Associations of brain structure and neuropsychological function with artificial intelligence estimates of biologic vascular age

Abstract

Background:

Accelerated vascular aging, assessed as artificial intelligence-based vascular age (AIVA), is associated with small vessel disease that may impact brain structure and neuropsychological function.

Methods:

In a cross-section of Framingham Heart Study participants, AIVA was estimated using a validated convolutional neural network trained to predict carotid-femoral pulse wave velocity from a normalized pressure waveform. Brain structure was assessed using MRI with diffusion tensor imaging, and neuropsychological function was assessed via a standardized test battery. Analyses included MRI (N=2313) and neuropsychological (N=3001) samples. We used multivariable linear and logistic regression to relate AIVA with brain structural and neuropsychological functional measures.

Results:

The mean±SD age across participants was 62±11 years; 56% were women. In multivariable models, higher AIVA was associated with worse markers of cerebral small vessel disease (mean white matter free water: β [per SD], 0.09; 95% CI, 0.04 to 0.14; P<0.001; peak width of skeletonized mean diffusivity: β, 0.05; 95% CI, 0.00 to 0.10; P=0.045; white matter hyperintensity volume: β, 0.08; 95% CI, 0.03 to 0.12; P<0.001) and higher odds of cerebrovascular injury (presence of extensive white matter hyperintensities: OR [per SD]=1.22; 95% CI, 1.02 to 1.47; P=0.03; presence of brain infarcts: OR=1.44; 95% CI, 1.02 to 2.04; P=0.04). Additionally, higher AIVA was associated with worse performance on trails B-A (β, −0.05; 95% CI, −0.10 to −0.01; P=0.03), similarities (β, −0.06; 95% CI, −0.11 to −0.01; P=0.02), and global cognition (β, −0.07; 95% CI, −0.12 to −0.03; P=0.002) and higher odds of prevalent depressive symptoms (OR=1.24; 95% CI, 1.08 to 1.41; P=0.002) and high CES-D score (OR=1.20; 95% CI, 1.00 to 1.44; P=0.047). Vascular brain injury markers partially mediated relations of AIVA with global cognition score and presence of depressive symptoms.

Conclusion:

Peripheral pressure waveform AIVA may be a novel, noninvasive indicator of subclinical vascular brain injury and neuropsychological function.

Graphical Abstract

Introduction

As the US population ages, the estimated Alzheimer’s disease burden will increase to almost 14 million by 2060.¹ Although dementias are typically classified as vascular and Alzheimer’s disease types based on dominant pathology, vascular dysfunction is common across all forms of Alzheimer’s disease and related dementias.² Notably, community-based studies have shown associations of elevated aortic stiffness and pressure pulsatility with alterations in brain structure and Alzheimer’s disease and related dementias,^3–18 including greater amyloid burden in older adults^14,15 and higher tau burden in younger adults.¹⁷ In addition, we have shown that higher carotid-femoral pulse wave velocity (CFPWV), the reference standard for aortic stiffness, is closely associated with microvascular injury (e.g., greater white matter free water and white matter hyperintensity [WMH] volume), highlighting the susceptibility of brain white matter to excessive pulsatility.¹⁸ Importantly, white matter free water reflects an early manifestation of brain tissue damage, capturing extracellular water that is not restricted by surrounding structures and is independent of directionality.

Accumulating evidence indicates that lifestyle and pharmacological interventions can modify aortic stiffness.^19–26 Yet, clinical usage of aortic stiffness measurements is limited by a need for specialized equipment and training. Recently, we have described a novel artificial intelligence-based model that estimates aortic stiffness from an easily obtained uncalibrated peripheral pressure waveform and rescales that predicted value as a biologic vascular age (AIVA).²⁷ The AIVA model strongly predicts cardiovascular disease events in community-based samples with a broad range of chronological age. Although CFPWV has been associated with measures of brain microvascular injury, the current analysis extends the foregoing findings to a more accessible measure of premature aortic stiffening. Thus, we aimed to assess relations of AIVA with measures of brain vascular injury and neuropsychological function. We hypothesized that higher AIVA is associated with more markers of brain vascular injury and worse neuropsychological function.

Methods

Data Availability

The data that support the observations and findings of this study are available from the Framingham Heart Study. The procedures for requesting data can be found at https://www.framinghamheartstudy.org/. Please see the Major Resources Table in the Supplemental Materials.

