Sex and APOE ε4 genotype modify risk factor associations with cerebral amyloid angiopathy: a multi-cohort autopsy study

Abstract

Background

This study aimed to identify risk factors and develop statistical models to predict cerebral amyloid angiopathy (CAA).

Methods

Associations between demographic, cognition, cardiovascular, and AD-related neuropathology and CAA were analyzed using data from three longitudinal cohorts of aging and dementia. Logistic regression with LASSO was used for feature selection. Predictive performance was assessed using ROC-AUC and decision curve analysis (DCA). Predictor importance was quantified using Shapley Variable Importance Cloud (ShapleyVIC), which provides a robust estimate of individual feature contribution in prediction.

Results

Stratified analyses showed that the strength of association between episodic memory or tau pathology and CAA was greater in males, while the amyloid pathology-CAA association was stronger in females. Among APOE ε4 carriers, the amyloid/tau pathology-CAA associations were pronounced. Episodic memory and amyloid/tau pathology were identified as key factors in our predictive model. DCA demonstrated the model’s clinical utility, and SHAP values confirmed the importance of individual features.

Conclusion

We identified sex- and APOE-specific risk factors for CAA and developed models to support CAA risk stratification.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-025-01948-8.

Background

Cerebral amyloid angiopathy (CAA) is a cerebrovascular disease characterized by amyloid-β deposition in the walls of small- to medium-sized cerebral blood vessels [1, 2]. Currently, the Boston criteria are used to establish a probable diagnosis of CAA, while a definitive diagnosis requires postmortem histopathological examination [3]. Accurate prediction of CAA risk is critical, as it is associated with a higher likelihood of intracerebral hemorrhage, cognitive decline, and complications from anticoagulant or thrombolytic therapies [1, 4]. Identifying individuals at an elevated risk of CAA during life could inform clinical decision-making, improve individualized treatment planning, and enhance participant stratification in prevention or intervention trials. CAA is also a known risk factor for amyloid-related imaging abnormalities (ARIA), a serious side effect seen in AD patients treated with anti-amyloid monoclonal antibodies (MABs, e.g. lecanemab, aducanumab) [5–8]. Identifying individuals at higher risk of CAA may guide eligibility decisions for MAB therapy, inform monitoring strategies, and support patient counseling.

The risk of CAA increases with age and frequently coexists with Alzheimer’s disease (AD). It is also known that sex and APOE ε4 gene are risk factors for CAA. Women appear relatively protected from cerebrovascular disease until menopause, likely due to the vascular effects of estrogen, while in men, low androgen levels have been linked to increased risk [9–12]. Despite these known associations, other potential risk factors for CAA remain underexplored. To address this gap, we utilized harmonized clinical and autopsy data from three large, prospective cohort studies: the Religious Orders Study (ROS), the Rush Memory and Aging Project (MAP), and the Minority Aging Research Study (MARS). Conducted by the Rush Alzheimer’s Disease Center (RADC; https://www.radc.rush.edu/home.htm), these longitudinal cohorts are harmonized through shared study staff, unified study designs and protocols, and extensively overlapping assessment measures coordinated within the RADC [13–15]. We conducted a retrospective analysis of prospective associations between demographic factors, APOE genotype, cognitive function, and cardiovascular conditions (including stroke, heart disease, claudication, and hypertension) and the risk of CAA. In addition, we assessed cross-sectional associations between CAA risk and both vascular pathologies (arteriosclerosis, cerebral atherosclerosis, and macro- or microinfarctions) and AD-related neuropathology (AD neuropathologic changes [ADNC], Aβ plaque score, Braak stage, neuritic plaque score, and amyloid and tau densities). We further examined whether the associations between risk factors and CAA differed by sex and APOE ε4 to better identify high-risk groups. Lastly, we developed statistical predictive models incorporating both noninvasive and pathological features to evaluate their effectiveness in identifying individuals at elevated risk for CAA. Although the pathological data were obtained postmortem, they were included to explore potential surrogate biomarkers that may reflect these pathological hallmarks in vivo and inform future clinical applications.

