Cross‐sectional and longitudinal cognitive trajectories associated with dementia comorbidity clusters: Results from a 10‐year follow‐up real‐world outpatient memory clinic in Brazil

Abstract

INTRODUCTION

We identified comorbidities associated with dementia and assessed the impact of comorbidity clusters on dementia patients’ cognitive trajectories in a real‐world Brazilian memory‐clinic cohort.

METHODS

We analyzed 500 individuals. Ridge‐penalized logistic regressions assessed associations between comorbidities and dementia. Tetrachoric correlations and hierarchical clustering informed mixed‐effects analyses of longitudinal trajectories.

RESULTS

Fazekas score ≥ 2 (odds ratio [OR] 2.48, 95% confidence interval [CI] 1.64–4.54), stroke (OR 2.15, 95% CI 1.23–3.23), alcohol abuse (OR 2.13, 95% CI 1.22‐4.92), hearing loss (OR 1.99, 95% CI 1.06–3.94), visual loss (OR 1.94, 95% CI 1.04–4.01), and low education (OR 1.21, 95% CI 1.09–1.54) were associated with dementia cross‐sectionally. Comorbidities formed four clusters: cerebrovascular, cardiometabolic, neurosensory, substance abuse. Higher cerebrovascular burden predicted a 158% faster rate of cognitive decline in individuals with dementia.

DISCUSSION

Our findings show four dementia comorbidity clusters and identify cerebrovascular burden as a prognostic marker of faster cognitive decline.

Highlights

• Fazekas score ≥ 2, stroke, alcohol abuse, hearing loss, visual loss, and low education each showed independent associations with dementia.

• Associations of Fazekas score ≥2 and stroke with dementia persisted after excluding vascular‐dementia cases.

• Comorbidities formed four clusters in a real‐world Brazilian memory‐clinic cohort: cerebrovascular, cardiometabolic, neurosensory, and substance abuse.

• Higher cerebrovascular burden predicted a faster rate of cognitive decline in individuals with dementia.

1. BACKGROUND

In low‐ and middle‐income countries (LMICs), population aging often coincides with multimorbidity – particularly cardiometabolic and cerebrovascular diseases – and with lower cognitive reserve from limited educational access. ¹ , ² These conditions can increase brain vulnerability and accelerate cognitive deterioration through direct neuronal injury, vascular compromise, and inflammatory processes. ³ , ⁴ , ⁵ Evidence shows that comorbidities interact synergistically rather than additively, yielding distinct cognitive phenotypes with different prognoses. ⁶ , ⁷

The latest Lancet Commission identified 14 key risk factors for dementia in high‐income settings, and models replicated in Latin America have shown a higher overall burden in lower‐resource populations. ⁸ , ⁹ , ¹⁰ , ¹¹ These single‐factor analyses estimate population‐level risk and prevention potential. However, evidence remains limited on how comorbidities interact in routine care and how those interactions affect clinical outcomes. Because comorbidities account for a large share of dementia cases – reflected in high population‐attributable fractions – the open question is whether and how much multimorbidity influences cognitive trajectories following a dementia diagnosis. This remains uncertain, particularly in Latin American memory clinics where comorbidity patterns and health‐system constraints differ, underscoring the need for direct investigation in real‐world cohorts. ¹²

Brazil provides a suitable setting to address this question as a large LMIC with demographic heterogeneity and universal coverage through the Unified Health System (SUS). In a 10‐year real‐world retrospective cohort from a Brazilian tertiary memory clinic, we aimed to (1) estimate prevalence and association of comorbidities across the cognitive continuum, and (2) test whether comorbidity clusters could impact cognitive trajectories in individuals with dementia.