Study Samples

The present study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.²⁸ The samples consisted of participants from the Framingham Offspring at examination 9 (2011–2014), Omni 1 at examination 4 (2011–2014), and Generation 3, New Offspring Spouses, and Omni 2 cohorts at examination 3 (2016–2019); these cohorts have been described.^29–31 Among the 6251 participants, hemodynamic assessment and subsequent AI modeling were completed in 5950 participants. We excluded participants without neuropsychological assessment or who were missing all neuropsychological variables (N=2737), participants with a history of dementia or stroke (N=101), and participants missing cardiovascular disease covariates (N=103). Therefore, 3009 participants, without stroke or dementia, had complete neuropsychological assessments and cardiovascular covariates. We further excluded participants missing all MRI assessment variables (N=611) and participants with a history of other neurological conditions (N=85) from the brain MRI sample (N=2313). For the neuropsychological test sample (N=3001), we excluded participants missing assessment on depressive symptoms (N=7) and education data (N=1). Participants were included in each sample if they had complete data for at least one relevant dependent variable measure. To maximize sample sizes, we excluded participants missing individual neuropsychological or brain MRI data on an analysis-by-analysis basis. Participants provided written informed consent, and protocols were approved by the Boston University Medical Center Institutional Review Board.

Artificial Intelligence Vascular Age Model

AIVA was trained to predict a negative inverse transformed carotid-femoral pulse wave velocity value (niCFPWV) from only the shape of a normalized radial arterial pressure waveform. We utilized radial arterial pressure waveforms for the current analyses since they can be obtained noninvasively, rapidly, and with minimal technical expertise using widely available devices. The full model architecture for the 8-layer convolutional neural network, including hyperparameters, training protocol, data pre-processing, and regression of predicted niCFPWV to an equivalent age scale for AIVA, have been described in detail.²⁷ Notably, the model was developed in the Age, Gene/Environment Susceptibility-Reykjavik Study (Iceland) and then externally validated in the independent Framingham Heart Study (United States). The AIVA model strongly predicts cardiovascular disease events in community-based samples. Importantly, although the model was trained using niCFPWV as the reference label, the resulting model-predicted AIVA was more strongly associated with incident clinical events than native CFPWV.²⁷ This reflects one key strength of machine learning, where models trained on large datasets can reduce measurement noise and may better capture the underlying physiological signal relevant to disease risk.

Neuroimaging

We performed brain MRI with diffusion tensor imaging using previously described algorithms.^32–36 We segmented WMH and total brain parenchymal volumes (both normalized to intracranial volume) from fluid-attenuated inversion recovery and T1-weighted images, respectively, using automated procedures.^37,38 Peak width of skeletonized mean diffusivity was calculated as the difference between the 95th and 5th percentiles of diffusion tensor imaging-derived mean diffusivity values within the white matter tract skeleton.^36,39 We defined the presence of extensive WMHs in cases where the 10-year age-bin specific z-score of the natural log of the ratio of WMH volume to total cranial volume was >1.⁴⁰ We identified the presence of covert brain infarcts based on their size, location, and imaging characteristics.⁴⁰

Neuropsychological Assessment

Framingham Heart Study has ongoing surveillance for dementia and cognitive impairment.^41,42 A neuropsychological battery was administered by trained research assistants and neuropsychologists to participants. A global cognitive score was derived from a principal component analysis of weighted loadings of Similarities, Visual Reproduction, Logical Memory, and Trail Making Part B tests.⁴³ The Center for Epidemiologic Studies Depression Scale (CES-D) was used to assess depressive symptoms; we natural log transformed the score to normalize its skewed distribution. Participants with a score ≥16 were classified as having a high CES-D score. Furthermore, we defined presence of depression symptoms as having a score of ≥16 on the CES-D or use of antidepressant medications.

Clinical evaluation and covariates

Medical history and physical examination were performed routinely at each research examination.²⁹ Age, sex, hypertension treatment, education level, and smoking were assessed via questionnaires. Current smoking was defined as self-reported regular use of cigarettes in the year preceding examination. Height (meters) and weight (kilograms) were assessed during examination. Body mass index was calculated as the ratio of weight and the square of height. Heart rate and mean arterial pressure were assessed during tonometry. Serum cholesterol levels were measured from a fasting blood test. Criteria for diabetes were a fasting glucose ≥126 mg/dL (7.0 mmol/L), random glucose≥ 200 mg/dL (11.1 mmol/L), or treatment with insulin or oral hypoglycemic agent. Metabolic syndrome was defined as meeting ≥3 of 5 criteria:^44,45 (1) high waist circumference (≥102 cm in men; ≥88 cm in women); (2) high fasting triglyceride (≥150 mg/dL/≥1.7mmol/L or treatment for elevated lipids); (3) high blood pressure (≥130 mm Hg systolic blood pressure, ≥85 mm Hg diastolic blood pressure, or treatment for hypertension); (4) low high-density lipoprotein (HDL) cholesterol (<40 mg/dL in men; <50 mg/dL in women); and (5) high fasting glucose (≥100 mg/dL or treatment for elevated glucose).

