This systematic review followed the procedures of the Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) statement (Page et al., 2021). The present study was registered in PROSPERO. Available at: https://www.crd.york.ac.uk/prospero/display_record.php? ID=CRD42023316549. Since Harrison reported related research in 2014 (Harrison et al., 2014), we have systematically reviewed articles from 1 January 2014 until 16 February 2023. PubMed, Web Of Science, PsychINFO, Embase, Cochrane Library, CNKI, Sinomed, and WanFang were searched. Combinations of the following terms were explored: “cognit*,” “dementia,” “Alzheimer’s disease,” “cognitive dysfunction,” “cognitive impair*,” “cognitive decline,” “heart disease risk factors,” “cardiovascular risk,” “cardiovascular risk model,” “cardiovascular risk score,” “cardiovascular health,” “cardiovascular health metrics,” and “Framingham”. The search strategy of PubMed is presented in Table 1.

The inclusion criteria for studies were the following: (1) assessed CVD risk models; (2) had access to specific information on cognitive test results (or how they changed); and (3) used a cross-sectional, longitudinal cohort study design. Articles were disqualified if they only examined a single risk factor or included dementia or cognitive impairment data in their risk models. Studies written in languages other than English or Chinese, those with incomplete abstracts, and those lacking full texts were also disqualified.

From each eligible study, we extracted data on (1) article information: authors, publication year, and country; (2) cohort information: name of the study, number of participants, percentage of women, mean age, and race/ethnicity; (3) study design: cross-sectional study, longitudinal cohort study (follow-up time); (4) CVD risk models: details of the risk model used and any modifications made; (5) primary outcome (dementia diagnoses, cognitive decline, or cognitive function); and (6) analytical strategy: details on models’ prognostic performance including odds ratio (OR), 95% confidence interval (95% CI), hazard ratio (HR), β coefficients, the area under the receiver operating characteristic curve (AUC), or C-statistic.

Cohort studies were evaluated for quality using the Newcastle-Ottawa scale (NOS) (Lo et al., 2014). In more detail, the scores of the related items in three domains–four items for the chosen domain, one item for the comparability domain, and three items for the resulting part–were added up to create a total score of the risk of bias. High-quality studies were those with a total score of six or higher. The Agency for Healthcare Research and Quality (AHRQ) provided an 11-item checklist for evaluating the methodological quality of the cross-sectional studies (Newman, 2000). If a question was answered “NO” or “UNCLEAR,” the item received a score of “0” and if the reply was “YES,” the item received a score of “1”. Low-quality articles were given a score of 0–3, moderate quality ones of 4–7, and high-quality ones of 8–11. The prediction model Risk Of Bias ASsessment Tool (PROBAST) includes 20 signaling questions across four domains: participants, predictors, outcome, and analysis (Moons et al., 2019). It is used to assess a prediction model’s risk of bias and applicability concerns. It can be downloaded from www.probast.org.

Due to the variation among studies, including variations in the variables included in the CVD risk models, the small number of studies using the same models, and even variations in adjusted variables across studies using the same models, as well as differences in cognitive outcomes and variations in the reporting of effect sizes, synthesis with meta-analysis is not possible.

The search yielded a total of 9,718 articles. Of these, 3,250 were duplicates, and 6,417 were irrelevant after title and abstract reviews. A total of 51 articles underwent full-text screening and 34 were excluded due to unmet criteria. Other 5 extra studies were obtained from the references of the included articles and the citation retrieval of related systematic reviews, and 22 full-text articles, including 14 cohort studies and 8 cross-sectional studies, were included. Figure 1 summarizes the study selection process.

The included research was conducted in 10 different countries. The most significant articles came from the United States, China, Italy, France, Thailand, the Republic of Korea, and the United Kingdom. A total of 19 studies were derived from datasets based on data from a population-based cohort. Most studies used the National Alzheimer Coordinating Center (NACC) database (Jefferson et al., 2015; Viswanathan et al., 2015; Pelcher et al., 2020). Sample sizes ranged from 300 to 429,033 and were racially diverse as they included White people, Black people, Hispanic/Latino (H/L) people, and non-H/L people. The mean age at baseline ranged from 45 to 79.5 years. In cohort studies, the follow-up duration ranged from 1 to 26.2 years. Detailed information on the study cohort, sample traits, CVD risk models, cognitive function or dementia assessments, and model details are presented in Table 2.

After the studies’ quality assessment, the 14 cohort studies were rated as good quality, and the 8 cross-sectional studies were rated moderate to high. The deduction items are concentrated in 6 and 9–11. The key concern is that several cardiovascular parameters (physical activity and diet) were dependent on self-reports. In addition, some studies substituted similar variables because existing cohorts did not contain the desired variables. For example, Washington Heights-Inwood Community Aging Project Risk Score (WHICAP) studies used Body Mass Index (BMI) instead of the waist-to-hip ratio (Torres et al., 2020). The quality assessment of the reviewed studies is shown in Tables 3, 4.