2. METHODS

2.1. Participants and data collection

We retrospectively reviewed electronic clinical reports of individuals evaluated at the Neurology Memory Clinic of São Lucas Hospital, a tertiary institution in Porto Alegre, Brazil, that provides public healthcare services under the SUS. We included all consecutive patients with a baseline clinical evaluation between January 2014 and December 2024. We excluded patients with fewer than two cognitive evaluations during follow‐up or those who were inappropriately referred at initial assessment. Of 521 patients, 21 were excluded: 13 for insufficient follow‐up and 8 for inappropriate referral. The final sample comprised 500 individuals. Patients were referred to the memory clinic primarily from primary healthcare services. Three independent authors reviewed the electronic medical records between October 2023 and January 2025. To ensure data consistency, each author reviewed the extracted information twice.

2.2. Clinical evaluation

At baseline, all participants underwent a standardized clinical evaluation and were stratified into three clinical syndromes: subjective cognitive decline (SCD), mild cognitive impairment (MCI), and dementia. Clinical syndromes were defined by a discussion panel comprising two cognitive and behavioral neurologists, neurology resident physicians, and neuropsychologists. All individuals underwent diagnostic work‐up with cognitive assessment, laboratory tests, and brain magnetic resonance imaging (MRI). Structural MRI scans were formally reported by a board‐certified neuroradiologist, who scored medial temporal atrophy on T1‐weighted sequences using the medial temporal atrophy (MTA) scale ¹³ and white matter hyperintensity burden on T2–fluid‐attenuated inversion recovery (FLAIR) sequences using the Fazekas scale. These sequences were part of the routine clinical MRI protocol and were available for all patients included in the study. Additionally, two cognitive‐behavioral neurologists also reviewed the images and the neuroradiologist clarified uncertainties when identified by the reviewing neurologists. Given the low educational level of the study population, the Mini‐Mental State Examination (MMSE) was the primary cognitive test employed, although additional neuropsychological tests were used when necessary.

2.3. Comorbidity assessment

Comorbidities were identified from International Classification of Diseases (ICD) codes and routine‐care chart documentation and coded as binary indicators (present/absent) for the clustering procedure and burden‐score construction. ¹⁴ These conditions reflect clinically documented history and were treated as baseline prevalent conditions. For alcohol use and smoking, we coded patients with documented history and ongoing use at baseline. For hearing and visual loss, we coded impairments documented as persistent. For hypertension, diabetes, and dyslipidemia, we coded the documented clinical diagnosis regardless of current control status. We used binary coding because severity and duration were not consistently available across the cohort. This approach maximized completeness and harmonized measurement across conditions within the same analytic framework. When a graded measure was routinely available (e.g., Fazekas score), we preserved clinical granularity through anchored thresholds and derived ordinal burden categories for the longitudinal models. We defined low education as ≤ 6 years of formal schooling. ¹⁰ , ¹⁵ We quantified white matter lesion burden with the Fazekas scale, where a score ≥ 2 denoted high burden. ¹⁶ , ¹⁷ Ocular conditions such as cataract, glaucoma, or age‐related macular degeneration were used as proxy measures for visual loss, and visual impairment was classified as yes/no. ¹⁸ Lastly, stroke (ischemic or hemorrhagic) was defined by documented clinical history plus imaging evidence – MRI – of acute or chronic infarction, or intracerebral hemorrhage; we included all TOAST ischemic subtypes (large‐artery, cardioembolic, small‐vessel/lacunar, other determined, undetermined), in both cortical and subcortical topographies. ¹⁹

2.4. Cluster analysis and burden scores

We first computed a tetrachoric correlation matrix across the 11 binary comorbidities and removed variables with |ρ| ≤ 0.30 after Benjamini–Hochberg correction (p < 0.05) to discard weak/unstable latent associations and to improve cluster cohesion and interpretability. We then applied hierarchical agglomerative clustering using Jaccard distance as the dissimilarity metric and complete linkage as the linkage criterion. Among k = 2–5, the four‐cluster solution achieved the highest average silhouette (0.51). We assessed stability with 1000 bootstrap resamples (80% subsampling) and retained clusters with mean Jaccard bootstrap stability ≥ 0.75 (observed 0.79–0.85). This two‐step workflow yielded four clinically interpretable comorbidity domains for the burden‐score analyses. These domains (cerebrovascular, cardiometabolic, substance abuse, and neurosensory) should be interpreted as descriptive, data‐driven co‐occurrence clusters rather than mutually exclusive clinical entities. For example, the cerebrovascular domain captures manifest vascular injury (stroke history and white matter hyperintensity burden), whereas the cardiometabolic domain captures established vascular risk factors (hypertension, diabetes, dyslipidemia) that confer elevated risk but do not necessarily result in overt brain damage. Because these processes overlap biologically, longitudinal models were adjusted for comorbidities from other clusters to isolate each domain's independent association with cognitive slope.

