Associations of dementia polyexposure scores to Alzheimer's disease endophenotypes in a diverse population

Abstract

INTRODUCTION

Dementia clinical risk scores (CRSs) provide accessible tools for identifying individuals at risk for Alzheimer's disease (AD) and related dementias, yet their performance across diverse populations and relationships to AD endophenotypes remains unclear.

METHODS

We evaluated four CRSs, modified Cardiovascular Risk Factors, Aging, and Incidence of Dementia (mCAIDE), Washington Heights–Inwood Columbia Aging Project (WHICAP), Lifestyle for Brain Health (LIBRA), and Cognitive Dementia Risk (CogDRisk), in relation to cognitive impairment (CI) and AD endophenotypes, including tau phosphorylated at threonine 217 (pTau₂₁₇)/amyloid beta 42 (Aβ₄₂) positivity defined using a Youden index–derived cutoff for amyloid positron emission tomography (PET) positivity. Logistic and linear regression models stratified by self‐reported race/ethnicity were used to assess the associations of CRS with endophenotypes and CI and to evaluate predictive performance.

RESULTS

CogDRisk showed the strongest and most consistent performance across endophenotypes, pTau₂₁₇/Aβ₄₂ positivity, and CI, with mCAIDE performing the worst and lacking associations with plasma biomarkers. Higher CRS were consistently associated with increased odds of dementia across all races/ethnicities.

CONCLUSIONS

CRSs capture AD‐related risk across diverse populations and modestly reflect underlying biological endophenotypes, supporting their utility in community‐based risk assessment.

Highlights

• CRSs are strongly predictive and associated with cognitive outcomes.

• CRSs with broader incorporation of risk factors are more predictive of endophenotypes.

• CogDRisk shows the highest model performance across racial/ethnic groups.

1. BACKGROUND

Alzheimer's disease (AD) presents substantial clinical and economic challenges, yet early identification of individuals at risk remains limited in diverse, community‐based settings. Current AD risk and diagnostic assessment involve genetic and cognitive screening, cerebrospinal fluid (CSF) and plasma biomarkers of AD pathology, and neuroimaging; however, these approaches, though valuable, remain costly, invasive, and limited to those with access to specialized care in developed countries. ¹ , ² , ³ They also only capture risks once pathology has begun, even if clinical symptoms of AD have not yet manifested. ⁴ , ⁵ , ⁶ In addition, many older adults, especially those from underserved populations, are evaluated in primary care contexts where advanced diagnostic resources are unavailable. ⁷ , ⁸ , ⁹ Given these limitations and the growing prevalence of AD, there is a pressing need for effective risk assessment tools that can identify at‐risk individuals earlier and support timely preventive interventions.

Clinical risk scores (CRSs) have been developed as accessible tools that estimate AD and dementia risks based on demographics, lifestyle, and clinical profiles. ¹⁰ Many CRSs can be readily implemented in community and primary care settings, often using self‐reported information, and can help identify individuals who may benefit from further confirmatory testing with biomarkers or genetic screening. ¹¹ , ¹² CRSs are designed to capture mid‐ to late‐life risk factors, enabling individualized prevention strategies and risk stratification. These tools are particularly relevant for individuals with a family history of dementia or known genetic susceptibility, as they provide actionable targets for risk reduction and support precision health approaches. Many CRSs have been proposed ¹³ and validated in mid‐ and late‐life populations across diverse settings, including the modified Cardiovascular Risk Factors, Aging, and Incidence of Dementia (mCAIDE) score, ¹⁴ the Washington Heights–Inwood Columbia Aging Project (WHICAP) score, ¹⁵ the Lifestyle for Brain Health (LIBRA) index, ¹⁶ and the Cognitive Dementia Risk (CogDRisk) score. ¹⁷ Among them, mCAIDE and WHICAP were independently developed using US‐based cohorts, while LIBRA and CogDRisk were developed based on systematic reviews and meta‐analyses. Each integrates slightly different risk domains such as vascular health, lifestyle, psychosocial, and cognitive engagement factors, reflecting distinct methodological approaches and population contexts in which they were developed. ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰

External validation studies of CRSs have reported variable performance across cohorts and settings. LIBRA and CogDRisk have been validated in multiple independent studies ²¹ , ²² , ²³ , ²⁴ , ²⁵ , ²⁶ ; however, most validation studies were conducted in populations of predominantly European descent, limiting their generalizability to other racial and ethnic groups. The WHICAP score, which uniquely incorporates ethnicity‐specific weighting, has only been evaluated in a small diverse population. ²⁷ The original CAIDE index has been validated in large, multiethnic samples, ²⁸ but few studies have independently validated mCAIDE. To assess utility in diverse populations, CRSs need to be validated across different racial and ethnic groups. Furthermore, only a handful of studies have examined the relationships between these risk indices and AD endophenotypes, ²⁹ , ³⁰ , ³¹ an important step in determining whether CRSs reflect underlying disease pathology. Because abnormalities in AD biomarkers precede clinical syndromes of dementia, ³² , ³³ , ³⁴ the growing emphasis on AD pathophysiology highlights the need to validate CRSs in detecting early abnormalities in biological endophenotypes.

This study examines how four widely used CRSs relate to dementia diagnosis and to AD endophenotypes within a multiethnic cohort, the Health and Aging Brain Study–Health Disparities (HABS‐HD). By investigating the utility of CRSs across racial/ethnic groups and their sensitivity to early changes in AD endophenotypes, we aimed to clarify the extent to which CRSs could serve as equitable and biologically informative tools for AD and dementia risk assessment in real‐world, community‐based settings.