Statistical Analyses

We tabulated characteristics for the study samples. We used multivariable linear and logistic regression models to relate AIVA with continuous and binary brain structural and neuropsychological functional measures, respectively. We analyzed continuous independent and dependent variables as standardized z-scores (mean=0, standard deviation=1). We selected covariates a priori as follows: age, age², sex, cohort, body mass index (natural log), triglycerides (natural log), total-to-high density lipoprotein cholesterol ratio (natural log), fasting glucose (natural log), diabetes, hypertension treatment, current smoking, and time between clinic exam and MRI or neuropsychological testing. These covariates were included in the linear regression, logistic regression, interaction, and mediation models, as applicable to the corresponding outcome. For neuropsychological functional models, we further adjusted for education and depression score (except where the dependent variable was depression related). To assess potential collinearity, we calculated variance inflation factors simultaneously for AIVA and all covariates included in representative brain MRI and neuropsychological regression models. Because age and age² are mathematically related, we also calculated variance inflation factors in sensitivity models excluding age² (Table S1). In secondary analyses, we assessed presence of effect modification (interaction) by age (below vs. at/above median), sex, or presence of metabolic syndrome by incorporating corresponding interaction terms into the analyses. Interaction models included the same covariates as the corresponding primary regression models. For significant interactions, we performed stratified analyses. In addition, we used structural equation modeling to assess the potential mediating effects of brain MRI measures on the relations of AIVA with continuous and binary neuropsychological functional measures.^46,47 All analyses were performed with SAS version 9.4 for Windows (SAS Institute, Cary, NC). Two-tailed P<0.05 were considered statistically significant, except for tests of interaction, for which P<0.1 was considered significant. Bonferroni-corrected thresholds were additionally reported for transparency and contextual interpretation of multiple comparisons.

Results

A flow chart for the analytical samples is presented in Figure 1, and characteristics of the samples are presented in Table 1. Our sample was composed of relatively healthy middle-aged and older adults without a history of dementia or stroke, characterized by a low prevalence of diabetes, smoking, and hypertension treatment, along with a moderate prevalence of metabolic syndrome, as compared to the general population. We present a summary of the brain MRI and neuropsychological test variables in Table S2 in the Supplement.

Figure 1.

Flow chart of analysis sample.

Table 1.

Clinical characteristics of the sample.

Variable	Neuropsychological test sample, N=3001	Brain MRI sample, N=2313
Age, years	62±11	61±11
Women, No. (%)	1673 (56)	1312 (56)
Cohort
Offspring, No. (%)	1278 (43)	988 (43)
New Offspring Spouses, No. (%)	17 (1)	1 (<1)
Third Generation, No. (%)	1456 (49)	1135 (49)
Omni 1, No. (%)	151 (5)	132 (6)
Omni 2, No. (%)	99 (3)	57 (3)
At least some college, No. (%)	2458 (82)	1909 (83)
Body mass index, kg/m2	27.7 [24.6, 31.5]	27.5 [24.5, 31.2]
Triglycerides, mg/dL*	96 [71, 137]	95 [70, 135]
Total to high-density lipoprotein cholesterol ratio	3.1 [2.5, 3.8]	3.1 [2.5, 3.8]
Presence of diabetes, No. (%)	339 (11)	240 (10)
Hypertension treatment, No. (%)	1179 (39)	867 (38)
Current smoking, No. (%)	144 (5)	105 (5)
Fasting blood glucose, mg/dL*	97 [90, 105]	96 [90, 104]
Presence of metabolic syndrome, No. (%)	1304 (44)	966 (42)
AI-vascular age, years	62±15	61±15

Values are mean±standard deviation, No., number (%), or median [25th, 75th percentile].

^*To convert to mmol/L, multiply values by 0.0259.

Associations of continuous measures of brain MRI with AIVA are presented in Table 2. Higher AIVA was associated with higher WMH volume (β per standard deviation [SD], 0.08; 95% confidence interval [CI], 0.03 to 0.12; P<0.001), mean white matter free water (β per SD, 0.09; 95% CI, 0.04 to 0.14; P<0.001), and peak width of skeletonized mean diffusivity (β per SD, 0.05; 95% CI, 0.00 to 0.10; P=0.045). Additionally, higher AIVA was associated with higher odds of extensive WMHs (odds ratios [OR] per SD, 1.22; 95% CI, 1.02 to 1.47; P=0.033) and prevalent brain infarcts (OR per SD, 1.44; 95% CI, 1.02 to 2.04; P=0.041). We did not observe significant effect modification for associations of AIVA with brain MRI measures (Table S3 in the Supplement).