Within the cerebrovascular cluster – which combined stroke history and Fazekas score – we designated low burden for patients without stroke history and Fazekas < 2, moderate burden for those with either a prior stroke or Fazekas ≥ 2, and high burden for patients meeting both criteria, thereby isolating the highest‐risk group without unduly shrinking reference sizes. In the cardiometabolic cluster (hypertension, diabetes, dyslipidemia), we defined low burden as zero or one condition and high burden as two or three conditions, aligning with established multimorbidity thresholds and dose–response evidence linking multiple cardiometabolic factors to accelerated cognitive decline. ¹⁰ , ²⁰ , ²¹ , ²² , ²³ Finally, for clusters of rare or inherently binary conditions – such as substance abuse (alcohol abuse and smoking) and neurosensory (hearing loss and visual loss) – we grouped all affected patients into a single high burden category to ensure stable model estimates. Depression and low education did not form stable domains and were modeled separately. A full description of this clustering workflow and variance inflation factor (VIF) diagnostics appears in the Supplementary material (E1, Tables S1–S6, Figure S1).

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional sources. Prior work has focused on individual risk factors with limited clinic‐based evidence on how clustered comorbidities shape cognition after a dementia diagnosis. Few studies integrated cluster methodologies with longitudinal cognitive assessment in real‐world settings.

2. Interpretation: We found that comorbidities consistently reported in prior population studies were each independently associated with dementia in a real‐world memory clinic. We identified comorbidity clusters that mirror routine care. Higher cerebrovascular burden was a predictor of faster cognitive decline in patients with dementia, indicating a potential prognostic marker.

3. Future directions: Future studies should test cluster‐level synergies beyond single‐factor effects. Large multi‐setting cohorts can validate the prognostic value of cerebrovascular burden and quantify effects on other pragmatic outcomes, increasing external validity across health systems.

2.5. Statistical analysis

All statistical analyses were conducted using R version 4.2.1 (R Foundation). At baseline, we compared demographic and clinical characteristics across clinical syndromes – SCD, MCI, and dementia – using one‐way analyses of variance (ANOVAs) with Tukey post hoc tests for continuous variables (age and educational attainment), and Pearson's chi‐square test for categorical variables (sex). Additionally, we performed an analysis of covariance (ANCOVA) to compare MMSE baseline scores across diagnostic groups, adjusting for age and sex as covariates.

2.6. Cross‐sectional analysis

We performed three ridge‐penalized (L2) logistic regression models to investigate the association between baseline comorbidities and clinical syndrome. The first model compared individuals with dementia to those with SCD (reference group); the second model repeated the first regression model excluding participants with a clinical diagnosis of vascular dementia, thereby providing a sensitivity analysis to test whether associations for stroke and white matter lesions persist in non‐vascular dementia cases. The third model compared individuals with MCI to those with SCD (reference group). Each model included binary comorbidities – hypertension, diabetes, dyslipidemia, stroke, white matter lesions, depression, smoking, alcohol abuse, hearing loss, visual loss and low education – adjusted for age and sex. The rationale for these models and the performance comparison with unpenalized models are provided in the Supplementary material (E2, Table S7, Figures S2–S4).