Methods

Cohort description and data availability

The present study used data from older adults enrolled in three ongoing community-based cohorts: the ROS (since 1994) [13], the MAP (since 1997) [13], and the MARS (since 2004) [14]. The Strengthening the Reporting of Observational Studies in Epidemiology guidelines [16] for observational research were followed. All three cohorts enroll participants without known dementia at baseline. The ROS recruits older Catholic nuns, priests, and brothers across the United States. MAP includes older adults from retirement communities and subsidized senior housing facilities in the Chicagoland area and northeastern Illinois. MARS focuses on older African American participants. All participants provided informed consent for annual clinical and cognitive assessments, brain autopsy, and optional whole-body donation under the Anatomical Gift Act. Study protocols were approved by the Institutional Review Board of Rush University Medical Center. These studies share a large, item-level harmonized core of data [13–15]. Each cohort was designed to ensure comparability through common inclusion criteria, standardized clinical and neuropathologic assessments, and consistent annual follow-up schedules. All employ a unified and comprehensive data collection framework encompassing demographics, lifestyle factors, medical history, cognitive performance, and detailed postmortem neuropathologic evaluations. This harmonized structure facilitates pooled analyses across cohorts, increases statistical power, and enables robust investigation of risk factors and neuropathological correlates of aging and dementia. To access these data, investigators are required to submit a brief research proposal. Collected data are available to qualified researchers upon application and approval. For detailed information, please visit the Rush Alzheimer’s Disease Center Research Resource Sharing Hub: https://www.radc.rush.edu.

Measures

The targeted outcome in the present study was CAA. The other variables of interest were categorized into five groups: (1) demographic and genetic factors; (2) cognitive domains; (3) cardiovascular-related conditions; (4) vascular-related pathologies; and (5) AD-related neuropathology. Detailed variable information is available on the official dictionary website (https://www.radc.rush.edu/docs/var/variables.htm), and a brief description of the outcome and each group’s variables is provided below.

Cerebral amyloid angiopathy

CAA pathology was assessed semi-quantitatively in four neocortical regions (midfrontal, midtemporal, parietal, and calcarine cortices) [17]. Amyloid deposition in meningeal and parenchymal vessels was scored from 0 to 4: 0 (no deposition), 1 (scattered segmental deposition without circumferential involvement), 2 (circumferential deposition in up to 10 vessels), 3 (circumferential deposition involving up to 75% of the region), and 4 (circumferential deposition involving more than 75% of the region). Paraffin-embedded sections were immunostained for beta-amyloid using one of the three monoclonal anti-human antibodies: 4G8, 6F/3D, or 10D5. For each region, the maximum score was recorded. These regional scores were averaged to generate a continuous severity measure, which was then categorized into four levels: None (0), Mild (< 1.5), Moderate (1.5–2.5), and Severe (> 2.5). In the present study, CAA severity was dichotomized as present (mild to severe) or absent (none).

Demographic and genetic factors

Age at baseline (in years, z-scored), sex, education level (in years, z-scored), and APOE ε4 carrier status (Yes/No) were considered.

Cognitive domains

Composite scores for each cognitive domain, episodic memory, working memory, semantic memory, perceptual speed, and visuospatial ability/perceptual orientation, were calculated by converting raw test scores to z-scores based on the cohort’s baseline mean and standard deviation (SD), and then averaging the z-scores within each domain. Baseline scores were considered in the present study [18]. The tests included in each domain are detailed on the official dictionary website (https://www.radc.rush.edu/docs/var/variables.htm).

Cardiovascular-related conditions

History of hypertension, heart conditions (including heart attack, coronary thrombosis, coronary occlusion, or myocardial infarction), stroke, and claudication were all based on baseline self-report.

Vascular-related pathologies

Arteriolosclerosis was assessed histologically in small vessels of the anterior basal ganglia [19]. Histopathological features included intimal deterioration, smooth muscle degeneration, and fibrohyalinotic thickening of arteriolar walls, resulting in narrowing of the vascular lumen. The semiquantitative scale was re-coded into four levels: None (0), Mild (1), Moderate (2), and Severe (3). In the present study, arteriolosclerosis was coded dichotomously as present (mild to severe) or absent.