Table 2.

Relations of brain MRI measures with artificial intelligence vascular age.

Continuous brain measures (linear regression models)
Variable	n	Est. β±SE (95% CI)	P	Adjusted R2	Delta R2 due to the addition of AIVA
Total brain volume*	2312	−0.03±0.02 (−0.06 to 0.01)	0.13	0.70	0.0003
WMH volume*	2304	0.08±0.02 (0.03 to 0.12)	<0.001	0.46	0.0028
Mean white matter free water	1897	0.09±0.03 (0.04 to 0.14)	<0.001	0.42	0.0037
Peak width of skeletonized mean diffusivity	1897	0.05±0.03 (0.00 to 0.10)	0.045	0.43	0.0012
Binary brain measures (logistic regression models)
Variable	cases/n	OR (95% CI)	P	c-statistic	Delta c-statistic due to the addition of AIVA
Presence of extensive WMHs	300/2304	1.22 (1.02 to 1.47)	0.03	0.614	0.010
Presence of brain infarcts	88/2313	1.44 (1.02 to 2.04)	0.04	0.757	0.007
Presence of small (<1cm) brain infarcts*†	71/2296	1.31 (0.89 to 1.92)	0.17	0.758	0.005

Abbreviations: MRI, magnetic resonance imaging. AI, artificial intelligence. CI, confidence interval. WMH, white matter hyperintensity.

^*Volumes were normalized to intracranial volume.

^†Analyses excluded those with brain infarcts >1cm, n=17.

All estimated β±standard error (SE) represents standard deviation difference in AI-vascular age per standard deviation difference in brain MRI measures. Models are adjusted for age, age2, sex, cohort, body mass index (natural log), fasting glucose (natural log), triglycerides (natural log), total to high-density lipoprotein cholesterol ratio (natural log), diabetes, hypertension treatment, smoking status, and time interval between clinic exam and MRI. The variance inflation factors (VIFs) were calculated for AIVA and all covariates in the models and are reported in Table S1. VIF for AIVA is 2.24. VIFs are even lower for the covariates, except for age and age2 (because they are collinear). If age2 is removed, the VIF for AIVA is 2.22, the VIF for age is 2.41; and all other VIFs are <2.15. Note: Applying a Bonferroni correction would yield a significance threshold of α=0.0071.

Associations of continuous neuropsychological tests with AIVA are presented in Table 3. Higher AIVA was associated with worse performance for Trails B-A (β per SD, −0.05; 95% CI, −0.10 to −0.01; P=0.028), similarities (β per SD, −0.06; 95% CI, −0.11 to −0.01; P=0.016), and global cognition (β per SD, −0.07; 95% CI, −0.12 to −0.03; P=0.002). Additionally, higher AIVA was associated with higher odds of prevalent depressive symptoms (OR per SD, 1.24; 95% CI, 1.08 to 1.41; P=0.002) and high CES-D score (OR per SD, 1.20; 95% CI, 1.00 to 1.44; P=0.047).

Table 3.

Relations of neuropsychological tests with artificial intelligence vascular age.

Continuous neuropsychological measures (linear regression models)
Variable	n	Est. β±SE (95% CI)	P	Adjusted R2	Delta R2 due to the addition of AIVA
Trails B-A*	2814	−0.05±0.02 (−0.10 to −0.01)	0.03	0.23	0.0013
Logical memories, delayed	2958	−0.03±0.03 (−0.08 to 0.02)	0.28	0.10	0.0004
Visual reproductions, delayed	2953	−0.03±0.02 (−0.08 to 0.01)	0.17	0.25	0.0005
Similarities	2961	−0.06±0.03 (−0.11 to −0.01)	0.02	0.13	0.0017
Global cognition	2726	−0.07±0.02 (−0.12 to −0.03)	0.002	0.33	0.0025
CES-D score†	3001	0.04±0.03 (−0.02 to 0.09)	0.18	0.04	0.0006
Binary neuropsychological measures (logistic regression models)
Variable	cases/n	OR (95% CI)	P	c-statistic	Delta c-statistic due to the addition of AIVA
Presence of depressive symptoms	680/2999	1.24 (1.08,1.41)	0.002	0.676	0.004
High CES-D score	291/3001	1.20 (1.00,1.44)	0.047	0.621	0.006

Abbreviations: AI, artificial intelligence. CES-D, Center for Epidemiologic Studies Depression Scale.

^*The mathematical sign for Trail Making B-A was natural log-transformed and reversed such that higher scores indicated faster processing speed and superior executive functions.

^†CES-D score was natural log-transformed.

