshold (< 27). Models were adjusted for covariates including age, sex, and education, and the proportional hazards assumption was verified. Kaplan–Meier (KM) survival curves were used to visualize between-group differences.

Nonlinearity and dynamic associations

Restricted cubic spline (RCS) analyses were used to explore nonlinear relationships of continuous predictors with MMSE. Finally, linear mixed-effects models (LMM) assessed how time-varying risk factors related to longitudinal change in MMSE. Model selection used the Akaike information criterion (AIC) to compare goodness of fit.

In brief, we first used multivariable binary logistic regression to identify factors associated with cognitive impairment (MMSE < 27) at baseline and at the most recent examination; next, we applied Cox proportional hazards models to evaluate time-to-event risk, incorporating RCS to examine potential nonlinearities; and finally, we used LMM to quantify the dynamic associations of risk factors on MMSE trajectories over time. All tests were two-sided with α = 0.05.

Results

Completeness of health examination data among older adults

In this study, we included 6,462 older adults who underwent physical examinations between 2019 and 2024, with a total of 11,286 examination instances (Table 1). Among these, 10,737 examinations had completely recorded data, resulting in a data completeness rate of 95.1%. The proportion of missing data for all analyzed indicators was low (all under 3%). Specifically, missing data were found for marital status (129 cases, 1.1%), serum LDL-C (75 cases, 0.70%), serum HDL-C (40 cases, 0.40%), TG (30 cases, 0.30%), education level (300 cases, 2.7%), and FBG (102 cases, 0.90%). For other variables – including examination age, height, weight, waist circumference, sex, TC, smoking status, alcohol consumption, exercise habits, denture use, missing teeth, dental caries, and BMI category – the data were complete with no missing records.

Table 1.

Completeness of health examination data and basic indicator statistics for residents aged 65 years and above

Variable	Number of Cases	Mean ± Standard Deviation (SD)	Missing Count	Missing Percentage(%)
Age (Years)	11286	74.5 ± 6.8	0	0.0
Height (cm)	11286	159.3 ± 7.8	0	0.0
Weight (kg)	11286	54.6 ± 8.0	0	0.0
Waist circumference (cm)	11285	79.1 ± 6.7	1	0.0
Gender	11286		0	0.0
Marital status	11157		129	1.1
LDL-C (mmol/L)	11211		75	0.7
HDL-C (mmol/L)	11246		40	0.4
TG (mmol/L)	11256		30	0.3
TC (mmol/L)	11286		0	0.0
Educational Level	10986		300	2.7
Smoking Status	11286		0	0.0
Drinking Status	11286		0	0.0
Exercise	11286		0	0.0
FBG (mmol/L)	11184		102	0.9
Dentures	11286		0	0.0
Missing Teeth	11286		0	0.0
Caries	11286		0	0.0
BMI (kg/m²)	11286		0	0.0

Although the missing proportions were small, we performed Little’s test for missing completely at random (MCAR), which yielded a P-value > 0.05, indicating that the missingness was approximately MCAR. Subsequently, multiple imputation by chained equations (MICE) was used for sensitivity analysis. The post-imputation analysis showed that the directions of estimated associations and statistical conclusions were consistent with those of the complete-case analysis (Supplementary Table 1).

Among 7,398 participants aged ≥ 65 years, 309 (approximately 4.18%) were excluded due to hearing impairments that still severely hampered data collection despite the use of hearing aids or other assistive methods. This exclusion proportion was relatively small. We compared the baseline characteristics of these 309 excluded individuals (whose severe hearing loss precluded effective data collection despite assistance, and thus were not included in the main analysis) with those of the included participants. The two groups showed minimal differences (all SMD < 0.1; Supplementary Table 2).

Univariate analysis

We performed univariate analysis on 6,462 older adults undergoing their first health examination to explore associations between cognitive impairment and various demographic and behavioral characteristics (Table 2; Fig. 2). The results showed that among those with cognitive impairment, the proportion of women was significantly higher than that of men (59.39% vs. 40.61%, χ² = 23.831, P < 0.001). Similarly, the proportion of individuals with lower educational levels (illiterate or semi-literate) was significantly higher in the cognitive impairment group than in the cognitively normal group (χ² = 37.489, P < 0.001). Furthermore, exercise frequency was significantly associated with the presence of cognitive impairment (χ² = 207.023, P < 0.001). In the cognitive impairment group, the proportion of individuals with elevated TC and elevated FBG was significantly higher than non-impaired group (χ² = 294.276 and 122.126, respectively; both P < 0.001). Additionally, the distribution of HDL-C across five categories differed significantly between the two groups (χ² = 99.215, P < 0.001). Normality tests for age, height, weight, and waist circumference showed that none of these variables were normally distributed (all P < 0.001) (Supplementary Table 3). Mann–Whitney U tests indicated that older adults with cognitive impairment differed significantly from those without impairment in height and weight (height in cm: U = 2,183,538, z = − 5.693, P < 0.001; weight in kg: U = 2,362,522, z = − 2.223, P = 0.026) (Supplementary Table 4).

Table 2.

Demographic and behavioral characteristics of subjects at first health examination

