Age- and sex-specific modifiable risk factor profiles of dementia: evidence from the UK Biobank

Dementia constitutes a worldwide concern. To characterize the age- and sex-specific modifiable risk factor profiles of dementia, we included 497,401 UK Biobank participants (mean age = 56.5 years) without dementia at baseline (2006–2010) and followed them until March 2021. Cox proportional hazard models were used to estimate the age- and sex-specific hazard ratios (HRs) of incident dementia associated with socioeconomic (less education and high Townsend deprivation index), lifestyle (non-moderate alcohol intake, current smoking, suboptimal diet, physical inactivity, and unhealthy sleep duration), and health condition factors (hypertension, diabetes, cardiovascular diseases, and depressive symptoms). We also calculated the population attributable fractions (PAFs) of these factors. During follow-up (mean = 11.6 years), we identified 6564 dementia cases. HRs for the risk factors were similar between the sexes, while most factors showed stronger associations among younger participants. For example, the HRs of smoking were 1.74 (95% CI: 1.23, 2.47) for individuals aged < 50 years, and 1.18 (1.05, 1.33) for those aged ≥ 65 years. Overall, 46.8% (37.4%, 55.2%) of dementia cases were attributable to the investigated risk factors. The PAFs of the investigated risk factors also decreased with age, but that for health condition risk factors decreased with lower magnitude than socioeconomic and lifestyle risk factors. The stronger associations and greater PAFs of several modifiable risk factors for dementia among younger adults than older participants underscored the importance of dementia prevention from an earlier stage across the adult life course.


Introduction
Worldwide, ~55.2 million people had prevalent dementia in 2019, and the number was expected to triple by 2050 [1,2]. Recent evidence also suggested that from 1990 to 2019, the incidence of dementia has increased in midlife and younger older adults [3]. The expanding dementia population means heavier economic burdens for individuals and health systems [4]. Since there are currently limited effective treatments for dementia [5,6], it is crucial 1 3 to identify the risk factors, especially the modifiable ones, of dementia.
According to the Lancet Commission on dementia prevention, intervention, and care, ~40% of dementia cases could be prevented or delayed by targeting 12 modifiable risk factors [7], including less education at an early age, midlife hearing loss, traumatic brain injury, hypertension, unhealthy alcohol intake, obesity, and smoking, and depression, social isolation, physical activity, diabetes, and air pollution in later life. Emerging evidence also suggested suboptimal dietary patterns [8] and unhealthy sleep duration [9,10] as potential risk factors. Subgroup analyses in previous studies suggested that the associations of risk factors with dementia might differ by sex and age [9,11,12], but few studies have characterized these disparities in sufficient detail based on a large cohort with long-term follow-up. Therefore, we characterized the ageand sex-specific modifiable risk factor profiles of dementia across age and sex groups in the UK Biobank (UKB), a nationwide cohort study in the UK.

Study population
UKB is a nationwide cohort study in the UK [13]. From 2006 to 2010, UKB recruited ~ 500,000 UK residents, as described on its website (http:// www. ukbio bank. ac. uk/ resou rces/). Ethical approval was granted by the North West-Haydock Research Ethics Committee (REC: 16/ NW/0274). Participants signed informed consent before data collection, and individuals who withdrew their consent were not included in the analyses.
Among 498,713 UK Biobank non-pilot participants, we excluded 158 participants who were lost to followup, 909 with baseline prevalent dementia, and 205 who developed dementia within the first two years. Finally, 497,401 middle-aged and older adults were included in analysis. We categorized participants with every 5-year increment in baseline age and combined the youngest two groups into one to ensure sufficient statistical power. Thus, we assigned participants into five age groups (40-< 50, 50-< 55, 55-< 60, 60-< 65, or ≥ 65 years) and two sex groups (female or male). We selected two socioeconomic, five lifestyle, and four health condition risk factors [8,[14][15][16][17][18][19] that are relatively prevalent in the UKB and might be directly or indirectly modified. We further defined the socioeconomic (0-2), lifestyle (0-5), and health condition risk scores (0-4). Each risk factor was scored as 1 for unhealthy status, and a higher score indicates a higher risk in the corresponding category.