Cerebral atherosclerosis was evaluated by visual inspection after paraformaldehyde fixation at the Circle of Willis and included assessment of the vertebral, basilar, posterior cerebral, middle cerebral, and anterior cerebral arteries and their proximal branches [20]. Severity was categorized into four levels: None (0); Mild (1, small amounts in up to several arteries, typically less than 25% vessel involvement without significant occlusion); Moderate (2, in up to half of all visualized major arteries, with less than 50% occlusion of any single vessel); Severe (3, in more than half of all visualized arteries and/or more than 75% occlusion of one or more vessels). In this study, cerebral atherosclerosis was treated as a dichotomous variable, defined as present (mild to severe) or absent.

Macroinfarcts were defined as gross chronic cerebral infarctions visible to the naked eye on fixed brain slabs during neuropathologic examination [21]. For analysis, macroinfarcts were coded dichotomously as present (one or more chronic infarctions, regardless of location) or absent.

Microinfarcts were defined as chronic microscopic infarctions identified on hematoxylin/eosin-stained 6 µm paraffin-embedded sections [22]. A minimum of nine regions in one hemisphere were examined, including six cortical areas (midfrontal, middle temporal, entorhinal, hippocampal, inferior parietal, and anterior cingulate cortices), two subcortical regions (anterior basal ganglia, thalamus), and the midbrain. Microinfarcts were classified dichotomously as present (one or more lesions, regardless of location) or absent.

AD-related neuropathology

ADNC was classified according to the National Institute on Aging–Alzheimer’s Association consensus criteria based on Aβ plaque score, Braak score, and neuritic plaque score. The neuropathologist assigned a four-level likelihood of AD pathology: no AD (0), low (1), intermediate (2), and high (3) [23]. For analysis, ADNC was dichotomized, with 'present' defined as intermediate or high likelihood, and 'absent' defined as no or low likelihood of ADNC. In addition, we separately investigated the subscores of ADNC as binary variables: (1) Aβ plaque score – yes/no; (2) Braak score – stage 0–II/stage III–VI; and (3) Neuritic plaque score – yes/no.

Amyloid density and tangle density are continuous variables (z-scored) derived from either historical or digital data. Each is calculated as the mean of the square root–transformed pathology scores across eight regions: angular gyrus, anterior cingulate, calcarine cortex, entorhinal cortex, hippocampus, inferior temporal gyrus, midfrontal gyrus, and superior frontal gyrus [24].

Statistical analysis

Group comparison methods

Data analyses were conducted using R (version 4.4.3) and Stata (version 18.0). The normality of continuous variables (age, education, cognitive domain scores, amyloid/tangle density) was assessed using the Shapiro–Wilk test. Group comparisons of continuous variables were performed using ANOVA or the Mann–Whitney U test, depending on the normality of the data, while differences in categorical variables were assessed using Pearson’s χ² test. P < 0.05 was considered statistically significant. All continuous variables were subsequently standardized using z-scores.

Participants selection and missing data management

The original dataset included 4,642 ROS/MAP/MARS participants, of whom 1,593 had first visit and death records, complete CAA information, and a follow-up duration of ≥ 3 years. The proportion of missing data, from lowest to highest, was as follows: cerebral atherosclerosis (3/1,593; 0.19%), APOE ε4 carrier status (32/1,593; 2.01%), arteriolosclerosis (34/1,593; 2.13%), macroinfarcts and microinfarcts (39/1,593; 2.45%), ADNC (39/1,593; 2.45%), Braak stage (30/1,593; 1.88%), neuritic plaque score (30/1,593; 1.88%), tangle density (59/1,593; 3.70%), stroke (60/1,593; 3.77%), and amyloid density (68/1,593; 4.27%). These variables were imputed using the missRanger package in R, which applies a random forest-based imputation method.

Statistical models

Binary logistic regression models were used to estimate odds ratios (ORs) for both crude and adjusted analyses, evaluating the association between variables of interest, which were categorized into five major groups as previously described, and the risk of CAA. All adjusted models included potential confounders: age (years), sex (female/male), education (years), and APOE ε4 carrier status (yes/no). For both crude and adjusted models, statistical significance was defined as p < 0.05.