2.7. Longitudinal analysis

For the longitudinal analysis, we employed linear mixed‐effects models to evaluate whether the interaction between time and baseline comorbidity burden scores significantly influenced cognitive trajectories over 18 months, while controlling for demographic and clinical covariates. The longitudinal component was restricted to those patients with dementia, given that MCI and SCD patients presented reduced sample sizes, high attrition rates, irregular follow‐up intervals, and a short observation period. We selected this interval for two primary reasons. First, in real‐world clinical settings, follow‐up intervals vary, and patient attrition is common, which requires a timeframe that accommodates such variability. Second, preliminary analyses indicated that an 18‐month duration provided optimal stability for our models. We fit four separate linear mixed‐effects models – one per burden domain (cerebrovascular, cardiometabolic, substance abuse, neurosensory) – to 18‐month cognitive trajectories. Each model included fixed effects for centered months since baseline, age, sex, education, and MTA score; the domain‐specific burden category; and a burden × time interaction. In each model, we added time interactions for comorbidities from other clusters and for age, sex, education, and MTA score, thereby isolating the cluster‐specific effect on cognitive slope. We specified subject‐level random intercepts and slopes with an unstructured covariance to account for individual variability. In sensitivity analyses, we evaluated potential non‐linear cognitive trajectories by incorporating quadratic time terms and their interactions with burden scores within the mixed‐effects framework for each comorbidity cluster. Full model specifications, including sensitivity analyses with quadratic time terms, can be found in the Supplementary material (Tables S8–S14). Trajectories of patients with depression (Figure S5) and low education (Figure S6) can also be found in the Supplementary material.

3. RESULTS

3.1. Demographic and clinical characteristics

Of the 500 individuals evaluated (mean [SD] age at baseline = 70.30 (11.20) years), 264 (52.80%) presented with dementia, 155 (31.00%) MCI, and 81 (16.20%) SCD. Women constituted the majority across all diagnostic groups (313/500, 62.60%), with similar proportions across diagnostic categories (SCD: 58/81, 71.60%; MCI: 101/155, 65.16%; dementia: 154/264, 58.33%). Educational attainment was notably low, with an overall mean (SD) of 5.08 (3.72); the lowest mean was observed in patients with dementia: 4.69 (3.52) years. Hypertension was the most prevalent comorbidity, affecting approximately 70% of participants across all groups (SCD: 67.90%; MCI: 73.55%; dementia: 73.48%). Depression showed the highest rates in MCI patients (70.97%) compared to those with SCD or dementia, while low education affected more than half of participants across groups (59.25%‐72.34%). Stroke demonstrated a clear severity gradient, increasing from SCD (22.22%) through MCI (24.52%) to dementia (40.53%). Similarly, alcohol abuse showed progressive increases with cognitive decline (SCD: 3.70%; MCI: 12.25%; dementia: 15.15%), as did visual loss (SCD: 7.40%; MCI: 11.61%; dementia: 22.34%) and hearing loss (SCD: 8.64%; MCI: 12.90%; dementia: 26.13%). In contrast, diabetes and smoking showed consistent prevalence across diagnostic categories (22.22%‐29.68% and 20.08%‐23.23%, respectively). Descriptive characteristics are shown in Table 1, and Supplementary material (Table S15) details less‐prevalent dementia diagnoses and secondary etiologies.

TABLE 1.

Baseline characteristics of the study population. Variables are expressed in mean (SD) (continuous), and frequency (%) (categorical).