Socioeconomic risk factors
We selected education levels lower than college or university degree and high Townsend deprivation index (TDI, above median value) as socioeconomic risk factors, according to definitions in previous studies [7,[14][15][16]. At recruitment, participants were asked about the highest degree they have achieved. TDI is an area-based proxy measure of socioeconomic status composed of ownership and employment as a state of observable and demonstrable disadvantage relative to the local community [20], and a higher TDI indicates a higher level of deprivation.

Lifestyle risk factors
We defined lifestyle risk factors using data from a webbased questionnaire at baseline [21,22] as in a previous study [23]. Non-moderate alcohol intake [17] was defined as zero alcohol intake or daily intake over 28 g (two drink equivalences) for men and over 14 g (one drink equivalence) for women [24]. Smoking status was self-reported (never, former, or current) and current smoking is defined as a risk factor. According to the American Heart Association (AHA) recommendations [25], suboptimal diet was defined as meeting less than 4 recommendations of 7 commonly eaten food groups related to cardiometabolic health and dementia: fruits (≥ 3 servings/day), vegetables (≥ 3 servings/ day), fish (≥ 2 servings/week), processed meat (≤ 1 serving/ week), unprocessed red meat (≤ 1.5 servings/week), whole grains (≥ 3 servings/day), and refined grains (≤ 1.5 servings/ day). Physical inactivity was also defined according to AHA recommendations [26] (≥ 150 min of moderate activity per week or 75 min of vigorous activity per week or an equivalent combination of engaging in moderate physical activity at least 5 days a week or vigorous activity once a week). According to the National Sleep Foundation, unhealthy sleep duration was defined as ≤ 6 or ≥ 10 h/d for adults aged 40-64 years and ≤ 6 or ≥ 9 h/d for adults aged 65 years or above [27]. Of note, different from the risk factor profiles of other diseases, such as CVD [28], diabetes [29], and heart failure [30], obesity was not included because of inconsistent evidence regarding its association with dementia [31,32].

Health condition risk factors
The UKB incorporated a multi-source health condition identification system. For diabetes, we used self-reports and linkage to electronic health records (EHRs) to identify prevalent cases at baseline. In the linkage to EHRs, we used any code mapped to 3-character ICD-10 E10-E14 before baseline. Similarly, for hypertension and cardiovascular diseases, we used self-reports and linkage to EHRs, and the ICD-10 codes were I10 and I15 for hypertension and I20-25 and I60-I69 for cardiovascular diseases. In the two-item Patient Health Questionnaires (PHQ-2) [33], participants were asked about their frequency of depressed mood and anhedonia in the past 14 days. The responses "not at all," "several days," "more than half the days," and "nearly every day," were scored as 0, 1, 2, and 3, respectively. A PHQ-2 score of 3 or more indicated depressive symptoms. All these conditions have been suggested as risk factors for dementia in previous research [7,12,[34][35][36][37][38].

Dementia ascertainment
In this study, the primary outcome was incident dementia. We ascertained incident dementia using the algorithms developed by UKB [39], which were generated using linkage to electronic health records (EHRs), including primary care, hospital admissions and the death registry, coded with the ICD-10 and Read coding system. With a similar strategy, we identified Alzheimer's Dementia (AD), vascular dementia (VaD), and other types or undefined dementia. The linkages were updated through March 2021.