All estimated β±standard error (SE) represents standard deviation difference in AI-vascular age per standard deviation difference in neuropsychological test score. Models are adjusted for age, age2, sex, cohort, education, body mass index (natural log), fasting glucose (natural log), triglycerides (natural log), total to high-density lipoprotein cholesterol ratio (natural log), diabetes, hypertension treatment, smoking status, CES-D score (except where CES-D score is the dependent variable), and time interval between clinic exam and neuropsychological testing (except where CES-D score is the dependent variable). The variance inflation factors (VIFs) were calculated for AIVA and all covariates in the models and are reported in Table S1. VIF for AIVA is 2.28. VIFs are even lower for the covariates, except for age and age2 (because they are collinear). If age2 is removed, the VIF for AIVA is 2.26, the VIF for age is 2.47; and all other VIFs are <2.20. Note: Applying a Bonferroni correction would yield a significance threshold of α=0.0063.

We summarize interactions for associations of neuropsychological tests with AIVA in Table S4 in the Supplement. We observed significant effect modification by age for associations of AIVA with Trails B-A and CES-D score (Figure 2A). In older but not younger participants (at/above median vs. below median age), higher AIVA was associated with worse performance on Trails B-A tests. Conversely, in younger but not older participants, higher AIVA was associated with higher CES-D score. Additionally, we observed significant interactions for relations of AIVA with the presence of depressive symptoms by age and sex, and with high CES-D score by age, sex, and presence of metabolic syndrome (Figure 2B). In younger participants, higher AIVA was associated with higher odds of prevalent depression and high CES-D score whereas these associations were not significant in older participants. In women but not men, higher AIVA was associated with higher odds of depression and high CES-D score. Additionally, in participants without metabolic syndrome, higher AIVA was associated with higher odds of high CES-D score whereas in participants with prevalent metabolic syndrome, there was no association.

Figure 2.

Effect modification for associations of neuropsychological measures with artificial intelligence estimates of vascular age (AIVA). NP, neuropsychological. Older participants were age≥62 years; younger participants were age<62 years. (A) Effect sizes (βs) and 95% confidence intervals from linear regression models that assessed relations of neuropsychological tests (continuous measures) with AIVA. All estimated βs represent standard deviation difference in AIVA per standard deviation difference in neuropsychological test score. βs <0 favor worse cognition (Trails B-A) or fewer depressive symptoms (CES-D score), whereas βs >0 favor better cognition (Trails B-A) or more depressive symptoms (CES-D score). Models are adjusted for age, age², sex, cohort, education, body mass index (natural log), fasting glucose (natural log), triglycerides (natural log), total to high-density lipoprotein cholesterol ratio (natural log), diabetes, hypertension treatment, smoking status, the CES-D score (except where CES-D score is the dependent variable), and time interval between clinic exam and neuropsychological (except where CES-D score is the dependent variable). (B) Odds ratios (OR) and 95% confidence intervals from logistic regression models that assessed relations of presence of depression with AIVA. High CES-D score was defined as participants with a CES-D score ≥16 who were classified as having depressive symptoms. MetS-, metabolic syndrome absent; MetS+, metabolic syndrome present. MetS+ was defined as meeting ≥3 of 5 criteria: (1) high waist circumference (≥102 cm in men; ≥88 cm in women); (2) high fasting triglyceride (≥150 mg/dL or treatment for elevated lipids); (3) high blood pressure (≥130 mm Hg systolic blood pressure, ≥85 mm Hg diastolic blood pressure, or treatment for hypertension); (4) low high-density lipoprotein (HDL) cholesterol (<40 mg/dL in men; <50 mg/dL in women); and (5) high fasting glucose (≥100 mg/dL or treatment for elevated glucose). ORs <1 favor absence of depression, whereas ORs > 1 favor presence of depression. All estimated ORs represent per standard deviation difference in AIVA. Models are adjusted for age, age², sex (except where model is stratified by sex), cohort, education, body mass index (natural log), fasting glucose (natural log), triglycerides (natural log), total to high-density lipoprotein cholesterol ratio (natural log), diabetes, hypertension treatment, and smoking status.

We summarize the mediation analysis of brain MRI measures on relations of AIVA with presence of depressive symptoms and global cognition score in Table S5 in the Supplement. Mean white matter free water and peak width of skeletonized mean diffusivity were significant partial mediators of the relations of higher AIVA with worse global cognition score (Figure 3). In addition, WMH volume, mean white matter free water, and peak width of skeletonized mean diffusivity were significant partial mediators of the relations of higher AIVA with higher odds of depressive symptoms (Figure 4).

Figure 3.

