Results 29 072 participants were included (mean age of 72.23 years; 48.54% (n=14 113) were women; and 20.43% (n=5939) were APOE ε4 carriers). Over the 10 year follow-up period (2009-19), participants in the favourable group had slower memory decline than those in the unfavourable group (by 0.028 points/year, 95% confidence interval 0.023 to 0.032, P<0.001). APOE ε4 carriers with favourable (0.027, 95% confidence interval 0.023 to 0.031) and average (0.014, 0.010 to 0.019) lifestyles exhibited a slower memory decline than those with unfavourable lifestyles. Among people who were not carriers of APOE ε4, similar results were observed among participants in the favourable (0.029 points/year, 95% confidence interval 0.019 to 0.039) and average (0.019, 0.011 to 0.027) groups compared with those in the unfavourable group. APOE ε4 status and lifestyle profiles did not show a significant interaction effect on memory decline (P=0.52).

Main outcome measures Participants were followed up until death, discontinuation, or 26 December 2019. Six healthy lifestyle factors were assessed: a healthy diet (adherence to the recommended intake of at least 7 of 12 eligible food items), regular physical exercise (≥150 min of moderate intensity or ≥75 min of vigorous intensity, per week), active social contact (≥twice per week), active cognitive activity (≥twice per week), never or previously smoked, and never drinking alcohol. Participants were categorised into the favourable group if they had four to six healthy lifestyle factors, into the average group for two to three factors, and into the unfavourable group for zero to one factor. Memory function was assessed using the World Health Organization/University of California-Los Angeles Auditory Verbal Learning Test, and global cognition was assessed via the Mini-Mental State Examination. Linear mixed models were used to explore the impact of lifestyle factors on memory in the study sample.

Studies have been conducted to identify factors that might affect memory, including ageing, the apolipoprotein E (APOE) ε4 genotype, chronic diseases, and lifestyle patterns. 6 7 Among these, lifestyle has received increasing attention as a modifiable behaviour because this factor is relatively easily amendable with potential benefits for overall health as well as memory. Studies on the effect of a healthy lifestyle on cognition are increasing. 8 9 10 11 12 However, few studies have focused on its effect on memory, and most were cross sectional in nature, 13 which is insufficient to evaluate the relationship between a long term healthy lifestyle and memory decline. Additionally, those studies did not consider the interaction between a healthy lifestyle and genetic risk; thus, the exact effect of a healthy lifestyle on memory decline in individuals with a higher genetic risk remains unknown. Therefore, longitudinal studies are needed to further investigate the effect of modifiable lifestyle factors on memory decline in older individuals, considering genetic risks, such as the presence of an APOE ε4 genotype.

Although a fundamental function of daily life, memory continuously declines as people age, 1 impairing both life quality and work productivity and increasing the risk of dementia. 2 3 4 However, age related memory decline is not always a prodrome of dementia; memory loss can merely be senescent forgetfulness, which is more prevalent among older individuals, and can be reversed or become stable rather than progress to a pathological state. 5 Thus, the prevention and slowing of age related memory decline in older individuals is paramount. Fortunately, memory decline can be mutable because various contributing factors are reportedly associated with memory loss.

Methods

Study design and participants The COAST is a nationwide, population based, cohort study on dementia in China (ClinicalTrials.gov identifier: NCT03653156). This investigation included individuals from 12 provinces from the north, south, and west of China, representing the geographical characteristics, degree of urbanisation, economic status, dietary patterns, and cultural and social differences in China. We conducted a multistage, stratified, cluster sampling procedure, which considered sex and age distribution, among these provinces. A total of 96 study sites (48 urban and 48 rural) were randomly selected (supplementary 1). Written or oral informed consent was obtained from all participants. The study enrolment procedure began on 8 May 2009 and comprised two sequential phases (lasting about six months together). In phase 1, individuals aged 60 years or older, who gave consent to participate in the study, were listed in the community census or village registry, and had resided there for at least one year preceding the survey date, were included in the study. We excluded participants with a life threatening disease, hearing loss, or vision loss. In phase 2, participants from phase 1 with available APOE genotyping and without mild cognitive impairment or dementia were included. Four rounds of follow-up were conducted—ie, a follow-up any time in 2012, 2014, 2016, and 2019. The end date of the study was 26 December 2019.