Statistical analyses
Baseline characteristics were presented by age group and by sex. Continuous variables were displayed as means (standard deviations, SDs), and categorical variables were shown as numbers (percentages). Cox proportional hazard models were used to estimate the hazard ratios (HRs) and confidence intervals (CIs) for incident dementia associated with the modifiable risk factors. Person-time was calculated from baseline to the diagnosis of dementia, date of death, the date of loss to follow-up, or the end of follow-up, whichever came first. We conducted stratified analyses to estimate the age-and sex-specific HRs. The proportionality assumption was verified by Schoenfeld residual methods. In this study, risk factors were mutually adjusted in the Cox models, and we further adjusted the models for age (not when assessing age differences), sex (not when assessing sex differences), and race. P for interactions were calculated by entering the cross-product terms of the main effect terms and sex or continuous age into the main models. We also tested the threeway interactions among age, sex, and risk factors.
In the secondary analyses, we investigated the associations of the three composite risk scores (as continuous variables) with incident dementia, with adjustments for age, sex, and race, as mentioned above. We further calculated the adjusted population-attributable fractions (PAFs) and 95% CIs of dementia attributable to each risk factor [40]. The PAFs estimated the percentage of cases of a disease (in this case, dementia) that would be prevented if the particular risk factors were eliminated [41].
To test the robustness of the main findings, we performed the following sensitivity analyses: (1) we further adjusted the models for APOE genotype (only available for 444,024 participants, categorized as "ε4", "non-ε4", and "missing") and family history of dementia, respectively; (2) we further excluded 835 participants who developed dementia within the first five years to further account for potential reverse causality; (3) we excluded participants with baseline CVD (n = 28,661 ) and repeated the primary analyses; (4) we performed competing risk analysis treating mortality of other causes as a competing event [42]; (5) we further adjusted the models for BMI categories (< 25 kg/m 2 , 25-30 kg/m 2 , and > 30 kg/m 2 ); (6) to validate the definition of non-moderate alcohol intake, we assessed the relation of alcohol intake to dementia using penalized splines by sex; (7) to validate the definition of smoking, we further categorized smoking status as current, former, and never. (8) we calculated the relative excess risk due to interaction (RERI) to assess the interactions in additive scales.
Statistical analyses were performed using R 3.6.0 and missing values were imputed using the "mice" package [43]. To account for multiple testing, we reported the original two-sided P-values and adopted the Benjamin-Hochberg methods when interpreting the results [44]. Corrected P-values reaching the significance level of 0.05 indicated statistical significance.

Participant characteristics
Of the 497,401 participants, the mean (SD) age at baseline was 56.5 (8.1) years, and 270,715 (54.4%) were female (Table 1). Older participants were more likely to be less deprived, less educated, non-current smokers, physically inactive, have unhealthy sleep duration, hypertension, diabetes, and CVD and less likely to have suboptimal diet and depressive symptoms. Female participants were less likely to be current smokers and to have hypertension, diabetes, and CVD and more likely to have depressive symptoms.
Among the overall participants, most risk factors were associated with a significantly higher risk of later-life

Table 2
Hazard ratios (HRs) and 95% confidence intervals (CIs) for associations between modifiable risk factors and dementia by age and sex Cox proportional models were adjusted for age (not in the age-stratified models), sex (not in the sex-stratified models), and mutually adjusted for individual risk factors. P-interaction was calculated as the significance of the multiplication term of individual risk factors and sex or age in the overall population Overall Age