Variable		Cognitive Impairment Presence (%)		Total	χ 2	P
		No	Yes
Gender	Male	2754(49.417)	361(40.607)	3115(48.205)	23.831	< 0.001
	Female	2819(50.583)	528(59.393)	3347(51.795)
Marital Status	Single	82(1.493)	20(2.307)	102(1.604)	5.229	0.156
	Married	4415(80.360)	701(80.854)	5116(80.428)
	Divorced	10(0.182)	0(0.000)	10(0.157)
	Widowed	987(17.965)	146(16.840)	1133(17.812)
Educational Level	Illiterate	1110(20.655)	231(26.674)	1341(21.490)	37.489	< 0.001
	Elementary School	3290(61.221)	543(62.702)	3833(61.426)
	Junior High School/Technical School	863(16.059)	81(9.353)	944(15.128)
	High School/Secondary Specialized School	111(2.047)	11(1.270)	122(1.939)
Smoking Status	Never Smoked	5522(99.085)	886(99.663)	6408(99.164)	3.368	0.186
	Quit Smoking	8(0.144)	1(0.112)	9(0.139)
	Smoker	43(0.772)	2(0.225)	45(0.696)
Drinking Status	Never	5509(98.852)	883(99.325)	6392(98.917)	1.661	0.646
	Occasionally	28(0.502)	3(0.337)	31(0.480)
	Frequently	24(0.431)	2(0.225)	26(0.402)
	Daily	12(0.215)	1(0.112)	13(0.201)
Exercise	None	2889(51.839)	267(30.034)	3156(48.839)	207.023	< 0.001
	Occasional Exercise	501(8.990)	181(20.360)	682(10.554)
	Regular Exercise (≥ 2 times per week)	1319(23.668)	310(34.871)	1629(25.209)
	Daily Exercise	864(15.503)	131(14.736)	995(15.398)
Dentures	No	3576(64.167)	485(54.556)	4061(62.844)	30.329	< 0.001
	Yes	1997(35.833)	404(45.444)	2401(37.156)
Missing Teeth	No	4171(74.843)	689(77.503)	4860(75.209)	2.909	0.088
	Yes	1402(25.157)	200(22.497)	1602(24.791)
Caries	No	5203(93.361)	787(88.526)	5990(92.696)	26.465	< 0.001
	Yes	370(6.639)	102(11.474)	472(7.304)
TC(mmol/L)	Normal	4241(76.099)	430(48.369)	4671(72.284)	294.276	< 0.001
	High	1332(23.901)	459(51.631)	1791(27.716)
TG(mmol/L)	Normal	4248(76.321)	666(75.169)	4914(76.162)	0.558	0.455
	High	1318(23.679)	220(24.831)	1538(23.838)
HDL-C(mmol/L)	Very Low	1116(20.1)	260(29.4)	1376(21.4)	99.215	< 0.001
	Low	280(5)	63(7.1)	343(5.3)
	Moderate	3552(63.9)	415(46.9)	3967(61.6)
	High	425(7.6)	88(10)	513(8)
	Very High	185(3.3)	58(6.6)	243(3.8)
LDL-C(mmol/L)	Normal	4249(76.600)	412(46.818)	4661(72.522)	338.050	< 0.001
	High	1298(23.400)	468(53.182)	1766(27.478)
BMI(kg/m²)	Underweight	548(9.833)	76(8.549)	624(9.656)	2.821	0.420
	Normal	4416(79.239)	709(79.753)	5125(79.310)
	Overweight	542(9.725)	89(10.011)	631(9.765)
	Obese	67(1.202)	15(1.687)	82(1.269)
FBG(mmol/L)	Normal	4674(84.828)	607(69.610)	5281(82.748)	122.126	< 0.001
	High	836(15.172)	265(30.390)	1101(17.252)
Hypertension	No	3048(54.712)	538(60.517)	3586(55.511)	10.463	0.001
	Yes	2523(45.288)	351(39.483)	2874(44.489)

TC (mmol/L): normal ≤ 5.20, high > 5.20 ; TG (mmol/L): normal 0.40–1.81, high > 1.81 ; HDL-C (mmol/L): very low < 1.00, low 1.00–1.49, moderate 1.50–1.99, high 2.00–2.49, very high ≥ 2.50 ; LDL-C (mmol/L): normal ≤ 3.10, high > 3.10 ; BMI (kg/m²): underweight < 18.5, normal 18.5–23.9, overweight 24.0–27.9, obese ≥ 28.0 ; FBG (mmol/L): normal 3.9 – <6.1, high ≥ 6.1

Fig. 2.

Grouped proportions of significant variables associated with cognitive impairment based on the chi-square test

In the data from the most recent health examination, demographic and behavioral characteristics were compared between the cognitive impairment group and the cognitively normal group (Supplementary Table 5). The results showed no significant differences between the two groups in smoking status, alcohol consumption, TG level, or BMI category (all P > 0.05). However, the cognitive impairment group exhibited a significantly higher prevalence of female gender, illiteracy, denture use, missing teeth (P = 0.037), dental caries, and elevated TC, LDL-C, and FBG compared to the non-impaired group (all other P < 0.01). Exercise frequency distributions differed significantly (χ2 = 219.422, P < 0.001); however, the higher activity levels in the impairment group may stem from reverse causality or confounding factors typical of cross-sectional data. The distribution of HDL-C categories differed significantly between the cognitive impairment group and the non-impaired group (P < 0.001), and the prevalence of hypertension was significantly higher in the cognitive impairment group (P < 0.01). Additionally, Mann–Whitney U tests for examination age, height, weight, and waist circumference indicated that those with cognitive impairment were older, shorter, and lighter than those without impairment, whereas waist circumference did not differ significantly (P = 0.109) (Supplementary Table 6).

Univariate correlation analysis

We conducted Pearson and Spearman correlation analyses to examine the relationships among variables in the older population (Fig. 3 and Supplementary Table 7). Pearson analysis showed that examination age was significantly negatively correlated with height (r = − 0.112, P < 0.01) and weight (r = − 0.195, P < 0.01), whereas height and weight were significantly positively correlated with each other (r = 0.665, P < 0.01). Spearman analysis indicated that sex was significantly correlated with marital status (r = 0.227, P < 0.01) and educational level (r = − 0.361, P < 0.01). Exercise was positively correlated with dental caries (r = 0.078, P < 0.01) and elevated TC (r = 0.125, P < 0.01). Additionally, elevated TC and elevated LDL-C showed a particularly strong positive correlation (r = 0.630, P < 0.01).

Fig. 3.

Correlation heatmap of relevant indicators among older adults. Notes: *P＜ 0.05, **P＜ 0.01, ***P＜ 0.001

Multivariate logistic regression analysis

On the basis of the most recent examination data, we performed multivariate binary logistic regression analysis to investigate the independent associations of multiple variables on cognitive impairment in older adults (Table 3; Fig. 4). The analysis included factors such as age, marital status, education level, exercise frequency, denture use, missing teeth, dental caries, TC, LDL-C, HDL-C, FBG, and height. Among marital status categories, widowed individuals had a lower risk of cognitive impairment compared to unmarried individuals (OR = 0.547, 95% CI: 0.333 ~ 0.898, P = 0.017). Higher educational levels were associated with lower risk of cognitive impairment. Compared to the illiterate group, the ORs for primary school, middle school/vocational school, and high school/technical secondary school were 0.634, 0.478, and 0.273, respectively (all P < 0.001). Compared to non-exercisers, the likelihood of cognitive impairment was highest in occasional exercisers (OR = 2.880, 95% CI: 2.394~3.465) and was comparatively lower in those with higher exercise frequency: 1.629 (95% CI: 1.379~1.925) for ≥2 times/week and 1.586 (95% CI: 1.312~1.918) for daily exercisers (all P < 0.001). Poor oral health – including denture use (P < 0.001), missing teeth (P < 0.01), and dental caries (P < 0.001) – was associated with a higher risk of cognitive impairment. Elevations in TC, LDL-C, and FBG each significantly increased the risk of cognitive impairment (all P < 0.001). The association of HDL-C was more complex: using the moderate HDL-C group as the reference, the very low HDL-C group had a significantly higher risk (OR = 1.312, 95% CI:1.126 ~ 1.527, P < 0.001) and the low HDL-C group also had a significantly higher risk (OR = 1.644, 95% CI:1.309 ~ 2.064, P < 0.001). In contrast, the high and very high HDL-C groups did not differ significantly from the reference group (both P > 0.001). The Hosmer–Lemeshow goodness-of-fit test yielded χ² = 4.766 (P = 0.782), indicating an adequate fit of the model.