2. METHODS

2.1. Health and Aging Brain Study–Health Disparities

2.1.1. Study population

The study used cross‐sectional data from the HABS‐HD cohort. HABS‐HD is an ongoing community‐based AD study that comprises participants from self‐reported Black (African American [AA]), Hispanic/Latinx (LA) (e.g., Mexican American, Puerto Rican, Dominican Republic), and non‐Hispanic White (NHW) populations recruited at the University of North Texas Health Science Center, Fort Worth, Texas, USA. The study collects data on demographics, AD biomarkers, neuroimaging, clinical history, and genomics, providing a comprehensive resource for AD and related dementia research. The study recruited generally healthy individuals free of serious mental illnesses and medical conditions, and we further excluded participants younger than 55 years old and those without apolipoprotein E (APOE) genotype information. The visit 1 baseline measurements for all outcomes of interest were used for all downstream analyses. All HABS‐HD participants (and/or their legal guardians) signed written informed consent to participate in the study.

RESEARCH IN CONTEXT

Systematic review: A number of CRSs have been developed for AD and dementia that incorporate different risk and protective factors; however, their predictive performance in diverse populations and relationships with AD endophenotypes remain underexplored.

Interpretation: CRSs that broadly incorporate mid‐ to late‐life risk and protective factors such as cardiometabolic, lifestyle, and psychosocial variables enhanced the model performance for cognitive and endophenotypic outcomes across racial/ethnic groups.

Future directions: Future research should prioritize external replication of our findings in larger, socioeconomically diverse populations to further measure generalizability of these CRSs, along with other CRSs that were not included in this study. Furthermore, more granular stratification (e.g., sex, race, socioeconomic status) and interaction analysis to better characterize CRS–endophenotype associations. Additionally, longitudinal studies looking at temporal relationships between CRSs, biomarkers, and cognitive decline would be crucial for clinical implementation.

2.1.2. Endophenotype standardization

Plasma biomarkers: Fasting blood samples were collected, and the levels of plasma biomarkers (i.e., amyloid beta 42 [Aβ₄₂], Aβ₄₀, tau phosphorylated at threonine 217 [pTau₂₁₇], total tau, neurofilament light chain [NfL]) were quantified using commercially available kits, Quanterix, for all the participants of the HABS‐HD. ³⁵ Plasma biomarker levels were log‐transformed to reduce the skewness, followed by z‐score normalization (mean = 0, standard deviation [SD] = 1) for all participants to standardize and remove any batch effects. The Aβ₄₂/Aβ₄₀ ratio was calculated by dividing the raw values of Aβ₄₂ and Aβ₄₀ prior to transformation and normalization. ³⁶ We excluded participants with plasma biomarkers identified as outliers using the interquartile range methodology. ³⁷

Brain morphometry: Cortical thickness was defined as the surface area‐weighted average of cortical thickness in the right and left entorhinal cortex, fusiform, and inferior and middle temporal cortex. These regions of interest correspond to brain areas that are especially vulnerable to AD‐related neurodegeneration. Hippocampal volume was determined by taking the mean volume of the right and left hippocampal volumes determined on a T1‐weighted volume scan. ³⁸

Cognition: The memory domain was assessed using immediate and delayed recall from the Wechsler Memory Scale‐III (WMS‐III) Logical Memory ³⁹ and the Spanish‐English Verbal Learning Test (SEVLT). ⁴⁰ The verbal ability domain was assessed using Letter Fluency (FAS) and Animal Naming tests. ⁴¹ The executive function domain was assessed using the WMS‐III Digit Span ³⁹ and the Trail Making Test, Parts A and B. ⁴¹ All cognitive test scores were standardized using z‐score transformation, and these scores for each domain (memory, verbal ability, and executive function) were averaged to create composite scores. In addition, the Mini‐Mental State Examination (MMSE) ⁴² and Clinical Dementia Rating (CDR) scale ⁴³ were administered to all participants as part of the neuropsychological assessment.

2.2. Modifiable risk factors

It is important to note that covariate definitions vary across CRSs. For example, dyslipidemia is classified through total cholesterol in CogDRisk and LIBRA, high‐density lipoprotein (HDL) in WHICAP, and self‐reported high cholesterol in mCAIDE. We retained the covariates as specified by each CRS rather than harmonizing them across models to retain their original scoring scheme.

Education: Years of education were coded as a numeric variable, with values ranging from 0 (kindergarten) to 20 (postgraduate).

Body mass index (BMI): BMI at baseline was calculated as weight divided by height squared, kg/m². Obesity was classified based on BMI as underweight (<18.5 kg/m²), normal weight (18.5 to 24.9 kg/m²), overweight (25 to 29.9 kg/m²), or obese (≥30 kg/m²).

Hypertension: All participants underwent a series of anthropometric measurements to assess blood pressure. Hypertension was defined as a past medical history of the condition or consistently elevated blood pressure, with at least two measurements of systolic blood pressure (SBP) ≥140 mmHg or diastolic blood pressure ≥90 mmHg at the baseline visit.

Dyslipidemia/high cholesterol: Hyperlipidemia was defined as meeting any of the following criteria: low‐density lipoprotein cholesterol ≥120 mg/dL, total cholesterol ≥240 mg/dL, triglycerides ≥200 mg/dL, or a documented history of high cholesterol.

Diabetes: All participants had measures of hemoglobin A1C from fasting blood samples. Diabetes was defined as an A1C level ≥6.5% at the baseline visit or a documented history of the condition.

Depression: Participants were considered to have depression if they had a past medical history of the condition or scored ≥10 on the 30‐item Geriatric Depression Scale (GDS).

Stroke: Participants self‐reported (i.e., yes or no) a history of stroke or indicated if they had ever been told by a healthcare professional that they had experienced a stroke.

Traumatic brain injury (TBI): TBI with or without loss of consciousness information was collected through self‐report data.