Diagnosis of dementia and mild cognitive impairment The aim of this study was to investigate the association between lifestyle and memory in participants with normal cognitive function throughout the study period. Neuropsychological tests were done to determine the cognitive status of the participants at baseline and at each follow-up. For people who progressed to mild cognitive impairment or dementia during the follow-up period, the data after their diagnosis were excluded in the main analyses. Global cognitive function was assessed using the Mini-Mental State Examination,14 which evaluates orientation, attention, calculation, executive function, language, among others. The total score of the Mini-Mental State Examination ranges from 0 to 30, with higher scores representing a better cognitive function. Participants whose score was less than 26 were suspected of cognitive impairment and would receive further evaluation, including physical and other neuropsychological testing. The results and participants’ medical records were reviewed by neurologists who were masked to the genetic results. Dementia was diagnosed on the consensus of at least two neurologists, according to the Diagnostic and Statistical Manual of Mental Disorders (4th edition, text revision criteria).15 Mild cognitive impairment was diagnosed according to the Peterson criteria.16 Having normal cognitive function was defined as having a score of 0 for the Clinical Dementia Rating.17 People for whom a consensus could not be reached were considered controversial cases. These cases would be discussed and diagnosed by an expert group, which included senior neurologists, psychiatrists, and neuropsychologists.

Assessment of lifestyle factors Lifestyle information was collected at baseline and each follow-up based on each individual’s performance over the past year through a healthy behaviour questionnaire (supplementary 2). We assessed lifestyle status by six modifiable lifestyle factors: physical exercise, diet, drinking, smoking, cognitive activity, and social contact. For physical exercise, weekly frequency and total time were collected, and at least 150 min of moderate or 75 min of vigorous activity per week was considered a healthy factor, according to the American guidelines for physical activity in adults.18 For smoking, participants were categorised as smokes current, never (participants who had smoked <100 cigarettes in their lifetime) smokes, or used to smoke (participants who had quit smoking at least three years before). Never or former smoking was deemed a healthy lifestyle factor.19 For alcohol consumption, the current frequency and volume of alcohol consumption were recorded, and individuals were categorised into never drinking (never drank or drank occasionally), low to excess drinking (daily alcohol consumption of 1-60 g), and heavy drinking (daily alcohol consumption >60 g).20 The category of never drinking was deemed a healthy lifestyle factor.21 The remaining three lifestyle factors were deemed healthy based on the top 40% of the population distribution, according to previous studies.22232425 For diet, we recorded the participant’s daily intake of 12 food items (fruits, vegetables, fish, meat, dairy products, salt, oil, eggs, cereals, legumes, nuts, and tea) (supplementary 2).26 For cognitive activity (writing, reading, playing cards, mahjong, and other games) and social contact (participation in meetings or attending parties, visiting friends or relatives, travelling, and chatting online), engagement frequencies were investigated. Ultimately, these factors were deemed healthy when participants consumed appropriate daily amounts of at least 7 of the 12 food items for diet, and they engaged in cognitive activity or social contact at least twice weekly in this study. We evaluated the association of each single lifestyle factor with memory function. For further investigation of the combined effect of lifestyle factors on memory function, we categorised participants into three groups based on the third of the number of healthy lifestyle factors: favourable (4-6 healthy factors), average (2-3), and unfavourable (0-1). To investigate the effect of lifestyle on memory stratified by APOE genotypes, participants were also grouped into people who were APOE ε4 carriers and people who were not carriers.

Outcomes Memory function was measured at baseline and each follow-up using the World Health Organization/University of California-Los Angeles Auditory Verbal Learning Test (AVLT),27 which included measurement of immediate recall, short delay free recall (3 min later), long delay free recall (30 min later), and long delay recognition. During the test, a rater read a list of words consisting of 15 nouns; immediately after its completion, the participant attempted to repeat as many words as possible. The score of immediate recall was 0-60, and the scores of all other tests were 0-15. The standard z scores for the auditory verbal learning test were calculated based on the respective mean and standard deviation test scores. The composite z score for memory function was constructed by averaging the z scores for each test in the auditory verbal learning test.

Covariates Potential confounders were selected based on directed acyclic graphs (supplementary figure 1) depicting the best known relations between the variables in this study. Sociodemographic information obtained at baseline were sex, place of residence (rural or urban), region (north, west, or south), marital status (married, widowed, or divorced/separated/single), years of education (<1 year, 1-6 years, 7-12 years, >12 years), monthly household income (1000-2999, 3000-4999, 5000-10 000, >10000 Chinese Yuan), occupation (manual labourer, office worker, household, or others). Health related covariates were collected at baseline and each follow-up, including age, body mass index and medical illnesses (hypertension, diabetes, hyperlipidaemia, heart attack, head injury, cerebrovascular disease, and depression). We also included the learning effect of each participant as a covariate due to repeated cognitive assessments.

