There are more than 50 million people living with dementia globally, with a figure increasing to 152 million by 2050 Alzheimer’s Disease International [ADI] (2019). China has the largest population with dementia (9.5 million), followed by the United States (4.2 million) Alzheimer’s Disease International [ADI] (2015). The annual cost associated with dementia is estimated at $1.33 trillion in 2020, with a figure rising to $9.12 trillion in 2050 (Jia et al., 2020). There is no cure for dementia, but the Lancet Commission recently reported that the number of dementia could be reduced by 40% by targeting 12 modifiable risk factors (Livingston et al., 2020). The first step is screening out people at high risk of dementia, but screening for every older adult is not necessary and possible because of the heavy burden and ethical issues. Therefore, an easy-to-use prediction model for health professionals in various care settings is needed to roll out those with low-risk and select high-risk dementia for further management.

There are easy-to-use (i.e., without expensive or unavailable biomarkers) and nationally representative dementia prediction models to assess the risk of dementia in care settings quantitatively, but there might be shortcomings leading to distorted estimates of the predictive performances of the model (Li et al., 2018; Hou et al., 2019; Licher et al., 2019; Zhou et al., 2020). One important consideration is related to the effect of death. In one of the studies (Li et al., 2018) that used the Framingham Heart Study to develop a 5-, 10-, and 20-year dementia risk prediction model, the area under the curve (AUC) was 0.72 with predictors of demographic characteristics and diseases. Two studies (Zhou et al., 2020; Hu et al., 2021) used the Chinese Longitudinal Healthy Longevity Survey (CLHLS) to develop a cognitive impairment prediction model, with an average AUC of 0.80. However, these studies did not consider death, a critical competing risk of dementia. Given that dementia generally occurs in late life, failure to account for such an important competing risk could result in bias, limiting the practical use of current models (Fine and Gray, 1999).

Another consideration is related to incorporation bias. Some dementia prediction models (Licher et al., 2019; Hu et al., 2021) included predictors that are part of the assessment of dementia, such as the cognition score or physical ability, resulting in optimistic estimates of the model performances. The last consideration involved the generalization of the prediction model to different ethnic groups. Licher et al. (2019) developed a dementia model with the AUCs of 0.78 of the basic model and 0.86 of the augmenting one (such as cognition and genetic predictors) while adjusting for the effect of competing risk of dementia. However, this model was developed in Caucasian descent, and it was doubtful whether the predictors and model were applicable to Asians. To fill these gaps, we aimed to develop and validate the cognitive impairment prediction model that would take account of the competing risk of death and exclude predictors of the definition of dementia in Chinese people.

The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) was followed (Collins et al., 2015; Supplementary Table 5).

We developed a model to predict cognitive impairment with a follow-up duration of up to 6.6 years among older adults aged 65 years and above. The model included eight predictors (i.e., age, sex, education, marital status, activity duration, playing cards or mah-jongg, watching TV or listening to the radio, and stroke and cardiovascular diseases) with a good discriminative ability. This model could help practitioners and nurses in various care settings identify older adults at high risk of 6-year cognitive impairment and thus give them continuing monitoring and interventions.

Predictors in this model included age (Kivipelto et al., 2006; Anstey et al., 2013; Barnes et al., 2014; Walters et al., 2016), sex (Kivipelto et al., 2006; Anstey et al., 2013; Walters et al., 2016), education (Kivipelto et al., 2006; Anstey et al., 2013; Barnes et al., 2014), and stroke (Anstey et al., 2013; Barnes et al., 2014; Walters et al., 2016), which were also included in Licher et al. (2018) the Cardiovascular Risk Factors, Aging, and Dementia (CAIDE) from Finland (Kivipelto et al., 2006), Brief Dementia Screening Indicator (BDSI) from America (Barnes et al., 2014), Australian National University Alzheimer’s Disease Risk Index (ANU-ADRI) from Australia (Anstey et al., 2013), and Dementia Risk Score (DRS) from England (Walters et al., 2016).

The incidence rate of cognitive impairment in our study was similar to previous studies. Compared with previous studies that only selected older people living in their own homes, the selection criteria of participants did not contribute to a relatively higher percentage of cognitive impairment identified in our study. The 6-year incidence rate of cognitive impairment was ranged from 5.6 to 9.4% (compared with 6.6% in our study) among older adults aged 65–74 years in the Cardiovascular Health Study (Barnes et al., 2014). The 6-year incidence rate of the cognitive impairment was ranged from 15.2 to 26.8% (compared with 15.6% in our study) in the Sacramento Area Latino Study on Aging (Barnes et al., 2014). The reason might be that a higher proportion (98.6%) of older adults were living in their own homes in our study.