Renal dysfunction: Estimated glomerular filtration rate (eGFR) was calculated using the 2021 Chronic Kidney Disease Epidemiology (CKD‐EPI) creatinine equation, which estimates kidney function based on serum creatinine, age, and sex. eGFR values are reported in mL/min/1.73 m², and participants with eGFR < 60 mL/min/1.73 m² were classified as having renal dysfunction, consistent with standard clinical definitions of CKD.

Alcohol intake: Alcohol intake was assessed using a single item from the Alcohol Use Disorders Identification Test (AUDIT): “How often do you have a drink containing alcohol?” Response options included never, monthly or less, two to four times per month, two to three times per week, and four or more times per week.

Smoking: Smoking was classified as self‐reported current smokers, indicating yes or no.

Social support: Loneliness was approximated using the social support questionnaires in the HABS‐HD. The questionnaire consisted of 12 items assessing perceived availability of emotional, neighborhood‐level, and instrumental support.

Physical activity: Physical activity was assessed using the Rapid Assessment of Physical Activity (RAPA), a questionnaire designed to provide a simple and accessible measure of physical activity among adults aged 50 years and older. ⁴⁴ RAPA scores were derived according to established guidelines, with higher scores indicating greater levels of physical activity.

2.3. Clinical risk scores

For a detailed description of variables for each CRS, refer to Table S1.

2.3.1. Modified Cardiovascular Risk Factors, Aging, and Dementia

mCAIDE was derived from the CAIDE score developed in Finland, but modified to better reflect US cohorts and mid‐ to late‐life risk using two US‐based cohorts. ¹⁴ mCAIDE incorporates age, sex, education, BMI, SBP, self‐reported high cholesterol status, and physical activity measured by the mini Physical Performance Testing (miniPPT). The RAPA was used in place of mPPT to replace the physical activity measure. Participants whose total scores were in the lowest tertile were classified as “inactive,” and those above this threshold as “active.”

2.3.2. Washington Heights‐Inwood Columbia Aging Project

WHICAP is a CRS that was developed in an independent cohort of the Washington Heights and Inwood Community Aging project in Manhattan, New York City, to create a risk index for late‐onset AD. ¹⁵ It includes the variables sex, age, education, race/ethnicity, APOE ɛ4 carrier status, BMI, low HDL cholesterol (HDL‐C), diabetes, hypertension, and current smoker. Among the four CRSs considered in our analysis, WHICAP was the only one that explicitly incorporates race/ethnicity as a predictor. Therefore, we selected WHICAP to assess its transportability within HABS‐HD, a multiethnic cohort from a different geographic region of the United States. HDL‐C was categorized as low or normal based on sex‐specific cutoffs: males with HDL‐C < 50 mg/dL and females with HDL‐C < 40 mg/dL were classified as low, while values above these thresholds were classified as normal. APOE ɛ4 carrier status was excluded from the scoring to prevent inflation of WHICAP relative to other CRSs that do not incorporate APOE ɛ4 information. Additionally, since genetic information is not always readily available in primary clinical settings, CRSs that do not incorporate genetic factors may be more practical and more readily adopted.

2.3.3. Lifestyle for Brain Health

The LIBRA score was developed using an approach that combined a systematic review, meta‐analysis, and Delphi consensus studies to identify and weight modifiable lifestyle, cardiovascular, and metabolic risk factors for all‐cause dementia. ¹⁶ , ¹⁹ , ⁴⁵ LIBRA was specifically designed to support primary prevention by focusing purely on modifiable risk factors in mid to late life and does not include age, sex, or education. The LIBRA score includes a healthy diet (i.e., Mediterranean diet), physical inactivity, cognitive activity, low to moderate alcohol consumption, current smoker, heart diseases, diabetes, high cholesterol, BMI, hypertension, renal dysfunction, and depression. Participants who reported drinking two to three times per week on the AUDIT questionnaire were classified as low to moderate consumers, corresponding to >0 and <14 units of alcohol per week. For the physical activity measure, participants who answered “yes” to item 1 or item 2 on the RAPA questionnaire were classified as physically inactive. Healthy diet and cognitive activity measures were omitted from the scoring due to the lack of data in the HABS‐HD study.

2.3.4. Cognitive Health and Dementia Risk Reduction

CogDRisk was developed through a systematic review and meta‐analysis of studies identified from international databases to quantify the impact of modifiable factors on cognitive decline and dementia. ¹⁷ , ²⁰ The CogDRisk score includes sex‐specific age categories (i.e., age effect estimates differ by sex), education, midlife (≤65 years) obesity, dyslipidemia, diabetes stratified by sex, history of stroke, history of TBI, hypertension, atrial fibrillation, clinical diagnosis of insomnia, depression, physical inactivity, cognitive engagement, social engagement, diet, and current smoker. Participants who scored at least 1 SD below the mean on the social support questionnaire score were classified as “lonely,” and those above this threshold were classified as “not lonely” for the CogDrisk scoring. For physical activity measures, participants who answered “yes” to item 6 or 7 on the RAPA questionnaire were classified as physically active, reflecting an activity level of >150 min per week of moderate to vigorous activity. Baseline BMI at visit 1 was used in place of midlife obesity/BMI, which was not available in this cross‐sectional study. Insomnia, atrial fibrillation, cognitive engagement, and diet measures were excluded from the scoring due to unavailability in the HABS‐HD study.

2.4. Missingness

Missing data for variables used in the construction of the CRSs were imputed using the missForest algorithm ⁴⁶ implemented in R for both continuous and categorical variables. The imputation dataset included relevant demographics, clinical variables, biomarkers, and predictors. Variables with high substantial missingness (>40%), such as myocardial and heart disease, were excluded from imputation. To verify that imputation did not introduce bias, all analyses were compared against analyses using complete case data.