A total of 22 studies described 15 CVD prediction models (Kannel et al., 1976; D’Agostino et al., 1994, 2008; Conroy et al., 2003; Sacco et al., 2009; Lloyd-Jones et al., 2010; Reitz et al., 2010; Goff et al., 2014; Jee et al., 2014; Maclagan et al., 2014; Tsao and Vasan, 2015; Yang et al., 2016; Dufouil et al., 2017; Hageman et al., 2021; Buawangpong et al., 2022). Among them were two models without evaluation (Lloyd-Jones et al., 2010; Maclagan et al., 2014). Life’s Simple 7 (LS7) was defined as the concept and determining of the metrics done by the Goals and Metrics Committee of the Strategic Planning Task Force of the American Heart Association (AHA) (Lloyd-Jones et al., 2010). Furthermore, the Cardiovascular Health in Ambulatory Care Research Team (CANHEART) health index was a modified version of the AHA’s definition of ideal cardiovascular health (CVH) for Canadians (Maclagan et al., 2014). Besides, WHICAP claimed good prediction performance, with no specific values discovered (Reitz et al., 2010). Others found AUC or C-statistics ranging from 0.51 to 0.86. Except for the Thai Cardiovascular Risk score (TCVR) (Buawangpong et al., 2022), most of the models demonstrated good to acceptable discrimination (Receiver Operator Characteristic Curve>0.7). No studies exhibited a low risk of bias owing to insufficient analytical domain reporting (Tables 5, 6). The explanation could be that the CVD risk models were built and verified before the PROBAST report was proposed, and there needed to be a uniform and standard reporting checklist, resulting in a low assessment score. In contrast, its predictive performance was quite good, particularly for the Framingham series models.

Framingham risk models and CAIDE, and other nine CVD risk models were used to predict dementia or cognitive decline, namely, the Korean Coronary Heart Disease Risk score (KPS), Systematic COronary Risk Evaluation (SCORE), SCORE2, atherosclerotic cardiovascular disease pooled cohort equation (ASCVD-PCE), TCVR, CANHEART, Global Vascular Risk score (GVRS), China-PAR, and LS7. Generally, the included CVD risk models had 5 to 15 factors. GVRS had the most number of factors, whereas SCORE had the least (Conroy et al., 2003; Sacco et al., 2009). Except for CANHEART and LS7, other models all included sex, age (age-stratified), systolic blood pressure (SBP), and smoking habits. Notably, race was taken into consideration by ASCVD-PCE, GVRS, and WHICAP (Rundek et al., 2020; Tarraf et al., 2020; Torres et al., 2020; Schaich et al., 2022). WHICAP is a simple summary risk score for predicting Alzheimer’s disease (AD) in older adults (Reitz et al., 2010). Given that its construction was based on cardiovascular risk factors, we decided that WHICAP could be included. All of the risk factors can be divided into traditional factors, behavioral factors, and anthropometric factors, including demographic data (age and sex), laboratory data (BP, blood glucose, and APOE ℇ4), questionnaire data (physical activity and dietary habits) and self-report data (physical activity, dietary habits, and smoking habits). Please see Tables 7, 8 for more details.

Compared with previous research, more studies focused on cognitive function than dementia (Harrison et al., 2014). Dementia was further diagnosed in different subcategories, including AD and vascular disease (VD) (Zheng et al., 2022). There were two ways for included studies to assess outcomes: one was cognitive diagnoses, including International Classification of Diseases 9 or 10 (ICD-9, ICD-10), an expert committee, clinical examination, or Clinical Dementia Rating (CDR), and the other way was neuropsychological assessments including Mini-Mental State Examination (MMSE), Montreal Cognitive Assessment (MOCA), or domain-specific tests. Furthermore, global cognitive function could be acquired by calculating Z scores.

Except for one study (Jeon et al., 2021), the various risk models were found to be significantly connected with a higher chance of dementia in the future. In this study, CVH levels did not correlate with early cognitive decline in the Korean study population. The validity of the CANHEART model used to measure CVH was not established for the Korean population because it was initially developed by Canadian academics for the Canadian population. In addition, the CANHEART model used in Canadian samples for predicting cognitive decline was unclear.

The results were ambiguous. AUC was the prognostic performance metric in only four studies (Viticchi et al., 2017; Buawangpong et al., 2022; Schaich et al., 2022; Zheng et al., 2022). More studies just analyzed β efficiency. According to the individual risk factor profiles in midlife, the CAIDE risk score was explicitly created as an easy method to forecast the likelihood of dementia in the future (Kivipelto et al., 2006). With an initial C-index of 0.78 and a repeated C-index of 0.75 following external validation, the score accurately predicted dementia (Exalto et al., 2014).