Subgroup analysis

For the subgroup analyses, we focused on sex- and APOE genotype–specific associations identified from the significant findings in the binary logistic regression models. Likelihood ratio tests (LRTs) were used to assess effect modification and evaluate the statistical significance of heterogeneity across subgroups. Statistical significance was defined as p < 0.05.

Statistical predictive analysis

The utility of factors from the five major groups in predicting the risk of CAA was evaluated using statistical modeling. Logistic regression was applied, and feature selection was performed using the Least Absolute Shrinkage and Selection Operator (LASSO) to identify a parsimonious combination of predictors with optimal model fit. The dataset was divided into training and test subsets using an 80–20 split. Discrimination was evaluated by calculating the receiver operating characteristic (ROC) curve and the area under the curve (AUC) in both the training and test datasets. Finally, decision curve analysis (DCA) was conducted to quantify the net clinical benefit of the predictive models across a range of probability thresholds.

To further evaluate the robustness of the predictive model, we performed a sensitivity analysis excluding ADNC-positive participants. Since ADNC frequently co-occurs with CAA [25], its presence could confound the associations and inflate model performance. Therefore, by restricting the analysis to ADNC-negative participants, we aimed to minimize this bias and determine whether the model retained its performance.

Assessing feature importance using shapleyVIC

LASSO coefficients provide an interpretable measure of feature importance but can become unstable when predictors are correlated. The dataset was split into training and test subsets using an 80/20 split. To assess each predictor’s contribution to the model’s predictions, we applied the Shapley Variable Importance Cloud (ShapleyVIC) methodology [26] within the training set. Building on the SHapley Additive exPlanations (SHAP) framework, ShapleyVIC aggregates Shapley-based feature importance across a “cloud” of near-optimal models; here we pooled estimates from 300 near-optimal (loss ε ≤ 0.05) logistic-regression models, thereby accounting for model-selection uncertainty and yielding more stable and reliable estimates of variable importance. To assess individual feature contributions to predictions, we present separate bar and violin plots. The bar chart displays pooled ShapleyVIC estimates with 95% intervals. The violin plot depicts the distribution of importance versus model performance across candidate models, conveying variability and stability. Model development was conducted in R (version 4.3.1) and Python (version 3.11.11) using the ShapelyVIC (version 1.2.0).

Results

Study characteristics

Among the 1,593 eligible participants, the mean follow-up duration was 9.67 years (SD = 5.14). Cohort characteristics stratified by CAA status are summarized in eTable S1. A radar plot visualizing the differences and statistical significance of continuous variables between participants with and without CAA is presented in Fig. 1. A total of 1,244 participants (78.09%) were diagnosed with CAA. The mean baseline age was 78.91 years (SD = 6.90), with a significantly older age observed in the CAA-positive group compared to the CAA-negative group (p < 0.0001). Males comprised 29.32% of the cohort. The proportion of APOE ε4 carriers was significantly higher in individuals with CAA (p < 0.0001).

Fig. 1.

Standardized Continuous Risk Factors by CAA Status. Radar plot illustrates the mean standardized (z-scored) values of continuous risk factors among participants with and without CAA. Significance: *p < 0.05; **p < 0.01; ***p < 0.0001

Potential risk factors and correlates of CAA

We examined the association between CAA risk and five domains: (1) demographic and genetic factors, (2) cognitive domains, (3) vascular conditions, (4) vascular pathologies, and (5) Alzheimer’s-related neuropathology. Notably, after controlling for confounders, older age (OR 1.35 [1.19–1.54]), APOE ε4 carrier (OR 3.81 [2.65–5.47]), and ADNC-positivity (OR 7.23 [5.49–9.51]) were all significantly associated with an increased CAA risk. Similarly, ADNC subscores, including Aβ plaque score (OR 6.86 [5.03–9.35]), Braak score (OR 2.88 [2.13–3.89]), and neuritic plaque score (OR 9.31 [7.02–12.34]), were also associated with an elevated risk of CAA. In addition, each SD increase in amyloid density (OR 3.04 [2.59–3.58]) and tangle density (OR 2.53 [2.06–3.11]) was linked to a greater CAA risk compared to their respective reference groups. In contrast, each SD increase in episodic memory score (OR 0.77 [0.66–0.89]) and global cognitive function score (OR 0.81 [0.70–0.94]) at baseline was significantly associated with a decreased CAA risk. Detailed results are presented in Fig. 2 and eTable S2.