2.5. Mild cognitive impairment (MCI) and dementia diagnosis

Cognitive status was assessed using self‐report and informant reports, the CDR scale, and a neuropsychological battery covering memory, executive function, and verbal ability (animal naming, FAS, SEVLT, WMS‐III Logical Memory, Digit Span, Trail Making, and Digit Symbol Substitution) as previously described. ³⁵ , ⁴⁷ Participants without any complaints of cognitive change, reported either by the individual or an informant, with a CDR sum of boxes (CDR‐SB) of 0, were classified as cognitively unimpaired. If a participant has an isolated poor neuropsychological test performance, in the absence of any cognitive complaints and functional decline, they were assigned as cognitively unimpaired. MCI was the classification of participants with complaints of cognitive change, reported either by the individual or an informant, with a CDR‐SB score of 0.5 to 2.0 and performance at or below 1.5 SD below z‐score adjusted norms on at least one cognitive test. Dementia diagnosis was made based on CDR‐SB score ≥2.5 and cognitive test score at or below 2 SD below the mean on tests in two or more domains.

2.6. Statistical analysis

2.6.1. Cognitive impairment

Logistic regression models were used to assess the association of CRS with (i) MCI, (ii) dementia, and (iii) MCI + dementia using cognitively normal participants as the reference group. Age, sex, APOE genotypes, and years of education were adjusted for in LIBRA, and APOE genotype was included as a covariate for mCAIDE, WHICAP, and CogDRisk. APOE genotype was modeled as a categorical variable at the individual genotype level (ε2/ε2, ε2/ε3, ε2/ε4, ε3/ε4, ε4/ε4), with ε3/ε3 specified as the reference category. Thus, regression estimates reflect genotype‐specific effects relative to the ε3/ε3 group. Model performance was evaluated by calculating the AUC values to assess model discrimination, and the Nagelkerke R‐squared (R ²) was calculated to determine the goodness of fit of the models. The Delong test was used to assess whether there were statistical differences in the AUC of two models. ⁴⁸ Self‐reported race/ethnicity was used for stratified analysis. P values were adjusted for multiple hypothesis testing by applying the false discovery rate (FDR) at 5%.

2.6.2. AD Endophenotypes

Linear regression models were used to estimate the effects of CRS on AD endophenotypes, including (i) plasma biomarkers (Aβ₄₂/Aβ₄₀ ratio, pTau₂₁₇, total tau, and NfL), (ii) neuroimaging measures (cortical thickness and hippocampal volume), and (iii) clinical function (cognitive severity assessed by MMSE and clinical severity assessed by CDR, and memory, verbal ability, and executive function cognitive domains assessed by composite scores). For associations with plasma biomarkers, models for mCAIDE and WHICAP were adjusted for APOE genotype and eGFR, models for LIBRA were adjusted for age, sex, BMI, and APOE genotype, and models for CogDRisk were adjusted for APOE genotype, eGFR, and BMI. For associations with neuroimaging, models for mCAIDE, WHICAP, and CogDRisk were adjusted for APOE genotype and intracranial volume. In contrast, models for LIBRA were adjusted for age, sex, APOE genotype, and intracranial volume. For associations with cognitive function, models for mCAIDE, WHICAP, and CogDRisk were adjusted for interview language and APOE genotype, while models for LIBRA were adjusted for age, sex, years of education, APOE genotype, and interview language. Pvalues were adjusted for multiple hypothesis testing by applying the FDR at 5%.

For both logistic and linear regression analyses, we specified two baseline models: (i) a demographics model including age, sex, and years of education, and (ii) a demographics plus APOE genotype model. These baseline models were used to evaluate how much each CRS improved prediction beyond demographics alone and beyond demographics combined with APOE genotype.

2.6.3. Prediction of pTau₂₁₇/Aβ₄₂ amyloid positivity by CRS

To assess whether CRSs predicted plasma biomarker‐derived amyloid positivity, we examined the pTau₂₁₇/Aβ₄₂ ratio, an FDA‐cleared plasma measure ⁴⁹ that demonstrates high concordance with amyloid PET positivity. Plasma pTau₂₁₇/Aβ₄₂ levels were dichotomized using an empirically derived threshold optimized to predict amyloid PET positivity in the HABS‐HD cohort. The optimal cutpoint was determined using the Youden index, with stability assessed via bootstrapping (2000 iterations). Individuals were classified as positive if their pTau₂₁₇/Aβ₄₂ value was greater than or equal to the derived threshold.

We evaluated all CRSs, including sensitivity models (Section 2.4.5), as well as two demographic models, to assess predictive performance using OR, AUC, and Nagelkerke R ². Associations between CRSs and dichotomized pTau₂₁₇/Aβ₄₂ were estimated using logistic regression. For mCAIDE, WHICAP, and CogDRisk, models were adjusted for APOE genotype. For LIBRA, models additionally included age, sex, years of education, and APOE genotype. The same demographic model used in the endophenotype analyses served as the baseline comparison. P values for odds ratios were adjusted by applying the FDR at 5%.

2.6.4. Racial/ethnic stratified analysis

To evaluate whether the associations between each CRS and outcomes differed across racial/ethnic groups, we conducted pairwise z‐tests comparing regression coefficients between subpopulations. For each score, the difference between standardized coefficients (β₁–β₂) was divided by the square root of the sum of their squared standard errors to calculate a z‐statistic and two‐sided p‐value. Comparisons were performed for all pairwise combinations of racial/ethnic groups, excluding the overall population, with significant differences highlighted at a FDR‐adjusted p < 0.05.