Conceptual models and pathway analyses for the effect of artificial intelligence-based vascular age (AIVA) with global cognition score. Candidate mediators: (A) mean white matter free water and (B) peak width of skeletonized mean diffusivity. Models adjusted for age, age², sex, cohort, body mass index, triglycerides, total-to-high density lipoprotein cholesterol ratio, fasting glucose, diabetes, hypertension treatment, current smoking, education, time between clinic exam and MRI, time between clinic exam and neuropsychological testing, and Center for Epidemiologic Studies Depression (CES-D) score. Parameter estimates were computed using maximum likelihood estimation. Bias-corrected 95% confidence intervals (CI) for direct and indirect effects were calculated using 5000 bootstrap samples within the structural equation modeling framework.

Figure 4.

Conceptual models and pathway analyses for the effect of artificial intelligence-based vascular age (AIVA) with presence of depressive symptoms. Candidate mediators: (A) white matter hyperintensity volume; (B) mean white matter free water; and (C) peak width of skeletonized mean diffusivity. Models adjusted for age, age², sex, cohort, body mass index, triglycerides, total-to-high density lipoprotein cholesterol ratio, fasting glucose, diabetes, hypertension treatment, current smoking, education, and time between clinic exam and MRI. Parameter estimates were computed using maximum likelihood estimation. Bias-corrected 95% confidence intervals (CI) for direct and indirect effects were calculated using 5000 bootstrap samples within the structural equation modeling framework.

Discussion

In middle-aged and older adults without a history of dementia or stroke, higher AIVA was associated with higher markers of cerebral small vessel disease, worse performance on neuropsychological tests, and higher odds of depression. Associations of neuropsychological outcomes with AIVA were modified by age, sex, and presence of metabolic syndrome. In addition, higher levels of subclinical vascular brain injury mediated the relations of higher AIVA with lower global cognition score and higher odds of depressive symptoms. Our data indicate that convolutional neural network-derived AIVA may be a novel, noninvasive indicator of subclinical cerebral small vessel injury and neuropsychological function that may have utility in early detection and prevention of vascular brain injury and cognitive decline.

AIVA and Brain Structural Integrity

We observed that higher AIVA is associated with increased mean white matter free water, indicative of cerebral edema, neuroinflammation, or breakdown of the blood-brain barrier. This observation is consistent with our prior work using CFPWV (a native measure of vascular aging), which shows similar associations with higher free water in a FHS sample.¹⁸ Additionally, higher AIVA was associated with MRI markers/imaging endophenotypes of small vessel disease such as peak width of skeletonized mean diffusivity (reflecting white matter microstructural damage) and WMH (reflecting macrostructural white matter injury). These MRI markers have consistently been associated with cognitive decline, and conditions such as vascular dementia and mixed dementia. Given the established associations between CFPWV and the risk of downstream brain injury,^{3,6,7,9–13,48} our study suggests that this novel, noninvasive method to detect and monitor aortic stiffness and biological vascular aging could potentially identify individuals at risk for cerebral small vessel disease who might benefit from earlier interventions. Further longitudinal studies are needed to test this hypothesis.

AIVA and Neuropsychological Measures

We observed that higher AIVA was associated with worse cognition. These data are consistent with prior work showing associations of higher age-related aortic stiffness and greater cardiac dysfunction with cognitive decline.^{3–5,8,16,49–51} Specifically, AIVA was associated with worse executive function and processing speed (lower Trails B-A score), worse abstract reasoning, verbal concept formation, and semantic knowledge (lower Similarities score), and lower overall cognition (lower global cognition score). Elevated aortic stiffness increases hemodynamic pulsatile load and reduces cerebrovascular reactivity and cerebral perfusion, particularly in deep cerebral small vessels arising directly from the circle of Willis.⁴⁸ As we observed, older adults may be particularly vulnerable, possibly due to their stiffer aortas, higher pressure pulsatility, and increased transmission of excessive pulsatile power into the brain microcirculation. The prefrontal and temporal lobes are particularly vulnerable to potential pulsatile damage or hypoperfusion caused by aortic stiffening and cardiac dysfunction.⁵² Chronic damage in these areas may reduce verbal and abstract reasoning abilities since they rely on tight neurovascular coupling to meet high metabolic demand.⁵³ Microvascular function underlies the well-known phenomenon of neurovascular coupling, wherein small resistance vessels dynamically regulate local cerebral blood flow in response to neuronal activity and metabolic demand. Impairment in this process leads to inadequate delivery of oxygen and nutrients to active brain regions. Consistent with this framework, impaired microvascular function has been linked to worse brain function in older individuals.⁵⁴ Elevated aortic stiffness may affect neuropsychological performance by contributing to reduced cerebral blood flow and flow reactivity, which may contribute to impaired neurovascular coupling and cognitive decline. Our mediation analysis provides preliminary support for this mechanistic framework. Mean white matter free water, an MRI marker of extracellular water accumulation and white matter microstructural integrity, and peak width of skeletonized mean diffusivity, an MRI marker of white matter microstructural injury, each partially mediated the association of higher AIVA with lower global cognition scores. Mean white matter free water accounted for 11.4% of the observed association, whereas peak width of skeletonized mean diffusivity accounted for 8.4%. These observations suggest that cerebral microvascular injury and related white matter microstructural injury may represent pathways linking accelerated vascular aging with cognitive dysfunction. However, subclinical aortic stiffening may have significant impact prior to the onset of cognitive impairment.⁵⁴ Therefore, early detection is crucial for effective intervention and prevention.