Table 3.

Summary of the results from multivariable binary logistic regression analysis

Variable	Regression Coefficient	Standard Error	Wald χ2	P	OR	95% CI for OR
Age (Years)	0.032	0.005	37.724	<0.001	1.032	(1.022 ~ 1.043)
Gender (Male)	0.115	0.092	1.559	0.212	1.122	(0.937 ~ 1.344)
Marital Status (Single)	reference
Marital Status (Married)	-0.384	0.244	2.483	0.115	0.681	(0.424 ~ 1.102)
Marital Status (Divorced)	-0.203	0.762	0.071	0.790	0.816	(0.183 ~ 3.636)
Marital Status (Widowed)	-0.604	0.253	5.685	0.017	0.547	(0.333 ~ 0.898)
Educational Level (Illiterate or Semi-Illiterate)	reference
Educational Level (Elementary School)	-0.455	0.084	29.241	< 0.001	0.634	(0.538 ~ 0.748)
Educational Level (Junior High / Technical School)	-0.737	0.129	32.622	< 0.001	0.478	(0.371 ~ 0.616)
Educational Level (High School / Secondary Technical)	-1.299	0.308	17.773	< 0.001	0.273	(0.149 ~ 0.499)
Exercise (None Exercise)	reference
Exercise (Occasional Exercise)	1.058	0.094	125.752	<0.001	2.880	(2.394 ~ 3.465)
Exercise (Regular Exercise (≥ 2 times per week))	0.488	0.085	32.931	<0.001	1.629	(1.379 ~ 1.925)
Exercise (Daily Exercise	0.461	0.097	22.641	<0.001	1.586	(1.312 ~ 1.918)
Dentures (Yes)	0.667	0.075	80.006	<0.001	1.949	(1.684 ~ 2.256)
Missing Teeth (Yes)	0.238	0.085	7.774	0.005	1.268	(1.073 ~ 1.499)
Caries (Yes)	0.424	0.102	17.296	<0.001	1.528	(1.251 ~ 1.866)
TC (High)	0.300	0.085	12.298	<0.001	1.349	(1.141 ~ 1.595)
HDL-C (Moderate)	reference
HDL-C (Very Low)	0.271	0.078	12.204	<0.001	1.312	(1.126 ~ 1.527)
HDL-C (Low)	0.497	0.116	18.260	<0.001	1.644	(1.309 ~ 2.064)
HDL-C (High)	-0.008	0.131	0.003	0.954	0.992	(0.768 ~ 1.283)
HDL-C (Very High)	0.184	0.190	0.933	0.334	1.202	(0.828 ~ 1.744)
LDL-C(High)	0.716	0.083	73.525	<0.001	2.046	(1.737 ~ 2.409)
FBG(High)	0.418	0.074	31.408	<0.001	1.518	(1.312 ~ 1.757)
Hypertension(Yes)	0.097	0.066	2.159	0.142	1.102	(0.968 ~ 1.254)
Height	-0.025	0.007	13.750	<0.001	0.976	(0.963 ~ 0.988)
Weight	0.003	0.006	0.256	0.613	1.003	(0.992 ~ 1.014)

Dependent Variable = Presence of Cognitive Impairment in older people; Hosmer-Lemeshow Test：χ2 (8)=4.766，P = 0.782

Fig. 4.

Forest plot of multiple independent variables on cognitive impairment in older people

Cox proportional hazards regression

This study’s data were collected from community-dwelling participants in southern China between January 1, 2019 and November 30, 2024. Baseline was defined as the date of the participant’s first examination that met the inclusion criteria; for individuals with multiple examination records, only the first qualifying examination was used as baseline. The follow-up period began at baseline and continued until the first occurrence of a cognitive impairment event; participants who did not develop cognitive impairment were censored at the end of the study period.

In this study, we employed a Cox proportional hazards regression model to examine the associations of multiple factors on the risk of cognitive impairment (Table 4). In terms of education level, compared to the illiterate group, the hazard ratios for the primary school, junior secondary/technical school, and senior secondary/vocational school groups were successively lower, at 0.730 (95% CI: 0.645 ~ 0.827, P < 0.001), 0.628 (95% CI: 0.517 ~ 0.762, P < 0.001), and 0.305 (95% CI: 0.157 ~ 0.595, P < 0.001), respectively. Regarding exercise frequency, compared with non-exercisers, occasional exercise showed no significant difference (P = 0.224). However, regular exercise (≥ 2 times per week) and daily exercise were associated with reduced risk, with HRs of 0.820 (95% CI: 0.717 ~ 0.938, P = 0.004) and 0.734 (95% CI: 0.631 ~ 0.855, P < 0.001), respectively. For oral health status, wearing dentures and having missing teeth were significantly associated with cognitive impairment, with HRs of 1.765 (95% CI: 1.564 ~ 1.991, P < 0.001) and 1.216 (95% CI: 1.062 ~ 1.393, P = 0.005), respectively. Among metabolic indicators, elevated FBG was significantly associated with cognitive decline (HR = 1.359, 95% CI: 1.225 ~ 1.506, P < 0.001). In a Cox proportional hazards model stratifying HDL-C into five levels (using the moderate HDL-C group as the reference), HDL-C level was significantly associated with the incidence of cognitive impairment (all P < 0.002). Specifically, compared to the moderate HDL-C group: the extremely low HDL-C group had a 20.4% higher risk (HR = 1.204, 95% CI 1.074 ~ 1.350, P = 0.002); the low HDL-C group had a 41.7% lower risk (HR = 0.583, 95% CI 0.474 ~ 0.718, P < 0.001); and the high HDL-C and extremely high HDL-C groups each had roughly double the risk (high HDL-C: HR = 2.012, 95% CI 1.690 ~ 2.396, P < 0.001; extremely high HDL-C: HR = 2.126, 95% CI 1.655 ~ 2.730, P < 0.001).

Table 4.