2.6.5. Sensitivity analysis

As a sensitivity analysis, demographic variables embedded within the CRSs (e.g., age, sex, and education) were removed from the scores and instead included as covariates in the regression models to evaluate the incremental predictive contribution of the clinical risk components beyond demographics and APOE genotype. Because these demographic factors are non‐modifiable and less clinically actionable, they were excluded from the mCAIDE, WHICAP, and CogDRisk scores when constructing the modified CRSs used in these analyses. For a detailed description and the model specification of covariates included in the CRSs for both the main and sensitivity analyses, please refer to the bottom of Table S1. We then conducted regression analyses to assess associations between the modified CRSs and both the endophenotypes and the pTau₂₁₇/Aβ₄₂ cutoff and to evaluate model predictive performance. In all linear and logistic regression models, age, sex, years of education, and APOE genotype were included as covariates. P values were adjusted for multiple hypothesis testing using an FDR of 5%.

3. RESULTS

3.1. Participant characteristics

Of the 2627 participants, 1907 were cognitively normal, and 720 were diagnosed with MCI or dementia (Table 1). Female participants (61.4%) comprised the majority in HABS‐HD. NHW individuals (N > 1062, 40.4%) were the largest racial/ethnic group, followed by LA (N = 948, 36.1%) and then Black/AA participants (N = 617, 23.5%). Mean CRS values ranged from 0 to 0.1 in cognitively normal and MCI participants, increasing to 0.3 to 0.5 in those with dementia (Table 1).

TABLE 1.

Sample characteristics of the Health and Aging Brain – Health Disparities (HABS‐HD) cohort, stratified by cognitive status.

Characteristics a	Cognitively normal	Mild cognitive impairment	Dementia
Sample size	1855	528	176
Demographics
Age	66.8 (± 7.3)	66.9 (± 8.2)	70.2 (± 8.9)
Female	1194 (64.4%)	281 (53.2%)	90 (51.1%)
NHW b	856 (46.1%)	138 (26.1%)	58 (33.0%)
LA b	649 (35.0%)	188 (35.6%)	69 (39.2%)
AA b	350 (18.9%)	202 (38.3%)	49 (27.8%)
APOE ε4 carrier	406 (25.8%)	132 (31.0%)	43 (29.7%)
Clinical risk score
mCAIDE	−0.069 (± 1.01)	0.14 (± 0.97)	0.32 (± 0.90)
WHICAP	−0.089 (± 0.97)	0.16 (± 0.93)	0.45 (± 1.27)
LIBRA	−0.090 (± 0.98)	0.19 (± 1.00)	0.36 (± 1.06)
CogDRisk	−0.066 (± 0.96)	0.05 (± 1.03)	0.53 (± 1.18)
Cognition
MMSE	28.20 (± 2.11)	26.63 (± 2.86)	21.52 (± 5.61)
CDR	0.00 (± 0.00)	1.23 (± 0.84)	4.57 (± 2.89)
Executive function	0.27 (± 0.63)	−0.52 (± 0.78)	−1.21 (± 0.91)
Verbal ability	0.24 (± 0.75)	−0.40 (± 0.75)	−1.28 (± 0.74)
Memory	0.33 (± 0.67)	−0.62 (± 0.67)	−1.57 (± 0.65)
Neuroimaging
Cortical thickness	2.76 (± 0.13)	2.73 (± 0.14)	2.60 (± 0.22)
Hippocampal volume	0.12 (± 0.92)	−0.14 (± 0.98)	−0.97 (± 1.34)
Plasma biomarkers
Aβ42/Aβ40	0.012 (± 0.99)	0.055 (± 1.05)	−0.25 (± 0.90)
Total tau	−0.016 (± 0.98)	−0.009 (± 1.01)	0.24 (± 1.22)
pTau217	−0.11 (± 0.90)	0.13 (± 1.06)	0.85 (± 1.43)
pTau217/Aβ42	−0.08 (± 0.91)	0.11 (± 1.07)	0.63 (± 1.50)
NfL	−0.078 (± 0.92)	0.055 (± 1.10)	0.66 (± 1.26)

^a Mean (SD), n (%).

^b Percentages calculated using total cohort as denominator for a given diagnosis.

For mCAIDE, mean scores for AA, LA, and NHW participants with dementia were 0.25, 0.61, and 0.060, respectively. For WHICAP, NHW participants had a mean of −0.40, AA participants had a mean of 0.1, and LA participants had a mean of 1.20 in participants with dementia. For LIBRA, mean scores were similar between LA and AA participants for MCI (0.30 and 0.20, respectively) and for dementia (0.51 and 0.50, respectively), while NHW participants had a mean of 0 for dementia and MCI diagnosis. For CogDRisk, AA, LA, and NHW participants had mean scores of 0.33, 0.63, and 0.65, respectively, in the dementia group (Table S2).

3.2. Higher clinical risk burden was associated with worsening AD endophenotypes

3.2.1. mCAIDE

Higher mCAIDE scores were associated with increased levels of NfL and total tau and decreased Aβ₄₂/Aβ₄₀ ratios in the total cohort, but not with pTau₂₁₇. In race/ethnicity‐stratified analyses, associations with higher NfL and lower Aβ₄₂/Aβ₄₀ were only significant in LA participants. For neuroimaging outcomes, mCAIDE was strongly associated with smaller hippocampal volumes across race/ethnicity groups and decreased cortical thickness in LA and NHW participants, but not in AA participants (Figure 1, Table S3). The effect sizes for mCAIDE on cortical thickness were statistically different in LA participants compared to NHW participants (z = −3.79, p = 2.10e‐3) (Table S4). For cognitive outcomes, higher mCAIDE scores were associated with lower MMSE performance and poorer verbal ability across racial groups (Figure 1, Table S3). There was a statistically significant difference in the effect size for mCAIDE on memory between LA and NHW participants (z = 2.82, p = 0.039), but no other significant differences were observed (Table S4).

FIGURE 1.

Associations of clinical risk scores (CRSs) with cognitive, neuroimaging, and plasma biomarker outcomes in the total cohort and stratified by self‐reported race/ethnicity (NHW: Non‐Hispanic White, LA: Hispanic/Latinx, AA: Black). Forest plots showing standardized beta coefficients (95% CI) for associations between four CRS (mCAIDE, WHICAP, LIBRA, and CogDRisk) and cognitive function, neuroimaging measures, and plasma biomarkers. The dashed line indicates the null (β = 0).