Higher AIVA was associated with higher odds of depression, particularly among women and younger participants. These observations are consistent with previous studies reporting significant associations between aortic stiffness and depression, as aortic stiffness is linked to subclinical markers of brain injury, including cerebral hypoperfusion, vascular brain injury, systemic inflammation, and neurohormonal dysregulation, which may contribute to depressive symptoms.^55–57 These findings are also consistent with the “microvascular depression” hypothesis, which posits that cerebrovascular small vessel disease may predispose, precipitate, or perpetuate depressive syndromes, particularly in later life.^58–60 Aortic stiffening reduces cerebral perfusion and flow reactivity, particularly in brain regions sensitive to ischemia.^61,62 Within this framework, microvascular injury and cerebral small vessel disease have been proposed as key contributors to depression risk. Consistent with this hypothesis, we observed that WMH volume, mean white matter free water, and peak width of skeletonized mean diffusivity each significantly and partially mediated the association of higher AIVA with greater odds of depressive symptoms, accounting for 6.0 to 15.7% of the observed association. These findings support the possibility that microvascular brain injury and related white matter microstructural injury represent pathways through which vascular aging contributes to depression risk. Furthermore, recent Framingham Heart Study observations from Araujo-Contreras et al. suggest that greater burden of cerebral small vessel disease and presence of visible perivascular spaces are associated with higher risk of incident depression.⁶³ Regarding differences by sex, women, particularly premenopausal women or those on hormone replacement therapy, have estrogen levels that can influence vascular stiffness and cerebral perfusion.^64–66 Beyond sex-related differences, age may further modify these associations. One possible explanation for the stronger association observed in younger participants is that vascular risk may be more readily detectable against a lower background burden of age-related multimorbidity and competing pathophysiology, whereas in older adults the contribution of any single vascular marker may be attenuated. This is consistent with evidence that the relative impact of individual cardiovascular risk factors diminishes with advancing age,^67,68 while older populations experience greater competing risks, including non-disease-specific mortality, which can obscure or bias associations with incident outcomes.^69,70 Moreover, age-dependent analyses demonstrate that vascular risk factors are more strongly associated with depressive symptoms in midlife compared with later life, further supporting this interpretation.⁷¹

Importantly, the relation of vascular dysfunction with depression is likely bidirectional and multifactorial. We cannot overlook the possibility of reverse causality since depression may exacerbate arterial dysfunction by encouraging unhealthy behaviors. For example, individuals with depression may be less physically active, have poorer diets, smoke, or adhere less consistently to prescribed medications, all of which can worsen vascular health. Additionally, depression and aortic stiffness may reinforce each other through autonomic dysregulation and inflammatory mechanisms, perpetuating a vicious cycle.^64,72 Regardless, depression poses a significant clinical burden and is linked to an increased risk of cognitive decline and cardiovascular disease, particularly later in life. Further research is needed to understand the pathobiology of depression and its links to vascular dysfunction.

Novelty and Clinical Relevance

Vascular dysfunction is increasingly recognized as a contributor to Alzheimer’s disease and related dementias. The AIVA model provides a novel, accessible approach to assessing aortic stiffness that may be more practical and robust than directly measured CFPWV, which requires specialized equipment and training. In contrast, AIVA can be derived from a brief (~30-second) noninvasive arterial waveform acquisition, such as radial tonometry, requiring minimal operator training and no complex calibration procedures, thereby enhancing feasibility in routine clinical settings. This approach may also extend to photoplethysmography waveforms obtained from widely available wearable or handheld devices. AIVA may enable assessment of risk of cerebral vascular injury and cognitive dysfunction associated with elevated aortic stiffness at the point-of-care. If confirmed in prospective cohorts, AIVA may facilitate early detection of risk and targeted prevention strategies for cerebral small vessel disease and cognitive decline. We acknowledge that the potential clinical utility of AIVA may be greatest in individuals who would otherwise be considered at low or intermediate risk, particularly younger or middle-aged adults in whom subclinical vascular aging is not routinely assessed. In contrast to high-risk individuals, where aggressive risk factor modification is already indicated, or older adults, in whom vascular damage may be more advanced and less reversible, AIVA may help identify a subset of individuals with disproportionate aortic stiffening who could benefit from earlier and potentially more impactful lifestyle intervention.