Summary of Cox regression model analysis results

Variable	Regression Coefficient	Standard Error	P	HR	95% CI for HR
Age (Years)	-0.036	0.005	<0.001	0.964	(0.955 ~ 0.973)
Height (cm)	-0.024	0.005	<0.001	0.977	(0.967 ~ 0.987)
Weight (kg)	0.008	0.005	0.090	1.008	(0.999 ~ 1.017)
Gender (Female)	-0.131	0.074	0.076	0.877	(0.758 ~ 1.014)
Marital Status (Single)	reference
Marital Status(Married)	-0.304	0.226	0.179	0.738	(0.473 ~ 1.15)
Marital Status(Divorced)	-0.323	0.551	0.558	0.724	(0.246 ~ 2.132)
Marital Status(Widowed)	-0.171	0.232	0.461	0.843	(0.535 ~ 1.328)
Educational Level (Illiterate or Semi-Illiterate)	reference
Educational Level (Elementary School)	-0.315	0.063	<0.001	0.730	(0.645 ~ 0.827)
Educational Level (Junior High/Technical School)	-0.465	0.099	<0.001	0.628	(0.517 ~ 0.762)
Educational Level (High School/Secondary Technical)	-1.186	0.340	<0.001	0.305	(0.157 ~ 0.595)
Exercise (None Exercise)	reference
Exercise(Occasionally Exercise)	0.087	0.072	0.224	1.091	(0.948 ~ 1.257)
Exercise(Regular Exercise ≥ 2 times per week)	-0.198	0.068	0.004	0.820	(0.717 ~ 0.938)
Exercise(Daily Exercise)	-0.309	0.077	<0.001	0.734	(0.631 ~ 0.855)
Dentures (Yes)	0.568	0.062	<0.001	1.765	(1.564 ~ 1.991)
Missing Teeth (Yes)	0.196	0.069	0.005	1.216	(1.062 ~ 1.393)
Dental Caries (Yes)	-0.102	0.080	0.201	0.903	(0.773 ~ 1.056)
TC (Yes)	0.006	0.063	0.921	1.006	(0.89 ~ 1.138)
HDL-C (Moderate)	reference
HDL-C (Very Low)	0.186	0.059	0.002	1.204	(1.074 ~ 1.35)
HDL-C (Low)	-0.539	0.106	<0.001	0.583	(0.474 ~ 0.718)
HDL-C (High)	0.699	0.089	<0.001	2.012	(1.69 ~ 2.396)
HDL-C (Very High)	0.754	0.128	<0.001	2.126	(1.655 ~ 2.73)
LDL-C (Yes)	-0.014	0.063	0.821	0.986	(0.872 ~ 1.114)
FBG (Yes)	0.306	0.053	<0.001	1.359	(1.225 ~ 1.506)
Hypertension (Yes)	0.096	0.052	0.064	1.101	(0.994 ~ 1.219)

KM survival analysis

KM survival analysis was performed to investigate the association of various factors on cognitive function in older adults (Fig. 5). Throughout the follow-up period, the event-free survival curves for all subgroups showed an overall decline over time, with visible separation across several covariates. Notably: Education level (Fig. 5A): A clear gradient was observed, with the senior secondary/vocational school group showing the highest survival, followed by the junior secondary/technical, primary, and illiterate groups; the illiterate group had the lowest survival. The gap between the groups widened as follow-up progressed. Exercise frequency (Fig. 5B): The curves for the four exercise frequency groups were generally close together, with only slight separation near the end of the follow-up. Specifically, the daily exercise group exhibited a marginally higher survival curve, while the never-exercise group was slightly lower. Denture use (Fig. 5C): Participants wearing dentures had consistently lower survival than those without dentures; this separation became more pronounced after the mid-point of follow-up. Tooth loss (Fig. 5D): Participants with missing teeth showed slightly lower survival than those without missing teeth, with differences emerging mainly in the later follow-up period. HDL-C categories (Fig. 5E): When stratified into five groups, the survival curves showed a nonlinear order: compared to the “moderate” group, the “low” HDL-C groups had higher survival probability, whereas the “very low”, “very high” and “high” groups all showed lower survival rates than the “moderate” group. FBG levels (Fig. 5F): The high FBG group had lower survival throughout compared to the normal FBG group. The gap between these two groups became stable from mid follow-up and widened toward the end of the study.

Fig. 5.

Event-free surviv curves stratified by covariates (KM curve). Notes: A: KM curve for cognitive impairment according to educational level; B: KM curve for cognitive impairment according to exercise frequency; C: KM curve for cognitive impairment according to denture use; D: KM curve for cognitive impairment according to tooth loss; E: KM curve for cognitive impairment according to HDL-C level; F: KM curve for cognitive impairment according to elevated FBG

RCS analysis

RCS analysis was conducted to further investigate nonlinear associations of continuous variables on cognitive impairment risk, given that Cox regression for lipid levels did not show obvious statistical differences. We applied the RCS method within the Cox model for age, height, weight, TC, TG, HDL-C, LDL-C, and FBG (Fig. 6). The RCS results showed: Age: There was a significant nonlinear association with the outcome (Nonlinear P < 0.001). The HR curve declined rapidly from a high value to near 1 in the age range of about 60–70 years, then remained approximately flat with only a slight uptick at the upper end of age. Height: The relationship was overall linear (Nonlinear P = 0.240). As height increased, the risk gradually decreased, approaching an HR of 1 around 165–170 cm. Weight: No significant nonlinearity was observed (Nonlinear P = 0.740). The curve was essentially flat overall, with a slight increase in HR after about 60 kg and wider confidence intervals at both ends. TC and TG: No nonlinear associations were observed (Nonlinear P = 0.957 and 0.469, respectively). HRs increased slightly with higher TC and TG levels, but uncertainty was large at the upper extremes. LDL-C: No significant nonlinearity (Nonlinear P = 0.103). The HR curve showed a small bump around 3–4 mmol/L then flattened thereafter. HDL-C: Nonlinearity was not significant (Nonlinear P = 0.128), but the curve indicated elevated HR at extremely low HDL-C levels, followed by a rapid decline and leveling off in the mid-high range, suggesting a possible risk peak at the low end. FBG: A significant nonlinear association was present (Nonlinear P < 0.001). The HR curve peaked around 7–8 mmol/L and then slowly declined, with wide confidence intervals at the high end.

Fig. 6.

Nonlinear dose-response relationship between continuous exposure and risk of cognitive impairment (RCS)

LMM analysis

Cognitive score trajectories

The comparison results of cognitive function scores under different numbers of medical examinations based on the LMM model analysis indicate a gradual decline in participants’ cognitive function scores with an increase in the number of medical examinations (Fig. 7). Specifically, among the participants who underwent medical examinations 2 to 6 times, the mean cognitive function scores showed a gradual decrease, with a noticeable decline observed in the group that underwent the 4th examination and beyond.

Fig. 7.

Comparison of mean cognitive function scores across different numbers of health examinations (LMM model, bar plot). Note: The figure displays the mean cognitive function scores and their 95% confidence intervals for participants grouped by the number of health examinations (2–6 times). The red dashed line indicates the reference thresho for cognitive impairment (Score = 27)

Figure 8 displays the results of the LMM model analysis, showing the trajectories of cognitive function scores for individuals under different numbers of medical examinations. The majority of the black lines exhibit a downward trend, indicating a decline in individuals’ cognitive abilities with an increase in follow-up visits. The red trend line further confirms the overall direction of cognitive function scores. These results provide crucial clues for the development of cognitive impairment, emphasizing the importance of continuous monitoring and intervention in cognitive function changes.