3.2.2. Washington Heights–Inwood Columbia Aging Project

The WHICAP score demonstrated trends similar to those of mCAIDE, with weaker or absent associations among LA and AA participants, specifically for plasma biomarker outcomes. In the total cohort, higher WHICAP scores were associated with increased NfL and total tau and decreased Aβ₄₂/Aβ₄₀. Across racial groups, NfL associations remained consistent, while other biomarkers varied. For neuroimaging outcomes, higher WHICAP was associated with decreased cortical thickness and decreased hippocampal volume (Figure 1, Table S3). The magnitude of the effect size was larger in NHW compared to AA and LA (z = AA: 11.6, p < 1.0e‐5; LA: 11.9, p < 1.0e‐5) (Table S4). For cognitive outcomes, WHICAP was associated with verbal ability and MMSE, both in the total cohort and across all racial/ethnic groups (Figure 1, Table S3).

3.2.3. Lifestyle for Brain Health

For plasma biomarkers, the LIBRA score was consistently associated with increased levels of total tau and pTau₂₁₇ and decreased levels of Aβ₄₂/Aβ₄₀ in AA participants. The effect of Aβ₄₂/Aβ₄₀ significantly differed between AA and NHW participants, with a negative association in AA and a positive association in NHW (z = −3.00, p = 2.54e‐2) (Figure 1, Table S4). Additionally, a higher LIBRA score was associated with substantially higher NfL levels in LA participants, with a significantly stronger positive association compared to the NHW participants (Table S3). For neuroimaging outcomes, a higher LIBRA score was associated with lower hippocampal volume in LA and NHW participants, but not in AA participants. A higher LIBRA score was not associated with either cortical thickness or white matter hyperintensity levels in any population. For clinical outcomes, a higher LIBRA score was associated with higher CDR and decreased executive function across racial/ethnic groups (Figure 1, Table S3).

3.2.4. CogDRisk

CogDRisk showed consistent associations across all AD endophenotypes, including plasma, neuroimaging, and cognitive domains, in the total cohort. Higher CogDRisk scores were associated with reduced hippocampal volume and cortical thickness, higher levels of plasma biomarkers, including total tau, pTau₂₁₇, and a lower Aβ₄₂/Aβ₄₀ ratio. Higher CogDRisk scores were also associated with poorer performance across all cognitive domains, including memory, verbal ability, and executive function, as well as with higher CDR scores and lower MMSE scores (Figure 1, Table S3).

For plasma biomarkers, a higher CogDRisk score was associated with increased levels of pTau₂₁₇ in LA participants (Figure 1, Table S3). This effect size was significantly greater in LA than in NHW participants (z = −2.91, p = 3.16e‐2) (Table S4). Higher CogDRisk was not associated with levels of pTau₂₁₇ or Aβ₄₂/Aβ₄₀ in AA participants. There was a significant difference in the effect size of CogDRisk on hippocampal volume between AA and NHW participants (z = 3.96, p = 1.23e‐3), with higher CogDRisk showing larger effects in AA than in NHW participants, a pattern not observed for cortical thickness (Table S4). For cognitive outcomes, a higher CogDRisk score was consistently associated with decreased MMSE, increased CDR, and decreased executive function across racial/ethnic groups (Figure 1, Table S3).

3.3. CogDRisk score demonstrates the strongest and most consistent prediction of ptau₂₁₇/aβ₄₂ positivity in a diverse cohort

We next assessed the predictive accuracy of CRSs for amyloid positivity from pTau₂₁₇/Aβ₄₂ levels, using its discriminatory threshold for amyloid PET positivity. Across the total cohort, the CogDRisk score was the only CRS associated with a 63% increase in the odds of pTau₂₁₇/Aβ₄₂ positivity, and this association was consistent across all racial/ethnic groups (NHW: 63%, LA: 60%, AA: 76%) (Figure 2, Table S5). WHICAP was also associated with increased odds of positivity (NHW = 47%, LA = 34%, AA = 61%). mCAIDE was associated with higher odds of positivity in LA and AA groups (31% and 40%, respectively).

FIGURE 2.

Predictive performance of clinical risk score (CRS) for pTau₂₁₇/Aβ₄₂ positivity, in the total cohort and stratified by race/ethnicity (NHW: Non‐Hispanic White, LA: Hispanic/Latinx, AA: Black). pTau₂₁₇/Aβ₄₂ positivity was defined using a cutoff optimized to predict amyloid PET positivity. (A) Forest plot of odds ratios (95% CI) for associations between CRS (with and without demographic components) and pTau₂₁₇/Aβ₄₂ positivity, stratified by race/ethnicity. The dashed line indicates the null (OR = 1). (B) Nagelkerke R ² values for CRS and demographic models. (C) Receiver operating characteristic (ROC) curves and corresponding area under the curve (AUC) values for each model.

The LIBRA score was not associated with pTau₂₁₇/Aβ₄₂ positivity; however, when demographic information was added to the score in sensitivity analysis, we observed significant associations across all racial/ethnic groups (all: 37%, NHW: 31%, LA: 51%, AA: 65%). This translated into modest AUC and R ² for CogDRisk in the total cohort, which are not quite as predictive as those of demographic models but are better than other CRSs (AUC: demographics = 0.72, demographics + APOE = 0.68, CogDRisk = 0.62, LIBRA = 0.52, WHICAP = 0.51, mCAIDE = 0.52; R²: demographics = 0.15, demographics + APOE = 0.095, CogDRisk = 0.0.047, LIBRA = 1.25 × 10⁻³, WHICAP = 8.01 × 10⁻⁴, mCAIDE = 2.56 × 10⁻³) (Figure 2, Table S5).