Our model showed that being single increased the risk of dementia compared with married individuals. This finding was consistent with previous studies. The previous meta-analysis (Sommerlad et al., 2018) and our experience (Hu et al., 2021) have suggested that being widowed or single could increase the risk of dementia compared with married individuals. One reason behind this relation might be due to dyadic coping (Falconier and Kuhn, 2019). Another reason might be that married people may be more likely to participate in physical and social activities conducive to cognitive reserve (Hu et al., 2021). However, this strong predictor has been overlooked in previous studies (Kivipelto et al., 2006; Anstey et al., 2013; Barnes et al., 2014; Walters et al., 2016). Further studies should examine the association between the marital status and cognitive impairment.

The activity duration was selected as an important predictor of future cognitive impairment. Physical activity could improve the hippocampal volume, induce neurogenesis, and generate more neurotrophic factors (Mariotti et al., 2014; Urbán and Guillemot, 2014). However, only a long period of activity was predictive of the impairment. There might be a dose-response relationship between physical activity and cognition. The dose could be determined by different intensities, durations, and types (Tolppanen et al., 2015; Palta et al., 2019). One longitudinal study (Palta et al., 2019) with 10,705 participants and a follow-up of 17.4 years showed that moderate and high-intensity physical activities were associated with low incidences of cognitive decline but not the low intensity. Another study (Tolppanen et al., 2015) followed more than 3,000 middle-aged and older adults up to 36 years, and results showed that, after midlife, maintaining high physical activity or increasing physical activity was associated with lower dementia risk.

Two mental stimulation activities, playing cards and mah-jongg and watching TV or listening to the radio, were predictors. These two cognitive activities were called novel and passive cognitive activities, respectively (Carlson et al., 2008). The difference between these two cognitive activities is that the former needs a response. Several studies have investigated the association between these two types of cognitive activity and cognitive impairment. One study (Carlson et al., 2008) in male twins controlling for genetic and environmental factors showed that the midlife cognitive activity was related to a 26% risk reduction for dementia, particularly among twin pairs at an elevated genetic risk. Another recent study (Najar et al., 2019) that followed 800 women with a median age of 44 years exhibited a similar result. This phenomenon could be explained by the cognitive reserve hypothesis of the brain and “use it or lose it” theory. Brain reserve indicates that individuals with a greater brain reserve require higher levels of change in cortical thickness, β-amyloid, and regional atrophy to exhibit clinical symptoms of cognitive impairment (Tucker and Stern, 2011; Najar et al., 2019). The “use it or lose it” theory means that disuse could lead to subsequent atrophy of cognitive skills (Stern and Munn, 2010). Future detailed information on the type, duration, intensity, and timing of that activity might improve the accuracy of the model (Stern and Munn, 2010; Sajeev et al., 2016).

Consistent with previous studies, our model performed better in those younger older adults aged 65–74 years and those with higher MMSE scores of 26–30. One of the studies (Licher et al., 2018) included 2,710 older adults with a mean age of 71.2 years, where the general discriminative ability of the dementia prediction model was 0.79, but the C-statistic decreased to 0.57 among individuals older than 80 years. The study by Hall et al. (2019) constructed the dementia prediction model among older adults aged 85 years and above, and results showed that the predictors were different from those in previous reports of the younger population. Predictors of age, sex, and vascular and lifestyle factors no longer had the predictive ability. Objective predictors, such as total brain volume and hippocampal volume, helped predict cognitive impairment in the oldest-old (Licher et al., 2018). More promising blood-based predictors such as the plasma total tau level and lipids with high accuracy of 90% have been explored (Mapstone et al., 2014; Pase et al., 2019). Their usability in various care settings deserves attention.

In 2020, the US Preventive Services Task Force (USPSTF) issued that the balance of benefits and harms of screening for cognitive impairment was not determined due to insufficient evidence, making the role of our prediction model controversial. On one hand, early identification allowed early interventions. The USPSTF recommended that early signs or symptoms of cognitive impairment should be noticed and evaluated as appropriate because it allowed for identification and treatment of reversible causes and more involvement from patients and families. One study (Ngandu et al., 2015) involving 2,654 individuals investigated the effect of a 2-year multidomain intervention on cognitive function in at-risk older adults, and results showed that cognitive function was improved or maintained. In contrast, even though the harm of screening is limited, potential harms should be addressed, such as ineffectiveness and adverse effects of pharmacological interventions, potential stress, and lower quality of life caused by the awareness of a diagnosis of cognitive impairment (Stites et al., 2017). However, the main aim of our study was related to early prevention and did not emphasize diagnosis, and thus harms due to diagnosis were not applicable. Theoretically, 47.7% of dementias could be prevented or delayed through early detection and interventions on modifiable risk factors. This number was similar to that published by the Lancet Commission that emphasized the importance of early intervention. Due to the relatively low incidence of cognitive impairment, the positive predictive value was low, whereas the negative predictive value was high. Nevertheless, since this model aims to screen out individuals at high risk of cognitive impairment for further monitoring and non-pharmacological interventions, this model could meet this demand.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Duke University and Peking University. Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin.

MH and YG conceived and designed the study. MH contributed to the statistical analysis and wrote the first draft of the manuscript. ZS contributed to the manuscript revision. TK, LX, and HF contributed to the writing of the final version of the manuscript. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.