Fig. 8.

Individual trajectories of cognitive function scores in lmm model (line plot). Notes: Each black line represents the trajectory of cognitive function scores for an individual participant across different numbers of health examinations. The red solid line shows the overall fitted trend

Figure 9 compares the fitting trends of linear regression and mixed-effects models in analyzing the relationship between the number of medical examinations and cognitive function. The results indicate that in panel A, the estimated βa from the linear regression model is -0.55, while in panel B, the estimated βb from the mixed-effects model is -0.57. This suggests that after controlling for individual differences, the LMM model more accurately captures the downward trend in cognitive function, demonstrating a better fitting effect. Specifically, the LMM model better accounts for the variability between individuals and thus provides a more precise description of the relationship between the number of medical examinations and cognitive function scores. Therefore, the subsequent analysis will focus on the LMM model to further investigate the relationship between the number of medical examinations and cognitive function.

Fig. 9.

Comparison of fitted trends by visit times using linear model and lmm (scatter plot). Note: Panel A shows the linear regression (red line), while Panel B shows the LMM model (blue line); the gray points represent the original observed values, and the numbers indicate the estimated slope (β value) of the models

Analysis of participants with ≥ 2 examinations

We constructed both a base model and an interaction model using LMM to assess the associations of lifestyle factors, physiological indicators, and oral health status on cognitive function.

In a multivariable linear regression model without time factors, the overall trends indicated that older age and higher body weight were associated with lower cognitive scores, whereas greater height and higher education level were associated with higher cognitive scores. Regarding oral health, denture use showed a strong negative association with cognitive score, and both dental caries and tooth loss were also negatively associated with cognitive performance. Among metabolic indicators, elevated LDL-C and elevated FBG were each associated with lower cognitive scores; the other covariates (including TC, TG, HDL-C, marital status, smoking, alcohol consumption, waist circumference, BMI, physical activity, and hypertension) were not statistically significant. The specific coefficients and 95% confidence intervals for each variable are presented in Table 5.

Table 5.

Results of the basic model (without adjustment for time)

Variable	Coefficient	Standard Error	z	P>|z|	Lower bound of 95% CI	Upper bound of 95% CI
Age	-0.063	0.01	-6.644	<0.001	-0.082	-0.045
Marital Status	-0.002	0.076	-0.027	0.978	-0.152	0.148
Educational Level	0.569	0.098	5.819	<0.001	0.377	0.761
Smoking Status	0.048	0.398	0.119	0.905	-0.733	0.828
Drinking Status	0.038	0.308	0.123	0.902	-0.566	0.642
Exercise	-0.051	0.047	-1.068	0.286	-0.144	0.042
Dentures	-1.259	0.125	-10.105	<0.001	-1.503	-1.015
Missing Teeth	-0.321	0.144	-2.235	0.025	-0.602	-0.039
Caries	-0.828	0.187	-4.425	<0.001	-1.195	-0.462
TG	-0.025	0.127	-0.195	0.845	-0.273	0.223
LDL-C	-1.063	0.145	-7.327	<0.001	-1.347	-0.778
HDL-C	0.043	0.048	0.91	0.363	-0.05	0.137
TC	-0.237	0.147	-1.613	0.107	-0.526	0.051
FBG	-0.66	0.126	-5.224	<0.001	-0.908	-0.413
Hypertension	-0.081	0.116	-0.699	0.484	-0.309	0.147
Height	0.043	0.013	3.2	0.001	0.017	0.069
Waist	0.012	0.01	1.206	0.228	-0.007	0.031
Weight	-0.035	0.016	-2.095	0.036	-0.067	-0.002
BMI	0.185	0.187	0.989	0.323	-0.182	0.553

In the multivariable linear regression model including time and its interaction term, we focused on interpreting the baseline and interaction associations. The baseline main effect represents the influence of each variable on the outcome variable at baseline.

At baseline, the model showed that both older age and taller height were positively associated with cognitive score; in terms of oral health, having dental caries was related to a lower cognitive score; and metabolically, a higher LDL-C level was associated with poorer cognitive performance. Additionally, engaging in regular exercise and having a higher FBG were related to lower baseline cognitive scores. With regard to time interactions, the education level × time and exercise × time interactions were positive, suggesting that higher education and regular exercise slowed the rate of cognitive decline over time. In contrast, the interactions of tooth loss, denture use, hypertension, and age with time were negative, indicating that these factors were associated with a steeper decline trajectory. The LDL-C × time interaction was positive, implying that the adverse impact of higher LDL-C on cognition weakened over time, and the BMI × time interaction showed a marginal statistical significance. The detailed coefficients and 95% confidence intervals are provided in Table 6.

Table 6.

Interaction model results (considering time factor)

Variable	Coefficient	Standard Error	z	P>|z|	Lower bound of 95% CI	Upper bound of 95% CI
Age	0.038	0.016	2.446	0.014	0.008	0.069
Marital status	-0.061	0.076	-0.807	0.420	-0.210	0.088
Educational level	0.028	0.155	0.179	0.858	-0.276	0.332
Smoking status	0.115	0.392	0.292	0.770	-0.654	0.884
Drinking status	0.019	0.304	0.063	0.949	-0.576	0.614
Exercise	-0.355	0.081	-4.373	<0.001	-0.515	-0.196
Dentures	-0.195	0.213	-0.916	0.360	-0.614	0.223
Missing teeth	0.671	0.243	2.758	0.006	0.194	1.147
Caries	-0.526	0.188	-2.794	0.005	-0.896	-0.157
TC	-0.121	0.145	-0.835	0.404	-0.405	0.163
TG	-0.027	0.124	-0.220	0.826	-0.270	0.216
LDL-C	-1.166	0.239	-4.871	<0.001	-1.635	-0.697
HDL-C	-0.029	0.047	-0.615	0.539	-0.121	0.063
FBG	-0.550	0.125	-4.411	<0.001	-0.794	-0.306
Hypertension	0.270	0.191	1.411	0.158	-0.105	0.644
Height	0.036	0.013	2.714	0.007	0.010	0.062
Weight	-0.030	0.016	-1.817	0.069	-0.061	0.002
Waist	0.006	0.010	0.646	0.518	-0.013	0.025
BMI	-0.103	0.246	-0.420	0.674	-0.585	0.378
time	1.381	0.426	3.244	0.001	0.547	2.215
time*Age	-0.029	0.005	-5.864	<0.001	-0.039	-0.020
time*Educational level	0.239	0.047	5.060	<0.001	0.146	0.331
time*Exercise	0.156	0.028	5.637	<0.001	0.102	0.210
time*Dentures	-0.340	0.068	-4.980	<0.001	-0.474	-0.206
time*Missing teeth	-0.314	0.079	-3.998	<0.001	-0.468	-0.160
time*LDL-C	0.178	0.065	2.741	0.006	0.051	0.305
time*Hypertension	-0.155	0.060	-2.585	0.010	-0.273	-0.037
time*BMI	0.119	0.064	1.852	0.064	-0.007	0.244

The symbol “*” indicates the interaction term between time and the respective covariate

Analysis based on participants who underwent three or more health examinations

In this study, we performed LMM analyses on datasets stratified by the number of health examinations, in order to explore the associations of various factors on cognitive scores and the robustness of their interactions with time. The analysis included comparisons of participants with at least two, three, and four examinations to test the stability of our conclusions. Data from participants with five or more examinations were excluded from further analysis because the sample size was markedly reduced to 795 participants and model instability was observed.