Fig. 2.

Forest plot illustrating the associations between demographic characteristics, cognitive domains, genetic and vascular-related comorbidities/pathologies, AD-related pathologies, and the risk of CAA. Statistically significant positive associations are shown in red, negative associations in blue, and non-significant associations in black. Significance: **p < 0.01; ***p < 0.0001. Abbreviations: Aβ amyloid-β; ADNC Alzheimer’s disease neuropathological changes; CI confidence interval; OR odds ratio

Subgroup analysis

We explored sex- and APOE ε4 carrier status-specific associations between key risk factors and CAA.

Sex-specific

Higher episodic memory scores were associated with a lower risk of CAA in both sexes, with a stronger inverse association observed in males. ADNC-positivity was associated with an increased CAA risk in both sexes, with a slightly stronger association in females. Both amyloid and tangle-related pathologies were associated with a higher CAA risk. Notably, amyloid-related pathologies, including Aβ plaque score, neuritic plaque score, and amyloid density, showed stronger associations with CAA risk in females. In contrast, tau-related pathologies, such as Braak stage and tangle density, showed stronger associations with CAA risk in males. The interaction by sex was statistically significant for Aβ plaque score (p = 0.0384), suggesting a differential effect on CAA risk (Fig. 3).

Fig. 3.

Forest plot illustrating the associations between risk factors and CAA, stratified by sex and APOE ε4 carrier status. Statistically significant positive associations are shown in red, and significant negative associations are shown in blue. Significance: *p < 0.05; **p < 0.01; ***p < 0.0001. Abbreviations: Aβ amyloid-β; ADNC Alzheimer’s disease neuropathological changes; CI confidence interval; OR odds ratio

APOE ε4 carrier status-specific

For episodic memory, higher scores were linked to reduced CAA risk in both APOE ε4 carriers and non-carriers, with a more pronounced effect in carriers. ADNC was related to an increased CAA risk across both groups, though the magnitude of the association was greater in APOE ε4 carriers. Amyloid and tangle pathologies, including Aβ plaque score, Braak score, neuritic plaque score, amyloid density, and tangle density, were all positively associated with CAA risk in both carriers and non-carriers. Stronger associations were generally observed in APOE ε4 carriers. Significant interaction effects were detected for ADNC, Aβ plaque score, Braak score, amyloid density, and tangle density, suggesting that APOE ε4 modifies the strength of these relationships (Fig. 3).

Statistical predictive analysis

Predictor selection and construction of base and expanded models

The top predictors of CAA identified by LASSO logistic regression are presented in eTable S3 and eFigures S1 & S2. We defined a base model (B), which included age, sex, education, and APOE ε4 carrier status. Building on the selected features, we constructed the final model (B + All), which included age, sex, education, APOE ε4 carrier status, episodic memory, hypertension, amyloid and tangle density, Aβ plaque score, and neuritic plaque score. We also developed nine additional models, by adding various predictive features to the base model. Detailed information is presented in eTable S4.

Predictive model performance

The predictive performance of all models is summarized in eTable S5 and visualized in Figs. 4 & 5. The base model (B) yielded an AUC of 0.6583 in the training set and 0.6177 in the test set. Adding episodic memory (B + E) significantly improved the test set performance (AUC = 0.6514, p = 0.0067), while inclusion of hypertension (B + H) did not enhance model performance. Models that employed AD-related pathologies, including amyloid and tau density, Aβ plaque score, and neuritic plaque score, substantially improved predictive accuracy. These models achieved AUCs ranging from 0.7325 to 0.7944 in the training set and from 0.7062 to 0.7785 in the test set, with all p < 0.0001. B + AβDen + TauDen and the B + AβPlaque + NeurPlaque models also significantly outperformed the B model (all p < 0.0001). The B + All model demonstrated the highest predictive performance, with an AUC of 0.8137 in the training set and 0.7955 in the test set (both p < 0.0001), indicating a substantial improvement in predicting CAA risk.