Sex-and age-specific modifiable risk factor profiles
The associations between individual risk factors and dementia were similar between the sexes (Table 2). For example, the HR (95% CI) for physical inactivity was 1.13 (1.05, 1.23) in women and 1.18 (1.10, 1.27) in men. Correspondingly, all three risk scores were consistently associated with dementia among women and men ( Fig. 1A and Supplemental Table 1).
In terms of the associations by age groups, we observed different associations between all risk factors but hypertension and incident dementia by age groups (Table 2). High deprivation (P-interaction < 0.001) demonstrated stronger associations in younger participants. The associations for lifestyle risk factors also showed similar trends. For example, non-moderate alcohol intake was associated with an 88% (95% CI, 38-157%) higher hazard of dementia in participants < 50 years compared with a 9% (2-16%) higher hazard in participants ≥ 65 years. The associations of all chronic diseases, including CVD, diabetes, and depressive symptoms but not hypertension, were stronger in younger participants than in older adults (P-interaction < 0.001 for all). Fig. 1 Hazard ratios (HRs) and 95% confidence intervals (CIs) of incident dementia associated with each one-point increment in risk scores (RSs) by age and sex. Socioeconomic risk factors included education below college or university degree and high Townsend deprivation index, with the risk score ranging from 0 to 2. Lifestyle risk factors included nonmoderate alcohol intake, current smoking behavior, suboptimal diet, physical inactivity, and unhealthy sleep duration, with the risk score ranging from 0 to 5. Health condition risk factors included hypertension, diabetes, cardiovascular diseases, and depressive symptom, with the risk score ranging from 0 to 4. Cox proportional models were adjusted for age (not in the agestratified models), sex (not in the sex-stratified models), and risk scores of other categories Consequently, all three composite risk scores demonstrated weaker associations with dementia among older participants (P-interaction < 0.001 for all) (Fig. 1B and Supplemental Table 1). In particular, each unit increment in health condition risk score was associated with 114% (78%, 158%) higher hazard of incident dementia among the youngest participants and 40% (35%, 46%) increased risk among the oldest participants. None of the three-way interactions was significant between the risk factors, age, and sex (P-interactions = 0.138 for TDI, 0.397 for education, 0.074 for smoking, 0.416 for physical activity, 0.204 for diet, 0.377 for alcohol intake, 0.748 for sleeping duration, 0.652 for hypertension, 0.412 for CVD, 0.128 for diabetes, and 0.760 for depressive symptoms).
These associations remained similar in the sensitivity analyses. When further adjusted for the APOE genotype or family history of dementia, we observed similar results as in the primary analysis (Supplemental Tables 2, 3). When we excluded participants who developed dementia within the first five years, the associations remained similar (Supplemental Table 4). The relations remained similar when we excluded participants with baseline CVD (Supplemental Table 5) or further adjusted the models for BMI categories (Supplemental Table 6). By treating mortality of other causes as the competing event, the estimated HRs were not substantially affected (Supplemental Table 7). The "U-curve" associations (P-nonlinear < 0.001) between alcohol intake and dementia for both women and men (Supplemental Fig. 1) confirmed our definition of non-moderate levels of alcohol intake. Compared to current smokers, both never smokers (0.73; 0.67, 0.79) and former smokers (0.76; 0.70, 0.82) had lower risk. The RERIs also confirmed the relatively stronger associations for younger individuals compared with older participants (Supplemental Table 8).
Although the HRs of each risk factor were similar between female and male participants, the PAFs showed sex-specific patterns ( Fig. 2 and Supplemental Table 9). For both female and male participants, lower education was the leading contributor to PAF (~ 15% for both sexes). PAF of non-moderate alcohol intake was higher among men (5.3%; 2.7%, 7.9%) than in women (3.2%; 0.5%, 5.8%). Also, the PAF of health conditions tended to be lower among women compared to men (Fig. 3A).
PAFs of these risk factors declined with age (Figs. 2, 3B). For example, the PAFs for less education were 32.1% (8.9%, 52.1%) in the youngest age group (< 50 years), and 16.4% (10.6%, 22.0%) in the oldest age group (≥ 65 years). In general, the PAFs for health condition risk factors (27.1% in the youngest group and 18.6% in the oldest group) decreased slower than socioeconomic (51.4% in the youngest group and 23.5% in the oldest group) and lifestyle risk factors (56.4% in the youngest group and 12.1% in the oldest group), partially because of the increased prevalence of health conditions with age.