Parameter	SCD (n = 81)	MCI (n = 155)	Dementia (n = 264)	Overall (n = 500)
Age, years	63.90 (11.60)	66.90 (10.50)	74 (10.30) a , b	70.30 (11.20)
Sex
Female, No. (%)	58 (71.60)	101 (65.16)	154 (58.33)	313 (62.60)
Education, years	5.58 (4.02)	5.48 (3.84)	4.69 (3.52)	5.08 (3.72)
MMSE baseline score	25.60 (3.40)	22.30 (5.30) a	17.30 (6.40) a , b	20.50 (6.40)
Fazekas score
0, No. (%)	34 (41.98)	31 (20.00)	49 (18.56)	114 (22.80)
1, No. (%)	28 (34.56)	54 (34.83)	44 (16.66)	126 (25.20)
2, No. (%)	14 (17.28)	55 (35.48)	82 (31.06)	151 (30.20)
3, No. (%)	5 (6.17)	15 (9.67)	89 (33.71)	109 (21.80)
MTA score
0, No. (%)	48 (59.25)	53 (34.19)	14 (5.30)	115 (23.00)
1, No. (%)	13 (16.04)	15 (9.67)	16 (6.06)	44 (8.80)
2, No. (%)	16 (19.75)	46 (29.67)	64 (24.24)	126 (25.20)
3, No. (%)	3 (3.70)	36 (23.22)	105 (39.77)	144 (28.80)
4, No. (%)	1 (1.23)	5 (3.22)	65 (24.62)	71 (14.20)
Clinical comorbidities
Alcohol abuse, No. (%)	3 (3.70)	19 (12.25)	40 (15.15)	62 (12.40)
Depression, No. (%)	53 (65.43)	110 (70.97)	144 (54.54)	307 (61.40)
Diabetes, No. (%)	18 (22.22)	46 (29.68)	66 (25.00)	130 (26.00)
Dyslipidemia, No. (%)	40 (49.38)	87 (56.13)	163 (61.74)	290 (58.00)
Hearing loss, No. (%)	7 (8.64)	20 (12.90)	69 (26.13)	96 (19.20)
Hypertension, No. (%)	55 (67.90)	114 (73.55)	194 (73.48)	363 (72.60)
Low education, No. (%)	48 (59.25)	109 (70.32)	191 (72.34)	348 (69.60)
Smoking, No. (%)	17 (21.00)	36 (23.23)	53 (20.08)	106 (21.20)
Stroke, No. (%)	18 (22.22)	38 (24.52)	107 (40.53)	163 (32.60)
Visual loss, No. (%)	6 (7.40)	18 (11.61)	59 (22.34)	83 (16.60)
Presumptive clinical diagnosis
Alzheimer's disease (AD)	NA	NA	97 (36.74)	97 (19.40)
Vascular dementia	NA	NA	44 (16.66)	44 (8.80)
Mixed dementia (AD & vascular)	NA	NA	42 (15.90)	42 (8.40)
Dementia with lewy bodies (DLB)	NA	NA	17 (6.43)	17 (3.40)
Frontotemporal dementia	NA	NA	28 (10.60)	28 (5.60)
Other dementia c	NA	NA	36 (13.63)	36 (7.20)
Psychiatric disorders	20 (24.70)	34 (21.94)	NA	54 (10.80)

Note: Tests: One‐way ANOVA with Tukey post‐hoc (age and educational attainment); analysis of covariance – ANCOVA (MMSE); Pearson's chi‐squared (sex).

Abbreviations: AD, Alzheimer's disease; ANCOVA, analysis of variance; DLB, dementia with Lewy bodies; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; NA, not applicable; SCD, subjective cognitive decline; SD, standard deviation; y, years; MTA, medial temporal atrophy

^a p < 0.0001 vs SCD.

^b p < 0.0001 vs MCI.

^c “Other dementia” includes unspecified etiology and dementia due to other conditions. Imaging scales: Fazekas 0‐3 (higher = greater white matter hyperintensity burden); MTA 0‐4 (higher = greater medial temporal atrophy). Presumptive diagnoses appear only for the Dementia group (NA for SCD and MCI).

3.2. Comorbidities associated with cognitive status

Stroke (OR 2.15; 95% CI, 1.23–3.23; p = 0.0022), Fazekas ≥ 2 (2.48; 1.64–4.54; p = 0.00047), alcohol abuse (2.13; 1.22–4.92; p = 0.038), hearing loss (1.99; 1.06–3.94; p = 0.033), visual loss (1.94; 1.04–4.01; p = 0.047), and low education (1.21; 1.09–1.54; p = 0.031) were associated with dementia (Figure 1, Model 1). After excluding vascular dementia, associations persisted for stroke (1.68; 1.09–2.79; p = 0.031), Fazekas ≥ 2 (2.34; 1.54–3.97; p = 0.00043), alcohol abuse (2.01; 1.14–4.80; p = 0.049), hearing loss (1.92; 1.10–3.94; p = 0.045), visual loss (2.01; 1.07–4.15; p = 0.043), and low education (1.18; 1.07–1.46; p = 0.037) (Figure 1, Model 2). In the MCI versus SCD comparison, only Fazekas ≥ 2 was associated with MCI (1.80; 1.20–3.60; p = 0.036) (Figure 1, Model 3).