A total of 12 studies used the Framingham risk model: five studies used the Framingham General Cardiovascular Risk Score (FGCRS) (Viticchi et al., 2017; Song et al., 2020; Tarraf et al., 2020; Torres et al., 2020; Ji et al., 2022), three studies used the Framingham Stroke Risk Profile (FSRP) (Jefferson et al., 2015; Harrison et al., 2017; Schaich et al., 2022), while three studies used the revised Framingham Stroke Risk Profile (rFSRP) (Viswanathan et al., 2015; Pelcher et al., 2020; McGrath et al., 2022), which eliminates left ventricular hypertrophy (LVEF), and one study used Framingham Vascular Age (Badran et al., 2019). All studies found higher Framingham risk scores associated with an increased risk of subsequent cognitive decline or impairment. Which cognitive abilities were linked to a higher risk of CVD or stroke varied among research due to various measures of cognitive function, sample sizes, or statistical performance.

Except for Framingham models, we concluded that nine specific CVD risk models may be associated with dementia and cognitive decline. Five CVD risk models were designed for a single country (Bao-Shan et al., 2021; Jeon et al., 2021; Buawangpong et al., 2022; Zheng et al., 2022; Mun et al., 2023). KPS, TCVR, the China-PAR, SCORE2, and the CANHEART health index were developed to measure cardiovascular risk among Korean, Thai, Chinese, European, and Canadian populations.

Four risk models were for the multiethnic cohort. A longitudinal cohort of non-Hispanic White people and African Americans served as the basis for the ASCVD-PCE (Schaich et al., 2022) and has since been validated for predicting atherosclerotic CVD risk in Asian and multiethnic populations, including Hispanics. A diverse group with almost equal representations of White people (34.2%), Black people (30.6%), and H/L (33.3%) participants created the WHICAP summary risk score to predict AD in older adults based on their vascular profiles associated with late-onset AD (Torres et al., 2020). GVRS is a model for assessing global vascular risk developed in the Northern Manhattan Study (NOMAS) that incorporates traditional, behavioral, and anthropometric risk factors and uses continuous variables (Rundek et al., 2020; Tarraf et al., 2020). A simple tool, LS7 only has seven modified factors, including four behavioral and three biological factors, similar to CANHEART (Lloyd-Jones et al., 2010). The AHA first suggested this metric for keeping CVH and has since been advised for brain health (Gorelick et al., 2017).

Other than CANHEART, specific CVD risk models were associated with a lower risk for dementia or cognitive decline (Covello et al., 2021; Jeon et al., 2021; Speh et al., 2021). CVD risk models from multiethnic cohorts were proven to predict dementia or cognitive decline in diverse race populations, such as White, H/L, Black, or Asian. Specifically, an article reported the association between QRISK and cognitive function, but we did not report it due to its poor quality (Hsu et al., 2015).

Only SCORE2, FSRP, ASCVD-PCE, and FCRP reported the models’ performance for dementia (Viticchi et al., 2017; Schaich et al., 2022; Zheng et al., 2022). The AUC or C-statistics of these models ranged from 0.65 to 0.75. Additionally, the C-indices of SCORE2 risk could distinguish between incident all-cause dementia, AD, VD, and all-cause death (Zheng et al., 2022).

Seven studies compared the predictive performance of two or three risk models (Harrison et al., 2017; Rundek et al., 2020; Tarraf et al., 2020; McGrath et al., 2022; Schaich et al., 2022; Zheng et al., 2022). CAIDE, as a reference substance, was used for five articles. Some studies reported that CAIDE is better for predicting dementia. For instance, one study that evaluated FSRP, ASCVD-PCE, and CAIDE concluded that CAIDE had more robust relationships with cognitive performance than FSRP and ASCVD-PCE but that ASCVD-PCE’s associations with the Digit Symbol Coding (DSC) and Digit Span (DS) were comparable to CAIDE’s (β = −0.57 and − 0.21, respectively) (Schaich et al., 2022).

Other studies showed that CAIDE had an inferior association with dementia or cognitive function compared with other risk models, such as SCOPE2, WHICAP, or GVRS. One study that compared SCORE2 with SCORE and CAIDE showed that the C-index of SCORE2 risk for discriminating against all-cause dementia was 0.750, which was significantly higher than the C-indices of SCORE and CAIDE risk (SCORE2 vs. SCORE: 0.014; SCORE2 vs. CAIDE: 0.093). The study further found that SCORE2 risk significantly improved the ability to distinguish between AD and VD (Zheng et al., 2022). Two studies showed that higher GVRS presented stronger associations with lower cognitive function than the FCRS or CAIDE (Rundek et al., 2020; Tarraf et al., 2020).