Sample size and model fit

The subsamples with at least 2, 3, and 4 health examinations included 7,422, 3,906, and 1,983 participants, respectively. Although sample size decreased as the number of examinations increased, both the base and interaction models converged stably. AIC indicated a marked improvement in model fit with additional examinations (Table 7): for the base model, AIC declined from 43,539.27 (≥ 2 exams) to 23,314.58 (≥ 3 exams) and 11,652.10 (≥ 4 exams); for the interaction model, from 43,263.70 to 23,193.15 and 11,572.70. In all subsamples, the interaction model outperformed the base model (AIC reductions of 275.57, 121.43, and 79.40, respectively), suggesting that inclusion of time×exposure interaction terms improves model fit.

Table 7.

AIC comparison among models with varying numbers of health checkups

Model	≥ 2 Checkups	≥ 3 Checkups	≥ 4 Checkups	Trend summary
AIC of the base model	43539.27	23314.58	11652.10	With the increase in the number of physical examinations, the AIC of the model decreased significantly, indicating improved model fit.
AIC of the interaction model	43263.70	23193.15	11572.70	The interaction model consistently outperformed the base model, demonstrating better goodness of fit.

Analysis of variable associations and trends

Using data from participants with two, three, and four examinations, we constructed baseline and interaction models to comprehensively evaluate the main associations of variables and their interactions with time.

Trends in the baseline model

Across data sets with increasingly stringent inclusion thresholds (requiring participants to have ≥ 2, ≥3, or ≥ 4 examinations), the baseline model yielded consistent trends. The direction of association for each variable with cognitive score remained essentially the same across these samples. Greater age was associated with lower cognitive scores, and the strength of this association increased from the ≥ 2 examinations sample to the ≥ 3 examinations sample, then plateaued. Height had a modest positive correlation with cognitive score, which became slightly stronger under stricter inclusion criteria, and education level was a consistently strong positive correlate of cognitive performance. In oral health factors, denture use maintained a strong negative association with cognitive score, whereas the negative associations of tooth loss and dental caries gradually weakened and were no longer evident in the ≥ 4 examinations sample. For metabolic factors, LDL-C continued to show a pronounced negative association with cognitive score, while the inverse association between FBG and cognitive score attenuated and was nearly absent in the most stringent sample. Although point estimates for HDL-C shifted slightly upward, overall HDL-C showed no clear association with cognitive score. BMI showed no significant overall association, with only a slight positive association emerging in the ≥ 4 examinations sample. Neither the lifestyle factors (smoking, exercise) nor hypertension demonstrated any significant or stable associations with the outcome in the analyses. The full association estimates and their 95% confidence intervals are provided in Supplementary Table 8.

Trends in the interaction model

In progressively stricter inclusion samples (requiring ≥ 2, ≥3, or ≥ 4 examinations), the time-interaction model results were highly consistent. Higher baseline age was associated with a faster decline in cognitive score over time (age × time < 0), whereas a higher education level was associated with a slower decline (education × time > 0). Physical exercise also showed a protective trajectory (exercise × time > 0), with a slightly stronger association size in the most stringent sample. Regarding oral health, both denture use and tooth loss were associated with a more pronounced decline over time (negative interaction terms). For metabolic factors, the LDL-C × time interaction was positive in the larger samples but diminished as the inclusion criteria became more stringent; notably, the FBG × time interaction turned negative only in the ≥ 4 examinations sample, suggesting a threshold (fragile) association. The hypertension × time interaction was negative in the larger samples but was not consistently retained, while the BMI × time interaction trended positive but remained statistically non-significant across samples, despite a marginal trend. Complete association estimates with 95% confidence intervals are presented in Supplementary Table 9.

Discussion

In a six-year follow-up, this study investigated the dynamic associations of educational attainment, physical activity, oral health, and metabolic factors on cognitive function among older adults in southern China. These findings provide potential strategies for screening and preventing cognitive impairment in community-based older populations. The results suggest that higher education, regular physical exercise, good oral hygiene, and stable metabolic status may play important roles in preventing or delaying cognitive decline in older adults. This is consistent with recent perspectives that multifaceted lifestyle interventions reduce dementia risk [12].

High educational attainment can increase cognitive reserve, thereby lowering the risk of dementia [32, 33]. Cognitive reserve refers to the phenomenon in which an individual does not exhibit clinical cognitive symptoms despite imaging or histological evidence of pathological changes consistent with progressive dementia [14]. Although some studies suggest that cognitive reserve cannot delay dementia onset or death [34], the mechanisms by which it affects disease onset remain unclear. Unlike prior studies that dichotomized education or only used years of schooling [33, 35], we stratified education into multiple levels (illiterate, primary, junior secondary/technical, senior secondary/vocational) and observed a clear gradient association. Multivariable logistic regression and Cox proportional hazards models indicated a dose–response relationship, with the association between education and cognitive impairment increasing as education levels. In our regression analyses, we also found that the association between higher education and reduced cognitive impairment risk appeared stable over time in our sample, though this may reflect long-standing social and cognitive advantages rather than direct causality. However, some studies have reported that individuals with higher cognitive reserve experience faster cognitive decline after dementia onset [14], suggesting that the mechanisms through which education influences disease progression are complex and warrant further investigation.