Discussion
In this large cohort of UK adults followed from different adult life stages, we observed no significant sex difference in dementia risk factor profiles, while most modifiable risk factors (high deprivation, non-moderate alcohol intake, current smoking, suboptimal diet, physical inactivity, unhealthy sleep duration, diabetes, cardiovascular diseases, and depressive symptoms) but not less education and hypertension showed significantly stronger associations with dementia among younger adults than in relatively older adults. The corresponding PAFs were also greater among adults of relatively younger ages. Taken together, this study identified a stronger association and greater attributable risk of several modifiable risk factors for dementia among younger adults than that among older individuals.
To the best of our knowledge, this study is one of the few to systematically investigate the age and sex disparities in risk factor profiles of dementia, although previous studies also frequently performed subgroup analyses by age and sex. Generally, the magnitudes of the associations between individual risk factors and dementia in this study were comparable with those in previous studies. For example, we observed that smoking was associated with approximately 30% increased dementia risk in the overall population (HR [95%CI] 1.28 [1.18, 1.40]), which was close to results in a previous meta-analysis [15] (1.37 [1.23, 1.52]). Similarly, we found that diabetes conferred a 1.9-fold hazard of incident dementia (1.91 [1.76, 2.07]), aligning with the pooled results of the previous findings [35].
Notably, we observed significant effect modifications on these associations by age, but not by sex, which echoed previous studies on individual risk factors, such as diabetes [45] and hypertension [46]. Biologically, our findings might represent differences in susceptibilities of the risk factors across adult life stages and the irreversible progression of dementia [47]. Also, considering that most dementia cases in the younger participants might be early onset dementia, which could be etiologically different from late onset dementia, future studies are needed to further confirm the observed age disparities and to understand the pathological mechanisms underlying the differences. For another, underlying methodological explanations might be possible. Older adults who were dementia-free at baseline might be healthier than the general population, and the secular trend of some risk factors such as education (which could better reflect intellectual ability in younger adults) might also contribute to the age disparities.
We also estimated the corresponding PAFs in the study population to represent the proportion of preventable or delayable dementia cases by modifying these risk factors, supposing the observed associations were causal. The overall PAF of the included risk factors in our study (46%) was similar to the Lancet Commission (~ 40%) [7]. In terms of the sex-specific PAFs, dementia cases attributable to health conditions were higher in men than in women, mainly due to the higher prevalence of diabetes, CVD, and hypertension in male participants. It is also noteworthy that the overall PAF was lower in older participants than in younger participants, suggesting that the percentage of incident dementia that could be prevented by controlling these risk factors is higher among younger participants. However, when interpreting our findings on the PAF, the possibility of non-independence could not be eliminated even if we have adopted a model-based strategy mutually adjusted for the risk factors. Also, although the PAF was lower in older age groups, the absolute risk of dementia was exponentially higher for older adults, so they might still benefit the most from eliminating the risk factors in the absolute context. For example, the absolute risk reduction for modifying all lifestyle risk factors was 6.6 cases/100,000 person-years for participants under 50 years, and that for participants aged ≥ 65 years was 43.4 cases/100,000 person-years.
The current study has several strengths. The large wellcharacterized cohort study with large variations in age at baseline enabled us to examine a wide range of modifiable risk factors in diversified age subgroups. The relatively long-term follow-up and low rate of loss-to-follow-up also increased the validity of the study results. However, our findings should be interpreted with caution due to a few limitations. First, the dominant European-ancestry population in our study may not be highly representative of the UK residents, and the incidence rate of dementia in UKB is lower than that in the general population. Second, some factors, such as depressive symptoms and lifestyle factors, might have bidirectional associations and thus confound each other in associations with dementia. Although we mutually adjusted for these variables in the same model, potential residual confounding might still exist. Therefore, our observations could not necessarily imply causal relations. Third, the non-significant association of suboptimal diet might result from the limited number of food groups assessed by the brief assessments of diet, although previous studies also reported inconsistent associations between dietary factors and dementia [48][49][50]. Future studies are needed to confirm the findings with more dementia-specific dietary patterns, such as the Mediterranean diet or the Mediterranean-DASH Diet Intervention for Neurodegenerative Delay (MIND diet). Furthermore, the underreporting of milder dementia cases is possible when patients did not acquire medical assistance and were thus not recorded in the electronic health system. In addition, we did not conduct analyses for subtypes of dementia because of the potentially inaccurate identification of subtypes and the lack of sufficient cases in younger age groups, and further studies are needed to understand whether the observed differences are comparable for subtypes of dementia.

Conclusions
In conclusion, this study supported stronger associations and greater attributable fractions of major modifiable risk factors, especially socioeconomic and lifestyle factors, for B Population attributable fractions for dementia associated with risk factor categories by age groups dementia among relatively younger adults. Our findings of the age-specific risk factor profiles of dementia provided information to guide the development and implementation of prevention measures to reduce the risk of dementia for adults at different life stages and underscored the importance of preventive strategies from an earlier age.