FIGURE 1.

Forest plots show adjusted ORs with 95% CIs from three ridge‐penalized logistic models: Model 1 (dementia vs. SCD), Model 2 (Dementia vs. SCD, excluding vascular dementia), and Model 3 (MCI vs. SCD). Points are ORs; bars are 95% CIs; the dashed line marks OR = 1. All models include the same binary predictors – hypertension, diabetes, dyslipidemia, stroke, white matter hyperintensities (Fazekas ≥ 2), depression, smoking, alcohol abuse, hearing loss, visual loss, and low education – and adjust for age and sex; λ was chosen by cross‐validation. p‐Values are from the ridge‐specific score test (Cule's method). CIs, confidence intervals; ORs, odds ratios; SCD, subjective cognitive decline.

3.3. Comorbidity clusters

Clustering analysis revealed four distinct comorbidity patterns (Figure 2, panel b). Cerebrovascular burden demonstrated the strongest clustering, with white matter hyperintensities (Fazekas score ≥ 2) and stroke history showing high correlation (rtet = +0.79, p < 0.0001). Cardiometabolic conditions clustered significantly, with hypertension–dyslipidemia showing the strongest association (rtet = +0.67, p < 0.0001), followed by diabetes–dyslipidemia (rtet = +0.51, p < 0.0001) and diabetes–hypertension (rtet = +0.49, p < 0.0001). Neurosensory impairments co‐occurred (hearing and visual loss: rtet = +0.41, p < 0.0001), as did substance abuse behaviors (alcohol and smoking: rtet = +0.63, p < 0.0001) (Figure 2, panel c).

FIGURE 2.

(A) The tetrachoric correlation matrix (rtet) for 11 binary comorbidities (cell values are coefficients). (B) Hierarchical agglomerative clustering of co‐occurrence using Jaccard dissimilarity (1–Jaccard) and complete linkage; the k = 4 solution maximized average silhouette (0.512) and showed bootstrap stability (Jaccard 0.79–0.85 across 1000 resamples with 80% subsampling). (C) Summarizes domain strength as the mean within‐domain rtet with Benjamini–Hochberg‐adjusted p values. Domains: cerebrovascular (stroke, white matter hyperintensities [WMH]; Fazekas ≥ 2), cardiometabolic (hypertension, diabetes, dyslipidemia), substance abuse (smoking, alcohol), and neurosensory (hearing, visual loss).

3.4. Longitudinal cognitive trajectories by comorbidity burden

Cerebrovascular burden had a significant impact on the rate of cognitive decline (Figure 3A). Individuals with both stroke history and a Fazekas score ≥ 2, considered as high burden, declined 0.43 MMSE points per month, representing a 158% faster rate of cognitive decline when compared to those with no burden (p < 0.0001 for the burden × time interaction vs. no burden). Participants classified as moderate burden (stroke history or Fazekas score ≥ 2) showed an intermediate decline rate 0.26 MMSE points per month (p = 0.048 for the burden × time interaction vs. no burden) (Figure 3A). Several additional comorbidity patterns trended toward association with altered cognitive decline, although they did not reach statistical significance (Figure 3B–D). Sensitivity analyses incorporating quadratic time terms showed non‐significant effects across all clusters. The substance abuse cluster showed the most quadratic activity (likelihood ratio test p = 0.20), though model fit did not significantly improve compared with the primary linear specification (ΔBIC +10.2 to +16.2 across clusters) (Table S14).

FIGURE 3.