Regular physical activity may be associated with better cognitive outcomes in older adults. It can improve cognitive function by promoting neural plasticity and maintaining neurovascular unit homeostasis, thereby exerting neuroprotective relationships and delaying cognitive decline [12, 36–38]. In multivariable logistic regression and Cox models, we observed that higher exercise frequency significantly reduced the risk of cognitive impairment, consistent with a large population study linking greater activity to lower dementia risk [39]. However, in the baseline LMM, the main association of exercise was not significant, whereas in a model including a time-by-exercise interaction, a consistently significant positive association emerged. Based on this finding, we believe that regular exercise may have a positive association on cognitive function; however, the cross-sectional or baseline association may be influenced by reverse causation and healthy volunteer bias, as individuals in a prodromal disease stage often reduce physical activity, which could bias the baseline exercise–cognition relationship. This has been supported by evidence in long-term cohort studies: when physical activity is measured ≥ 10 years before dementia diagnosis, or when dementia cases occurring in early follow-up (for example, within the first 2–10 years) are excluded, the association between exercise and dementia risk often markedly attenuates or disappears, indicating that reverse causation from prodromal decline is a major factor [40–42]. Although multiple covariates were adjusted for, residual confounding by comorbidities (such as cardiovascular disease and depression) and time-dependent confounders [12, 43, 44] may remain, potentially diminishing the observed association of exercise in the baseline model, whereas its time-varying interaction and cumulative long-term associations become more apparent. Moreover, the relationship of exercise on cognitive function requires sustained exposure and intensity; a one-time measure may not capture this. The time×exercise interaction indicates that participants who sustain or increase their exercise over time experience slower cognitive decline, which is consistent with the long-term cognitive benefits of exercise [45]. In addition, self-reported exercise is prone to measurement error and recall bias compared to accelerometer-based measures [46], which can affect association estimates; repeated measurements can better reveal the true signal. Nonetheless, reverse causality cannot be completely excluded [42]. Selection and survival biases inherent in older cohorts may also influence the observed association size [47]. Therefore, we believe exercise may benefit cognitive health, but these observational findings should be interpreted with caution. Supporting this caution, a long-term cohort excluding prodromal dementia found no significant association between exercise and dementia risk [40], and a randomized trial of moderate-to-high intensity exercise in patients with mild-to-moderate dementia failed to improve cognition [48]. Thus, the causal relationship between exercise and cognitive decline remains to be confirmed by interventional trials.

Although most evidence comes from observational studies [49–51], poor oral health is thought to potentially increase the risk of cognitive impairment. Our study found that tooth loss and denture use were associated with a higher risk of cognitive impairment, and the LMM analysis showed that the negative association of these factors on cognitive scores accumulated over time. Our findings suggest an association between poor oral health and greater cognitive decline over time, though this may reflect bidirectional influences or shared underlying risk factors rather than a direct causal pathway. First, older adults with cognitive decline often struggle to maintain oral hygiene, which may in turn raise the risk of dental disease [49], implying that poor oral status could be a consequence of cognitive decline. Second, although our study adjusted for many covariates, residual confounding by comorbidities such as diabetes, stroke, and depression may remain; these conditions affect both oral health and cognition, and may through inflammation, vascular damage, and neuropsychological mechanisms increase tooth loss risk and independently promote cognitive decline, thereby influencing the associations between tooth loss, denture use, and cognitive impairment [52, 53]. Furthermore, denture use is partly a remedial measure for tooth loss. A meta-analysis found that each additional missing tooth slightly increases the risk of cognitive impairment, but this association was not apparent in those with existing dentures, suggesting that timely denture replacement may mitigate the adverse cognitive association of tooth loss [54]. Therefore, we speculate that the association between the “denture” variable and cognitive impairment in our study essentially reflects severe tooth loss and the underlying long-term poor oral health, rather than a direct negative association of dentures themselves. On the other hand, in the baseline LMM model, dental caries were associated with cognitive impairment, but this was not significant in the LMM interaction model. This suggests that dental caries represent a treatable, reversible, and fluctuating condition (for example, fillings, silver diamine fluoride, and remineralization can rapidly alter lesion status during follow-up), and a one-time measure is easily affected by cycles of “treatment–remission–relapse,” which attenuates its apparent association [55, 56]. Moreover, these recurrent fluctuations in caries status may reflect the fact that early cognitive decline leads to weakened oral self-care, resulting in frequent oral health problems, consistent with the reverse causation notion that early cognitive decline leads to reduced oral self-care ability.

In terms of metabolic factors, studies have shown a complex relationship between blood lipid levels and cognitive impairment. A French study of 7,470 noninstitutionalized individuals aged 65 and older found that higher TC, TG, and LDL-C were significantly associated with the incidence of dementia [57]. Both cross-sectional and longitudinal studies have demonstrated that elevated TG or LDL-C levels are associated with cognitive impairment [3, 58]. However, other studies have reported inconsistent findings regarding the LDL-C–dementia association, with some showing positive associations, some none, and some inverse associations [59–61]. In our study, multivariable binary logistic regression indicated that elevated TC and LDL-C were risk factors for cognitive impairment; however, Cox proportional hazards regression did not find statistically significant differences. Further RCS analysis suggested a potential nonlinear trend between TC/LDL-C and cognitive impairment, but this did not reach statistical significance. TC remained non-significant in both the base and interaction models of the LMM. Previous research suggests that higher TC in midlife is associated with an increased long-term risk of dementia, whereas in older age the relationship between TC and dementia often weakens, becomes U-shaped, or even reverses direction [62–64]. Our results largely align with this age-dependent pattern: in cohorts comprised mainly of older adults, TC still showed some association in baseline/cross-sectional analyses, but the association weakened to non-significance in Cox regression and LMM, suggesting that the relationship between TC and cognitive impairment in later life may be influenced by time dependency and nonlinearity. In the base LMM, LDL-C consistently had a significant negative association, but this association gradually diminished over time; correspondingly, the time×LDL-C interaction was positive and significant at the second and third measurements, but was no longer significant at the fourth measurement. A possible explanation for this phenomenon is that individuals with high LDL-C start and continue lipid-lowering treatment (e.g. statins) earlier, which slows the subsequent slope of cognitive decline [65], leading to a positive time×LDL-C in the second and third measurements; once such treatment becomes widespread and stable, this interaction association disappears. At the same time, we must acknowledge that discrepancies between the logistic, Cox, and LMM results may not only be due to time or interventions, but also to differences in model estimators, measurement error and regression dilution, competing risks, and differing model assumptions or nonlinear specifications [66–69]. This suggests that further validation is needed.

However, the relationship between HDL-C and cognitive impairment is more complex. A study in California, USA, found a U-shaped association between both low and high HDL-C levels and increased dementia risk [9]. The ASPREE study indicated that in healthy older adults, high HDL-C was associated with increased dementia risk, and that the risk gradient became steeper with advancing age [8]. A 2024 analysis based on NHANES also supported a U-shaped association between high HDL-C and risk of cognitive impairment [10]. Our findings suggest that HDL-C and cognitive impairment in older individuals are not related in a monotonic or linear fashion. We found that extremely high or extremely low HDL-C was associated with a higher incidence of cognitive impairment, consistent with reports from recent large cohorts in the oldest-old subgroup [9]. However, in our Cox proportional hazards regression, the risk of cognitive impairment was significantly lower in the low HDL-C group compared to the moderate group; this may be due to differences in how models respond to variable ranges. Extremely high HDL-C can lead to dysfunctional HDL (such as impaired reverse cholesterol transport, post-translational modification of apolipoproteins, reduced antioxidant capacity, etc.) and can induce neurotoxic processes such as vascular endothelial damage and cholesterol crystallization [70]; whereas very low HDL-C is often associated with metabolic dysregulation and malnutrition [71]. KM curves and RCS in this study also suggested a complex nonlinear risk trend, but the RCS did not reach statistical significance, indicating that larger samples are needed for verification. While some evidence suggests both high and low HDL-C may be detrimental to cognitive health, the non-significant nonlinear trends in our data suggest this relationship requires further investigation. In conclusion, as this study is based on observational data, associations between lipid indices and cognition may reflect overall healthy lifestyle factors rather than direct causal relationships, and relevant clinical and public health recommendations require further validation through interventional studies.