Importantly, aortic stiffness is modifiable. Identification of a disproportionately high AIVA may inform more aggressive management of vascular risk, including optimization of blood pressure, lipid levels, physical activity, and diet. In this context, AIVA could serve as a clinically actionable marker to guide early intervention, even in individuals without overt disease. The potential integration of AIVA into wearable devices presents a compelling opportunity to advance the early detection and prevention of preclinical vascular brain injury and cognitive decline. The AIVA model, which relies solely on a standardized arterial pulse waveform without requiring pressure or time calibration, can be incorporated into wearable technology to facilitate self-assessment and long-term monitoring. This capability could facilitate early identification of individuals at risk for cerebral small vessel disease, accounting for 25% of ischemic strokes and representing the predominant vascular contributor to cognitive decline and dementia.⁷³ In addition, integration into wearable platforms may provide real-time feedback on the effects of lifestyle modification or pharmacologic treatment, potentially reinforcing adherence and enabling individualized risk reduction strategies. By tracking vascular health trends, AIVA potentially might inform and motivate lifestyle modifications that may mitigate the progression of vascular brain injury and promote cognitive resilience. Early identification of risk offers a window for preventive interventions to reduce the risk of depression and maintain cognitive health. Thus, AIVA could potentially serve as a powerful tool to detect preclinical markers of vascular brain injury and cognition, paving the way for early, targeted prevention strategies within a precision medicine framework.

Limitations

Our study has limitations to consider. Due to the cross-sectional design, further research is needed to determine prospective associations, potential mediation, and causality. Although we adjusted for confounders, residual confounding from unmeasured factors remains possible. The observed effect sizes were modest, suggesting that AIVA should be interpreted as a complementary marker of vascular brain injury risk rather than a stand-alone predictor. We did not account for multiple testing; therefore, our results are more susceptible to type-1 error. Notably, some observed associations remained above Bonferroni-corrected significance thresholds; therefore, our results should be considered hypothesis generating and replicated in additional cohorts. The AIVA model is complex, with over 6 million trainable parameters, which renders models susceptible to overtraining. However, during training, we used aggressive dropout, real-time evaluation in holdout validation and test sets and early stopping to prevent overfitting. Relations of AIVA values with niCFPWV were tested on holdout samples in the training set. Importantly, the model was trained using data from a pair of Icelandic cohorts and was then validated in the independent Framingham cohort to ensure robustness.²⁷ The findings of our study may not be generalizable to other ethnic or racial groups because our samples were composed mostly of white participants of Western European descent.

Conclusion

In middle-aged and older community-dwelling Framingham participants without a history of dementia or stroke, higher AIVA was associated with markers of cerebral small vessel disease, poorer performance on neuropsychological tests, and greater likelihood of depression. Some of these associations were modified by age, sex, and the presence of metabolic syndrome. Furthermore, relations of higher AIVA with lower global cognition score and higher odds of depressive symptoms were mediated by measures of subclinical vascular brain injury. Thus, this novel AI-derived metric of vascular function may provide insights into subclinical cerebral small vessel injury and its downstream effects on neuropsychological health.

Supplementary Material

• Online supplemental material file: Tables S1 – S5, Major Resources Table

What are the Clinical Implications?

Artificial intelligence-based vascular age (AIVA), derived from a brief, noninvasive peripheral arterial pressure waveform, was associated with MRI markers of cerebral small vessel disease, worse neuropsychological performance, and greater odds of depressive symptoms in adults without prior stroke or dementia. Measures of subclinical vascular brain injury partially mediated associations of higher AIVA with lower global cognition and depressive symptoms, supporting the possibility that accelerated vascular aging contributes to neuropsychological dysfunction through cerebral microvascular injury. Because AIVA can be estimated without the specialized equipment and training required for direct measurement of aortic stiffness, it may ultimately provide a scalable approach for identifying individuals with premature vascular aging who may be at increased risk for subclinical brain injury. Prospective validation is needed before clinical use. If confirmed, AIVA may complement established risk assessment and help target earlier lifestyle or pharmacological management of modifiable vascular risk factors to preserve brain and cognitive health.

Nonstandard Abbreviations and Acronyms

AIVA

artificial intelligence-based vascular age

CES-D

Center for Epidemiologic Studies Depression Scale

CFPWV

carotid-femoral pulse wave velocity

niCFPWV

negative inverse transformed carotid-femoral pulse wave velocity

WMH

white matter hyperintensity