Addition of the APOE genotype to the demographic model improved discrimination for pTau₂₁₇/Aβ₄₂ positivity in the total cohort (AUC: 0.740 vs 0.687; ΔAUC = 0.053). However, this improvement was not uniform across ancestry groups and was largely driven by NHW participants (AUC: 0.755 vs 0.658; ΔAUC = 0.097). In contrast, the incremental gain was attenuated in LA participants (AUC: 0.717 vs 0.678; ΔAUC = 0.039) and negligible in AA participants (AUC: 0.691 vs 0.690; ΔAUC = 0.001). A similar pattern was observed for model fit. The inclusion of APOE resulted in a notable increase in explained variance in the total cohort and NHW group (total: ΔR ²  = 0.08; NHW: ΔR ²  = 0.14), a more modest improvement in LA participants (ΔR ²  = 0.05), and a negligible change in AA participants (ΔR ²  = 0.01) (Table S5).

3.4. Higher clinical risk burdens are associated with increased odds of cognitive impairment across racial/ethnic groups

Across all CRSs, higher scores were consistently associated with greater odds of developing dementia in the total cohort, with each SD increase in CRS corresponding to a 35% to 73% increase in odds of having dementia in the HABS‐HD cohort. WHICAP demonstrated the largest effect (73%), followed by CogDRisk (68%) and LIBRA (61%). These observations were consistent across racial/ethnic groups. When MCI and dementia were combined, effect sizes were attenuated across all CRSs relative to the dementia‐only group, due to a weaker association of CRS with MCI (Figure S1, Table S3). These associations translated into only modest predictive performance. For dementia diagnosis, the demographics model had the highest AUC of 0.70 (R ² = 0.070), which was comparable to the highest AUC achieved by CogDRisk (AUC = 0.65, R ² = 0.046). All CRSs had marginal differences in AUC, with mCAIDE having the lowest performance (AUC = 0.61, R ² = 0.020) (Figure S2, Table S6). Consistent with the overall results, model performance was only modest across racial/ethnic groups. CogDRisk achieved the highest AUC and R ² across racial/ethnic groups, ranging from 0.70 in AA participants to 0.66 in NHW participants and 0.63 in LA participants. A general pattern of higher AUC and R ² was observed in AA participants across the four CRS compared to NHW participants, who had the lowest AUC and R ² (Figure S2, Table S6).

3.5. Model performance from sensitivity analysis

For the sensitivity analysis, we excluded demographic information (i.e., age, sex, and years of education) from mCAIDE, WHICAP, and CogDRisk scores to assess their contribution to model performance. For diagnostic outcomes, higher CRS was associated with increased odds of developing dementia in the total cohort across all diagnoses and all CRSs, but this association diminished in race/ethnicity‐stratified analysis (Table S7). This translated into reduced AUC and R ² values for all CRS (mCAIDE: AUC = 0.55, R ² = 0.0037; WHICAP: AUC = 0.53, R ² = 0.032; CogDRisk: AUC = 0.60, R ² = 0.015) (Table S8). The demographic model achieved the highest AUC in the overall cohort and across subgroups, with statistically significant differences compared to both CRS and CRS sensitivity models (Table S9).

For endophenotypic outcomes, substantial decreases in the associations of all CRS were observed after excluding demographic information. In the total cohort, significant associations were observed specifically for cognitive outcome measurements but did not translate consistently into neuroimaging and plasma biomarker outcomes. Higher CogDRisk was associated with increased levels of total tau across all races/ethnicities, but all other associations with endophenotypes for all CRSs diminished in sensitivity analysis (Table S7). Consistent with the overall results, the sensitivity analysis showcased decreased AUC and R ² across races/ethnicities. CogDRisk had the highest AUC for all races/ethnicities (AA: 0.63, LA: 0.57, NHW: 0.59) compared to WHICAP and mCAIDE, in which AUC ranged from 0.51 to 0.54 across races/ethnicities (Table S8).

4. DISCUSSION

In this study, we evaluated the associations of mCAIDE, WHICAP, LIBRA, and CogDRisk with cognitive impairment and AD endophenotypes in a cross‐sectional, multiethnic cohort. Across the total cohort, higher CRSs were associated with increased odds of dementia, although effect sizes varied by race/ethnicity. Among these CRSs, a higher CogDRisk score was consistently associated with higher odds of dementia, worse AD endophenotypes, and pTau₂₁₇/Aβ₄₂ positivity across racial/ethnic groups. LIBRA also demonstrated consistent associations with AD endophenotypes despite the absence of demographic information in its scoring. WHICAP and mCAIDE performed similarly overall; however, both CRSs lacked associations with plasma biomarkers in stratified racial/ethnic analysis. Across AD endophenotypes, higher CRSs were generally associated with more severe pathological abnormalities, particularly among AA participants. These stronger associations may reflect unaccounted‐for environmental factors in the model, such as social determinants of health (SDoH) (e.g., socioeconomic status, neighborhood and built environment, healthcare access) in the population. The wide confidence intervals were also apparent due to a smaller sample size in the AA group. pTau₂₁₇/Aβ₄₂ positivity findings arose as important implications for the clinical implementation of CRSs in diverse populations. Although CogDRisk demonstrated consistent associations with pTau₂₁₇/Aβ₄₂ positivity across racial/ethnic groups, its discriminatory power still remained modest and lower than the demographic + APOE model, limiting its utility as a standalone prediction of pTau₂₁₇/Aβ₄₂ positivity being the proxy to amyloid PET positivity. Together, these findings emphasize the need for replication in larger, racially diverse cohorts with more balanced distributions of group sizes.