In our study, multivariable binary logistic regression, Cox proportional hazards regression, KM analysis, and the base LMM all indicated that elevated FBG was associated with an increased risk of cognitive impairment. However, in the LMM interaction model, a significant negative correlation between FBG and cognitive score was observed only at the fourth measurement (β = -0.408, P = 0.002), and this “delayed association” may reflect the cumulative relationship of hyperglycemia on cognitive damage. Duration of diabetes is also an important factor for cognitive risk: Cao et al. reported in a meta-analysis that patients with disease duration < 5 years had a significantly increased risk of dementia [72], further supporting the notion that chronic hyperglycemia causes cumulative damage to cognitive function. Mechanistically, chronic hyperglycemia and insulin resistance can lead to oxidative stress, neuroinflammation, microvascular damage in the brain, and deposition of advanced glycation end products, ultimately resulting in progressive loss of neurons in critical brain regions such as the hippocampus and impairing cognitive function [72, 73]. Furthermore, RCS results showed a nonlinear relationship between FBG and cognitive risk (with risk peaking at 7–8 mmol/L), consistent with a threshold association: studies have indicated that even mild hyperglycemia can increase dementia risk [74], and glycemic variability is a significant independent risk factor for cognitive decline [75]. Another study observed an inverted U-shaped association between blood glucose and cognition: Liu et al. reported that among non-diabetic individuals, cognitive function was optimal at an FBG level of about 4–6 mmol/L, implying that adverse associations of hyperglycemia appear only beyond this threshold [76].

On the other hand, we found that hypertension was not significantly associated with cognitive impairment in either multivariable logistic regression or Cox proportional hazards regression. In the LMM analysis with an interaction model, the time×hypertension interaction term was significantly negatively associated with cognitive impairment in the first two examinations, but this associations was no longer significant by the fourth examination. We cannot exclude the possibility that this finding is related to antihypertensive treatment. As a result, our study cannot clearly define the relationship between hypertension and cognitive impairment, a finding that aligns with previous inconsistent findings [77, 78]. Apart from the primary results, other covariates did not show stable and significant associations across different models, and their associations directions were occasionally inconsistent. For example, we observed that widowed individuals had a lower risk of cognitive impairment than those classified as unmarried. This association may be affected by residual confounding arising from heterogeneity within the “unmarried” category (e.g., never married, cohabiting, in transition to remarriage) as well as differences in social support and living arrangements, and thus warrants cautious interpretation and further investigation.

However, this study still has certain limitations. First, measurement error and misclassification bias may affect the accuracy of our results. For example, physical activity was self-reported and may be subject to recall bias, and the lack of a graded scale for oral health indicators may lead to measurement error. Second, reverse causality and selection bias must be considered: cognitive decline can influence physical activity and oral health, and older cohorts are prone to selection and survival biases. For example, healthier and more mobile individuals may be more likely to return for follow-up examinations, which could further bias the association estimates; moreover, response or participation rates could not be calculated. We acknowledge that excluding older participants whose hearing impairments severely affected data collection despite assistive measures (e.g., hearing aids) may have introduced selection bias. Although our models included multiple covariates, residual confounding remains possible, as comorbidities, medication use, lifestyle, socioeconomic status, and genetic factors may not have been fully controlled. Overlap among oral health indicators makes it impossible to determine their independent contributions, and the inconsistent associations of lipid indicators suggest potential instability of these associations; we therefore take a cautious stance on the causal implications of HDL-C. Given the practical constraints of the study, we assessed dental caries status solely by visual-tactile examination and did not perform any radiographic evaluations (such as X-rays). Cognitive outcome classification was relatively coarse, relying solely on MMSE screening and lacking cognitive subtypes or biomarker support, which limits inferences about underlying pathophysiological mechanisms. We also acknowledge that our study did not include a comprehensive quantitative assessment of inter-examiner consistency among all examiners. Finally, this was a single-region retrospective cohort study, and the generalizability of the results may be affected by regional differences.

Conclusion

This study found that among older adults in rural areas, modifiable factors such as educational level, physical activity, oral health, and metabolic status were closely associated with the risk of cognitive impairment. Improving educational attainment, maintaining moderate physical activity, strengthening oral hygiene, and controlling metabolic indicators may help delay cognitive decline. We recommend implementing comprehensive, multifaceted interventions targeting older adults to promote cognitive health. Since these conclusions are based on observational data, further prospective studies and clinical trials are needed to validate them. In practice, attention should be focused on older adults with high-risk characteristics; it is important to balance the benefits of interventions against individual circumstances and to explore the associations of interventions in different subgroups and types of cognitive impairment. We anticipate that future research will deepen the understanding of the causal relationships between these factors and cognitive aging, providing evidence for the prevention and treatment of cognitive impairment in older adults.

Abbreviations

TG

Triglycerides

LDL-C

Low-density lipoprotein cholesterol

TC

Total cholesterol

HDL-C

High-density lipoprotein cholesterol

FBG

Fasting blood glucose

MMSE

Mini-mental state examination

AIC

Akaike information criterion

RCS

Restricted cubic spline

LMM

Linear mixed-effects models

MCAR

Missing completely at random

MICE

Multiple imputation by chained equations

SMD

Standardized mean differences

SD

Standard Deviation

T2DM

Type 2 diabetes

Authors’ contributions

ZWK and HRY conceived the study, performed data analysis, and contributed to manuscript writing. LNY, SL, QRX, ZXL, and ZJQ are responsible for data organization and analysis. HF and HRJ supervised the research, secured funding, and critically revised the manuscript. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Key Research and Development Program of Guangxi Zhuang Autonomous Region (Guike AB24010139), the Guangxi Natural Science Foundation (2025GXNSFHA069077), the National Natural Science Foundation of China (82360092), and a self-funded research project from the Administration of Traditional Chinese Medicine, Guangxi Zhuang Autonomous Region (GXZYC20240401).

Data availability

The raw data are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study adhered to the ethical principles outlined in the Declaration of Helsinki and followed the research protocols of the First Affiliated Hospital of Guangxi Medical University. The study received approval from the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University, with the approval number 2024-E706-01. All participants voluntarily joined the study and signed a written informed consent form.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.