Model‐predicted MMSE slopes from linear mixed‐effects models with random intercepts and slopes derived from the four clusters: (A) cerebrovascular, (B) cardiometabolic, (C) substance abuse, (D) neurosensory. Lines show fitted means; bands, 95% CIs. All models adjust for comorbidities from other clusters and for age, sex, educational attainment and MTA score; non‐cerebrovascular panels additionally adjust for stroke and Fazekas ≥ 2. Slope differences are tested via the time × burden interaction (reference noted in each panel). Burden definitions: cerebrovascular – none (no stroke & Fazekas < 2), moderate (stroke or Fazekas ≥ 2), high (both); cardiometabolic – low (≤ 1 of hypertension/diabetes/dyslipidemia) vs. high (≥ 2); substance – none versus any smoking/alcohol abuse; neurosensory – low versus any visual/hearing impairment. MMSE, Mini‐Mental State Examination; MTA, medial temporal atrophy.

4. DISCUSSION

In this study of 500 Brazilian memory clinic patients, we estimated comorbidity prevalence and associations across the cognitive continuum and tested whether comorbidity clusters influence cognitive trajectories in dementia. Overall, our data show that comorbidity was prevalent, and that stroke, white matter lesions, alcohol abuse, hearing loss, visual loss, and low education were each independently associated with dementia. Additionally, we found that higher cerebrovascular burden was a predictor of faster cognitive decline in patients with dementia. Altogether, our data suggest that comorbidities interact synergistically and continue to shape cognition after dementia diagnosis. This underscores the need for tailored primary‐care interventions that target the most impactful clusters to reduce the dementia burden in resource‐limited settings.

We found that comorbidities were common at every cognitive stage. Low educational attainment affected more than two‐thirds of participants across all diagnostic groups, representing a key brain health disparity factor in this Latin American cohort. Low education represents a critical finding given our sample's exceptionally low educational attainment (mean 5.08 years). Latin American studies identify education as one of the most powerful protective factors against dementia. ²⁴ , ²⁵ In our study, low education did not cluster significantly with other risk factors. Instead, it showed an independent effect on cognitive status, with a modest association with dementia in ridge‐penalized models. The limited educational range in our sample – nearly 70% had low education – likely reduced our ability to detect stronger effects. Notably, we found no significant differences in cognitive decline rates based on educational attainment. This aligns partially with cognitive reserve theory: education may delay dementia onset but accelerate decline once pathological processes reach a clinical threshold. ¹⁵ , ²⁶ Altogether, this result suggests that low education attainment reflects cognitive reserve vulnerability, potentially magnifying other risk factors’ impact.

The associations with stroke and moderate‐to‐severe white matter lesions persisted in sensitivity analysis that excluded patients with vascular dementia, indicating that cerebrovascular pathology contributes to dementia across etiologies. This pattern aligns with recent evidence that vascular injury and neurodegeneration co‐occur along the Alzheimer's disease continuum and supports considering cerebrovascular burden when interpreting diagnosis and progression in routine care. ²⁷ , ²⁸ By contrast, we did not observe a positive hypertension–dementia association in the cross‐sectional model. Although hypertension was more frequent in the dementia group, the multivariable model yielded a non‐significant inverse association, likely reflecting age adjustment and competing effects of stroke and white matter lesions. The life‐course dynamics of blood pressure indicate that it rises in midlife, plateaus, and declines prior to dementia diagnosis. Therefore, cross‐sectional ascertainment based on prior diagnoses may miss the risk window. ²⁹ , ³⁰ Because our hypertension measure relied on historical diagnostic records rather than contemporaneous blood pressure readings, it cannot capture these dynamics. Longitudinal tracking from midlife would clarify this temporal relationship.

We further demonstrate that patients with both clinical stroke and severe white matter hyperintensities experienced cognitive decline 158% faster than those without these conditions, suggesting synergistic interaction between large and small vessel disease in accelerating cognitive deterioration. Previous work has established that cerebral small vessel disease and stroke each contribute to cognitive decline through disrupted network connectivity and axonal injury. ³¹ , ³² , ³³ , ³⁴ The small vessel disease effect appears to be stronger in Latin America, where healthy adults and dementia patients carry a higher burden of white matter hyperintensities with broader coupling to cortical thinning. ³⁵ Our study expands these findings by examining stroke and white matter disease as a cluster rather than as isolated factors, supporting the notion that these factors combined, rather than individually, have a greater impact on post‐diagnosis cognitive trajectories. Ultimately, these findings offer valuable insights for public health policy‐makers to prioritize tailored interventions targeting the most impactful clusters in primary care settings.