Statistical analysis Baseline characteristics of the analytical sample were summarised across three lifestyle groups as a percentage for categorical variables and mean and standard deviation for continuous variables. Missing values are summarised in supplementary table 1. A multiple imputation chain-equation was used to impute missing data. Data were assumed to be missing at random. Missing continuous variables were imputed using predictive mean matching method, and binary variables were imputed by Logistic regression model. We generated five imputed datasets with 100th iteration, and analysed each dataset separately, then combined their results by use of Rubin’s method. Linear mixed effects models were used in this study, and all statistical assumptions were tested before interpretation of results. The data met the distributional requirements in each case. We examined the effects of different lifestyle profiles, APOE ε4 status, as well as their interactions on longitudinal memory trajectories in participants with normal cognition throughout the study by use of linear mixed effects models. The composite z score of the auditory verbal learning test was the dependent variable. The fixed effects were the following: lifestyle profiles (each respective lifestyle factor as well as the combination); APOE ε4 status; time (follow-up year from baseline); two way interactions of lifestyle with APOE ε4 status and time; and three way interactions between lifestyle, APOE ε4 status, and time. The random effects included intercept and time. Age (centred at 74 years to reduce the correlation between the age and age squared terms), age squared (to allow for the comparison of acceleration in the rate of memory decline between lifestyle groups), baseline memory score, learning effect (by assignment code to the first assessment as 1 versus subsequent assessments coded as 0), and other covariates were also adjusted in this model. In addition, the global cognitive trajectories in the cognitively normal population were evaluated using a linear mixed effects model with global cognition (standard z score of the Mini-Mental State Examination) as dependent variable. Furthermore, we investigated the interaction between lifestyle and age on memory decline. As lifestyle status and some of the covariates changed over time, a time dependent Cox regression model (non-proportional hazard model) was used to assess the effect of lifestyle groups on progression to mild cognitive impairment or dementia in the total study population and APOE ε4 stratified population. In this model death and drop-out data were treated as censoring data. Additionally, we evaluated whether dropping out from the study affected the effects of lifestyle on mild cognitive impairment or dementia using an inverse probability weighting model. We also assessed the influence of death as a competing risk for mild cognitive impairment or dementia via competing risk analyses.

Sensitivity analyses We conducted several additional sensitivity analyses to evaluate the robustness of our findings. Firstly, we included the number of lifestyle factors as a continuous variable in the mixed model to test the rationale of our grouping method. Secondly, we derived a weighted standardised healthy lifestyle score based on the β coefficient of each lifestyle factor in the linear mixed effects model adjusted for covariates.28 Specifically, the original binary lifestyle variables were multiplied by the β coefficients, and all the β coefficients were summed. The score of each lifestyle factor was obtained via the β coefficient of each lifestyle factor divided by the sum of the β coefficients, multiplied by 100. Participants were categorised into three groups based on the third of weighted standardised score into low, intermediate, and high score groups. Third, we excluded each lifestyle factor to identify possible factors that might drive the associations with memory. The excluded lifestyle factor was used as a confounder. Fourthly, we excluded participants who later developed mild cognitive impairment or dementia to determine whether the results were consistent with the main results. Fifthly, we excluded participants who died or dropped out throughout the study in order to evaluate the possible effects of death or drop-out on the results (if the results were consistent when participants were excluded with when they were not excluded, death or drop-out would have had no effect on the results). Finally, to assess the reverse causality of memory decline and healthy lifestyle profile, we performed a cross-lagged panel model, which is used when examining reciprocal causal processes in longitudinal data with at least two waves of data. Each analysis included a cross-wave component that estimated the change in each outcome from one wave, characterised by relationships between outcomes from each specific wave. We examined changes in composite z scores for memory in relation to changes in healthy lifestyle profiles over time. All statistical analyses were done in R (version 3.6.3). We used the following packages: mice, for imputing; lmerTest, for linear mixed effects models; cmprsk, for competing risk models; lavaan, for cross-lagged analysis; and ipw, for inverse probability weighting. The two tailed significance level was set at P<0.05.