Fig. 4.

ROC curves with AUC values for both training and test sets. Each colored curve represents a different model configuration, with performance assessed by the area under the curve (AUC). The diagonal gray line represents random chance (AUC = 0.5). B Base model (demographic + genetic); E episodic memory; H hypertension; AβDen amyloid density; TauDen tangle density; AβPlaque Aβ plaque score; NeurPlaque neuritic plaque score

Fig. 5.

Radar charts of model performance metrics in the training and test sets. Radar plots display five performance metrics: AUC, accuracy, sensitivity, specificity, and F1 score for each CAA risk predictive model in the training (left) and test (right) datasets. B Base model (demographic + genetic); E episodic memory; H hypertension; AβDen amyloid density; TauDen tangle density; AβPlaque Aβ plaque score; NeurPlaque neuritic plaque score

Evaluation of clinical utility

DCA was used to evaluate the clinical utility of predictive models. In the test set, the B and B + H models demonstrated a net benefit across a prediction threshold range of approximately 0.70–0.90, while the B + E model showed a benefit from 0.65 to 0.90. Notably, adding amyloid and tangle density to the base model extended the benefit range to 0.35–0.95 and 0.50–0.95, respectively. The B + AβPlaque and B + NeurPlaque models showed clinical utility across ranges of 0.50–0.90 and 0.35–0.95, respectively. B + E + H demonstrated a benefit across 0.70–0.95, while both B + AβDen + TauDen and B + AβPlaque + NeurPlaque showed a benefit between 0.40 and 0.90. B + All exhibited the widest range of clinical utility, from 0.35 to 0.95 (Fig. 6).

Fig. 6.

Decision curve analysis of predictive models in the training and test sets. Decision curve analysis shows the net benefit of each model across a range of threshold probabilities in the training set (left) and test set (right). The x-axis represents the threshold probability for predicting CAA risk, and the y-axis indicates the net benefit. B Base model (demographic + genetic); E episodic memory; H hypertension; AβDen amyloid density; TauDen tangle density; AβPlaque Aβ plaque score; NeurPlaque neuritic plaque score

Sensitivity analysis

After excluding ADNC-positive participants, the predictive performance of the models remained robust. The base model (B) yielded an AUC of 0.5920 in the training set and 0.4908 in the test set. After incorporating all selected predictors (B + All), the model’s discriminative ability improved substantially, achieving an AUC of 0.7233 in the training set and 0.7062 in the test set. This statistically significant improvement compared to the base model (p < 0.0001) indicated that the predictive model retained good performance even after the exclusion of ADNC-positive participants. Therefore, the observed associations were not solely driven by the comorbid ADNC pathology. However, the overall predictive performance was lower among ADNC-negative participants compared with the full cohort, which may be attributed to the limited sample size and statistical power in this subgroup.

ShapleyVIC analysis of important predictors

Figure 7 presents the findings from the SHAP analysis, displaying variables ranked by their importance in predicting CAA. The results indicate that APOE ε4 carrier status, higher amyloid and tangle burden (Aβ plaque score, neuritic plaque score, amyloid density, and tangle density), lower episodic memory performance, older age, and male sex are strongly associated with an increased CAA risk. These predictors remained robust across bootstrap replications and contributed substantially to the model performance.

Fig. 7.