Model performance metrics (AUC, R ²) were highest among AA participants and lowest among NHW participants, despite CRSs being developed primarily in NHW populations. This likely reflects sample size limitations and more balanced case–control distributions in the AA group (prevalence of MCI + dementia: AA: 41.7%, LA: 28.3%, NHW: 18.6%), which can inflate performance estimates in small samples. Because AUC and R ² are sensitive to outcome prevalence, the higher prevalence of cognitive impairment among AA participants may have further contributed to optimistic estimates. These findings further highlight the need for larger, more representative datasets to yield stable, generalizable performance estimates.

Consistent with previous studies, mCAIDE was the lowest‐performing model, exhibiting weaker discriminatory power and associations. ²¹ , ²² , ²⁵ CogDRisk score, on the other hand, consistently emerged as the best‐performing model across cognitive and endophenotypic outcomes, followed closely by LIBRA, which was also observed previously. ²¹ , ²⁴ , ²⁵ The better performance of the CogDRisk and LIBRA scores is likely reflected by its broader incorporation of mid‐ to late‐life cardiometabolic, lifestyle, and psychosocial risk and protective factors that are strongly linked to dementia risk and are amenable to intervention. ⁵⁰ Given the associations of these CRSs with plasma biomarkers, combining plasma biomarker and CRS screening may be an effective way to identify at‐risk individuals in the early preclinical stages. CRSs are inexpensive, non‐invasive, and largely scalable risk assessments; however, they have limited specificity to AD pathology. With the advancing sensitivity of AD plasma biomarkers, CRSs enhance the risk identification of individuals who are vulnerable to cognitive impairment based on clinical and lifestyle factors, while biomarkers improve specificity by confirming the presence of AD‐related pathology. Genetic factors, such as APOE, offer complementary information by capturing underlying susceptibility to AD that is not reflected in clinical or biomarker measures alone. Consistent with this, APOE provided incremental predictive value for pTau₂₁₇/Aβ₄₂ positivity compared to the non‐APOE demographic model. As the field continues to move away from a single‐modality approach into multimodal treatments and studies, ⁵¹ , ⁵² , ⁵³ the proposed combined strategies would allow for a framework that links risk exposure with biological disease processes. However, successful implementation will require further investigation and consideration of the model performance in large, diverse populations, assay standardization, and the clinical actionability of effective follow‐up interventions.

Several limitations should be acknowledged. First, only four CRSs were evaluated, which does not capture the full spectrum of both clinically relevant and community‐based CRSs that have been developed over the years. ¹³ Even among the four CRSs that were evaluated, some CRS components had to be modified due to unavailable information in HABS‐HD (e.g., diet, cognitive activity), which could have affected the overall results. Second, the cross‐sectional study used in this paper limits causal inference and the ability to assess the temporal relationships of CRSs and endophenotypes. In particular, reverse causality cannot be excluded for cognitive diagnoses, as early or prodromal disease processes may influence clinical and lifestyle factors captured by CRSs. Our results showcasing relationships between CRSs and cognitive impairment should be interpreted cautiously. Future studies should examine the longitudinal trajectories of biomarker changes associated with high baseline CRSs to strengthen evidence that an elevated CRS corresponds to abnormalities in AD pathology. Lastly, this study did not include participants from Native American/Alaska Native, Asian, or Native Hawaiian/Other Pacific Islander groups. Future work should extend these evaluations to cohorts that include these populations that reflect the broader racial and socioeconomic diversity of the US population.

Despite these limitations, this study contributes novel evidence to the development of clinical risk assessment tools. By incorporating racially diverse participants, we were able to evaluate the generalizability of CRSs across populations that are often underrepresented in AD research. More importantly, this work advances the field by linking CRSs to endophenotypes, particularly plasma biomarkers, which are minimally invasive and increasingly investigated for clinical implementation. ⁶ , ⁴⁹ , ⁵⁴ , ⁵⁵ , ⁵⁶ This study adds to the growing interest in adopting multimodal dementia prevention and risk identification efforts by supporting the use of CRSs as accessible tools to help identify individuals most likely to benefit from biomarker‐based assessment.

In conclusion, the current study found that the CogDRisk score outperformed other CRSs for both associations and model performance metrics compared to LIBRA, WHICAP, and mCAIDE, in which all three showed similar performances across different analyses. However, none of the CRS models outperformed the demographics or demographics + APOE genotype models, which is expected given that age, sex, years of education, and APOE ɛ4 allele are the strongest predictors of dementia risk. This underscores that CRSs are most useful as low‐cost, widely accessible tools for preliminary risk screening, rather than as standalone predictors, and highlights the potential value of combining them with other risk measures such as genetic screenings or biomarker assessments to create a multimodal approach that may improve overall accuracy and reliability. Taken together, these findings highlight both the promises and limitations of CRSs and clarify the role of CRSs as supportive tools to complement established demographic and genetic predictors in dementia risk assessments.

CONFLICT OF INTEREST STATEMENT

J.S.Y. serves on the scientific advisory board for the Epstein Family Alzheimer's Research Collaboration and the Charleston Conference on Alzheimer's Disease and is the editor‐in‐chief of npj Dementia. Other authors report no conflicts of interest. Author disclosures are available in the Supporting Information.

CODE AVAILABILITY

All codes developed and used are available at https://github.com/AndrewsLabUCSF/CRS‐analysis. Information regarding all the software and reference datasets used in the analysis can also be found in the repository.

CONSENT STATEMENT

All aspects of the HABS‐HD study protocol are managed by the North Texas Regional Institutional Review Board. All HABS‐HD partners (and/or legal guardians) undergo informed consent and provide written informed authorization to engage in the research study.

Supporting information

Supporting Information: alz71567‐sup‐0001‐ICMJE.pdf

Supporting Information: alz71567‐sup‐0002‐Figure.docx

Supporting Information: alz71567‐sup‐0003‐Tables.xlsx