We found no significant associations between the cardiometabolic cluster and cognitive outcomes, contrary to previous reports. ²³ , ³⁰ Several factors may explain these discrepancies. First, the very high rates of hypertension and dyslipidemia across all diagnostic groups create measurement challenges and potential ceiling effects, limiting our ability to detect independent cardiometabolic influences. Second, the relationship between metabolic disorders and cognitive outcomes often operates through complex pathways requiring longer follow‐up periods than our 18‐month window. ³⁶ Third, the pronounced impact of cerebrovascular pathology in our population may overshadow more subtle effects of metabolic dysregulation through competing risk mechanisms. Similarly, the neurosensory cluster showed a trend toward accelerated cognitive decline that did not reach statistical significance. The correlation between hearing and visual impairments suggests shared underlying mechanisms of age‐related sensory deterioration, supporting the “Common Cause” hypothesis where domain‐general neuronal degradation simultaneously affects both sensory modalities through age‐related changes in central processing capacity. ³⁷ These findings are relevant given growing evidence that sensory impairments represent significant modifiable risk factors for dementia. ¹⁰

Some limitations regarding our findings warrant consideration. First, our memory clinic population represents patients who successfully navigated healthcare system barriers to reach specialty care in a tertiary hospital setting, which limits generalizability even within the Brazilian context, as dementia care is mostly provided in primary care facilities. Second, our 18‐month follow‐up captures short‐term trajectories and may be insufficient to detect longer‐latency effects of cardiometabolic conditions; our estimates should be interpreted within this time horizon. Third, differential attrition across comorbidity groups may have influenced our longitudinal findings. Fourth, MRI ratings were anchored in the neuroradiologist's clinical report with adjudication when needed; we did not conduct a formal parallel independent rating workflow that would allow estimation of inter‐rater or intra‐rater reliability. Fifth, comorbidities identified through ICD codes and routine‐care documentation may be subject to misclassification or underreporting, particularly for alcohol use, smoking, and sensory impairments; such measurement error would likely bias our estimates. Sixth, binary coding of comorbidities may attenuate associations and precludes dose–response inference within each condition. Seventh, comorbidity variables should be interpreted as baseline prevalent burden rather than time‐varying exposure. Eighth, the MMSE may show floor effects in more impaired patients and education‐related measurement constraints in our low‐education population, which may compress observed change and attenuate estimated decline. Future studies with more detailed quantification, larger samples, and more balanced representation of comorbidity clusters are needed to confirm these findings. Despite these limitations, the use of a real‐world dataset from a public memory clinic in Brazil allows a broad characterization of comorbidity patterns in patients with dementia.

In conclusion, our study shows that patients with cognitive disorders in a Brazilian tertiary memory clinic carry a high comorbidity burden. Comorbidities co‐occur and interact synergistically, with cerebrovascular disease showing the strongest association with dementia. The combined effect of clinical stroke and white matter hyperintensities on cognitive trajectories supports incorporating both clinical and imaging cerebrovascular markers into prognostic assessment. Because the study was conducted in routine care, our findings may inform risk stratification in similar services where multimorbidity is common. Future studies with longer follow‐up and finer comorbidity phenotyping are needed to further clarify these relationships.

CONFLICT OF INTEREST STATEMENT

L.P.S. has received honoraria for educational activities from Aché, Apsen, Biogen, Knight, Libbs, Lilly, Novo Nordisk, and Roche, and has served on advisory boards for Biogen, Knight, Lilly, Novo Nordisk, and Roche, outside the submitted work. The other authors declare that they have no conflicts of interest. Author disclosures are available in the supporting information.

CONSENT STATEMENT

This work was approved by the Research Ethics Committee of the Pontifical Catholic University of Rio Grande do Sul.

Supporting information

Supporting Information

Supporting Information