ShapleyVIC summary plot for variable importance in CAA predictive models. Upper panel: bar chart of pooled ShapleyVIC estimates showing average model importance with 95% CIs; statistically significant features (p < 0.05) are marked with an asterisk. Lower panel: violin plots of SHAP value distributions across models for each variable. The equal-sized segments within each violin are colored to show the average model AUC from low (light blue) to high (dark blue)

Discussion

The present study investigated risk factors for CAA using data from 1,593 dementia-free older adults enrolled in the ROS, MAP, and MARS cohorts. We retrospectively analyzed prospectively collected data to assess the associations between demographics, APOE genetics, cognitive domains, and cardiovascular-related comorbidities and the risk of CAA. The findings indicated that older age and APOE ε4 carrier status were associated with a higher risk of CAA, consistent with previous studies [27, 28]. Notably, among the cognitive domains, higher baseline episodic memory scores in dementia-free individuals were associated with a lower risk of CAA, aligning with a previous study showing that greater CAA burden is linked to lower episodic memory [17]. Our subgroup analysis revealed that the inverse association between episodic memory and CAA risk was consistent across sex and APOE genotype but appeared stronger among males and APOE ε4 carriers. Although the interaction terms did not reach statistical significance, the observed differences suggest a possible effect modification. We also examined the cross-sectional associations between the risk of CAA, and vascular and AD-related pathologies. Individuals with ADNC had a significantly higher risk of CAA, consistent with prior findings [28]. Both amyloid-related (Aβ plaque score, neuritic plaque score, amyloid density) and tau-related (Braak stage, tangle density) pathologies were significantly associated with greater CAA risk. Notably, we observed sex-specific differences: amyloid pathology showed a stronger relationship with CAA among females, whereas tau pathology was more strongly associated with CAA among males. Sex significantly modified the association between Aβ plaque score and CAA presence. Stratified analyses by APOE genotype also revealed consistently positive associations between all AD-related pathologies and CAA, with stronger effects among APOE ε4 carriers, regardless of whether the pathology was amyloid- or tau-related. Interaction by APOE genotype was statistically significant for all AD-related markers except neuritic plaque score. Together, these findings highlight the potential vulnerability of specific subgroups to CAA-related cognitive decline and emphasize the clinical relevance of considering sex, age, APOE genotype, and cognitive performance, particularly episodic memory, as indicators that may help identify individuals at elevated risk of CAA in its early stages. This is especially important in healthcare settings where access to advanced imaging is limited. While the current study is based on post-mortem assessments of amyloid- and tau-related pathologies, advances in technologies now allow for the detection of these biomarkers through PET imaging and fluid-based assays. These tools hold promise for improving early risk stratification in clinical settings. Moreover, the observed sex- and APOE genotype–specific patterns suggest that biological and genetic differences may influence susceptibility to CAA and should be incorporated into future clinical risk models. These insights may support a more personalized monitoring and therapeutic strategy, particularly in the context of MABs for AD, where CAA burden has been linked to an increased risk of ARIA and cerebral hemorrhage/edema [6].

The current study also evaluated predictive models for CAA using demographic, genetic, cognitive, vascular, and AD-related neuropathology selected through LASSO. Adding episodic memory significantly improved test set performance compared to the base model. In contrast, adding hypertension did not enhance model performance in either the training or test sets. Models considering amyloid- and tau-related pathologies substantially improved predictive accuracy, with the comprehensive model (B + All) demonstrating the best performance and strong discrimination in both datasets. Our DCA further demonstrated that models incorporating episodic memory scores improved clinical utility, suggesting that episodic memory may serve as a promising marker for CAA risk. In addition, the inclusion of amyloid- and tau-related pathology yielded the greatest overall clinical benefit. Once reliable in vivo biomarkers for these pathologies are validated, they could serve as surrogates, further enhancing the performance and applicability of predictive models in clinical settings. SHAP analysis confirmed the importance of several features selected by LASSO, including amyloid and tau burden, APOE genotype, age, and episodic memory, which further supports the robustness of the model selection process. Although hypertension was selected as a feature by the LASSO algorithm, it did not enhance model performance and contributed minimally according to the SHAP analysis. This consistency across methods reinforces confidence in our final model.

Our study has several strengths. Potential selection bias was minimized through two key design features: a high autopsy participation rate and the blinding of neuropathologists to clinical diagnoses in the ROS/MAP/MARS cohorts. The high autopsy rate further enhances validity by enabling accurate neuropathological confirmation of disease status, including AD-related pathologies and CAA, thereby reducing misclassification and strengthening both epidemiological inference and predictive modeling. In addition, community-based recruitment improves the representativeness of the cohort, enhancing the generalizability of the findings. The large number of CAA cases increased statistical power, allowing for more precise effect estimates and robust stratified analyses by key subgroups such as sex and APOE genotype. Moreover, the consistency of results across epidemiological analyses, predictive modeling, and SHAP interpretation provides strong support for the robustness of our findings. Another notable strength is the inclusion of a sensitivity analysis excluding ADNC-positive participants, which enabled us to assess whether the model’s predictive performance was influenced by the comorbid AD pathology. The model maintained a good discriminative ability within the ADNC-negative subgroup. This finding underscores the robustness of the identified predictors and suggests that the associations with CAA were not solely driven by concomitant AD pathology. Nonetheless, predictive performance was somewhat attenuated compared with the full cohort, likely reflecting the smaller sample size and reduced statistical power in the ADNC-negative group. Future studies with larger and more diverse samples are warranted to confirm these results and to further disentangle CAA-specific mechanisms from AD-related pathology.

This study has several limitations. First, the cross-sectional nature of postmortem data precludes causal inference between pathologies and CAA risk. Second, selection bias may have arisen from participants who declined brain donation or lacked autopsy data. Third, the cohort’s predominantly Caucasian composition limits the generalizability of findings to more diverse populations. Fourth, self-reported baseline comorbidity histories may introduce recall bias. Fifth, although all participants were dementia-free at baseline, some had mild cognitive impairment, and the limited sample size prevented separate analyses excluding these individuals. In addition, the unbalanced sex distribution may restrict the generalizability of sex-specific findings and introduce bias in stratified analyses.

Conclusions

In conclusion, this study highlights the potential of integrating cognitive performance, APOE genotype, and pathological data to enhance CAA risk stratification. Future research should aim to validate these predictive models in more diverse populations and investigate the temporal progression of CAA using longitudinal, multimodal data. The discovery of accessible biomarkers, combined with machine learning approaches, could further increase the clinical utility of these tools. Ultimately, translating such models into routine clinical practice will be critical for enabling early intervention and personalized treatment strategies, particularly for individuals at elevated risk of CAA-related complications.

Abbreviations

AD

Alzheimer’s disease

ADNC

Alzheimer’s disease neuropathological changes

ARIA

Amyloid-related imaging abnormalities

AUC

Area under the curve

CAA

Cerebral amyloid angiopathy

DCA

Decision curve analysis

LASSO

Least Absolute Shrinkage and Selection Operator

LRT

Likelihood ratio test

MABs

Monoclonal antibodies

MAP

Rush Memory and Aging Project

MARS

Minority Aging Research Study

ROC

Receiver operating characteristic

ROS

Religious Orders Study

SHAP

SHapley Additive exPlanations

Authors’ contributions

Concept and design of the study: Liwei Ma (L.M.), Benjamin Goudey (B.G.), Liang Jin (L.J.), and Yijun Pan (Y.P.). Data analysis and interpretation: L.M., (Yihan Wang) Y.W., and Y.P. Drafting of the manuscript: L.M., Y.W., and Y.P. All authors critically revised the draft and contributed substantially to the final version. All authors have read and approved the final manuscript.

Funding

Dr Yijun Pan received support from the National Health and Medical Research Council (Australia) Investigator Grant (GNT2007912), NHMRC-AMED Dementia Collaborative Research Grant (GNT2022203), and Alzheimer’s Association USA (23AARF-1020292).

Data availability

Data requests should be made to Rush University (radc.rush.edu). The analysis files to support the results of this manuscript can be made available upon written request to the corresponding authors, which may also be subject to approval by the Rush University.

Declarations

Ethics approval and consent to participate

ROS, MAP, MARS studies were approved separately by the Rush University Medical Center Institutional Review Board. Written informed consent was provided to participants who signed an Anatomical Gift Act. This study involved secondary analysis of de-identified data and was approved by the Monash University Human Research Ethics Committee (Project 48297).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.