Relations of Artificial Intelligence Vascular Age With Cardiometabolic Disease Progression: The Framingham Heart Study Authors: David J Hamel‐Sellman, Brenton R Prescott, Timothy J Korzinski, Vanessa Xanthakis, Leroy L Cooper, Naomi M Hamburg, Connie W Tsao, Emelia J Benjamin, Ramachandran S Vasan, Gary F Mitchell Published: 2025 ABSTRACT AbstractBackgroundAortic stiffening increases pulsatile power transmission into the microvasculature, resulting in blunted microvascular reactivity, which can impair postprandial nutrient processing and glucose homeostasis. Cardiometabolic disease (CMD) contributes to aortic stiffening and risk for major cardiovascular disease events. Our goal was to assess potential bidirectional relations of aortic stiffness with CMD progression.MethodsRelations of artificial intelligence vascular age (AI‐VA), a measure of aortic stiffness, with new‐onset hypertension, diabetes, obesity, and lipid abnormalities were assessed in 7188 Framingham Heart Study (FHS) participants by using Cox proportional hazards regression. Short‐term bidirectional relations of AI‐VA with key continuous cardiometabolic measures were assessed by using autoregressive cross‐lagged panel models at 2 visits 6 years apart. Progression to CMD and cardiovascular disease states was visualized using multistate transition analysis.ResultsIn models adjusted for standard vascular risk factors, accelerated AI‐VA was associated with elevated risk for incident hypertension (hazard ratio [HR], 1.44 [95% CI, 1.24–1.61]), diabetes (HR 1.29 [95% CI, 1.07–1.56]), lipid abnormalities (HR 1.13 [95% CI, 1.00–1.28]), and obesity (HR 1.23 [95% CI, 1.06–1.42]). Autoregressive cross‐lagged panel models demonstrated bidirectional relations between AI‐VA and fasting blood glucose, triglycerides, and mean arterial pressure consistent with a self‐reinforcing cycle of CMD progression. Multistate transition analysis demonstrated that most cardiovascular disease events (70%) occurred in participants who had developed at least 2 CMD states.ConclusionsIn our community‐based sample with repeated measures, AI‐VA, a measure of aortic stiffness, had bidirectional relations with continuous cardiometabolic measures and incident CMD, highlighting the potential utility of AI‐VA as a novel screening tool for CMD risk.Keywords: aortic stiffness, artificial intelligence, cardiometabolic risk factors, carotid‐femoral pulse wave velocitySubject Categories: Cardiovascular Disease, Epidemiology Nonstandard Abbreviations and Acronyms artificial intelligence vascular age autoregressive cross‐lagged panel model carotid‐femoral pulse wave velocity cardiometabolic disease Research Perspective What Is New? In a community‐based cohort, a deep learning‐based measure of biologic vascular age (artificial intelligence vascular age) is bidirectionally related to progression of fasting blood glucose, triglycerides, and mean arterial pressure, leading to development of a self‐propagating cardiometabolic disease cycle. In a community‐based cohort, a deep learning‐based measure of biologic vascular age (artificial intelligence vascular age) is bidirectionally related to progression of fasting blood glucose, triglycerides, and mean arterial pressure, leading to development of a self‐propagating cardiometabolic disease cycle. Accelerated artificial intelligence vascular age is associated with increased risk for development of hypertension, diabetes, lipid abnormalities, and obesity in a cohort where 70% of major cardiovascular disease events occurred in participants who had progressed to concurrence of 2 or more of these cardiometabolic disease states. Accelerated artificial intelligence vascular age is associated with increased risk for development of hypertension, diabetes, lipid abnormalities, and obesity in a cohort where 70% of major cardiovascular disease events occurred in participants who had progressed to concurrence of 2 or more of these cardiometabolic disease states. What Question Should Be Addressed Next? Further research will be required to determine if artificial intelligence vascular age can be deployed for use as a widely accessible indicator of cardiometabolic and cardiovascular disease risk that can be used to identify and track the effects of interventions that mitigate or reverse accelerated vascular aging. Further research will be required to determine if artificial intelligence vascular age can be deployed for use as a widely accessible indicator of cardiometabolic and cardiovascular disease risk that can be used to identify and track the effects of interventions that mitigate or reverse accelerated vascular aging. Cardiometabolic disease (CMD) prevalence has risen dramatically over the past 4 decades during the obesity epidemic. 1 In light of known associations between CMD and cardiovascular disease (CVD), the increase in CMD represents an urgent public health priority. 2 , 3 High levels of systolic blood pressure, fasting blood glucose, body mass index (BMI), and low‐density lipoprotein cholesterol are leading risk factors by global burden at all levels of social deprivation, and their respective burdens have accelerated since 2000. 4 Modern primary CVD prevention risk equations, such as the Predicting Risk of CVD Events (PREVENT) equations developed by the American Heart Association, incorporate measures of blood pressure, blood glucose, BMI, cholesterol, and glomerular filtration rate in response to the growing burden of CMD. 5 The prevailing paradigm for the pathogenesis of CMD and related adverse CVD outcomes posits that elements of CMD, including hypertension, diabetes, lipid abnormalities, and obesity, develop because of various genetic and lifestyle factors and exert an adverse effect on aortic structure and function. The subsequent increase in aortic stiffness contributes to risk for major adverse CVD events. 6 Recent longitudinal studies support the hypothesis that aortic stiffness may contribute to the pathogenesis of various components of CMD, including hypertension, 7 insulin resistance and lipid abnormalities, 8 obesity, 9 and metabolic‐associated steatohepatitis. 10 The foregoing observations represent a major shift from the traditional paradigm wherein the various cardiometabolic conditions were thought to damage and stiffen the aorta unidirectionally. 11 Aortic stiffening increases transmission of pulsatile power into the periphery and damages the microcirculation. 12 , 13 Resulting alterations in microvascular function contribute to associations between aortic stiffness and CVD events. 14 Associations between aortic stiffness and incident CMD may therefore reflect microvascular target organ damage in the pancreas, skeletal muscle, adipose tissue, liver, kidneys, and other vital organs, which contributes to progressive cardiometabolic dysfunction due to the central role that microvascular reactivity plays in physiologic responses in tissues throughout the body. 13 We have reported that artificial intelligence vascular age (AI‐VA), a measure of aortic stiffness derived using a convolutional neural network model trained to predict carotid‐femoral pulse wave velocity (CFPWV) from a normalized radial pressure waveform, is a powerful predictor of new‐onset CVD. 15 CFPWV is an accurate method for assessing vascular stiffness and higher values are associated with microvascular dysfunction. 16 However, clinical utility of CFPWV is limited by the requisite specialized equipment and operator training. 17 Furthermore, relations between AI‐VA and CVD events in a nearly identical sample were stronger than relations between measured CFPWV, the training label, and CVD events. 15 We hypothesized that clinical performance of the prediction often exceeds the label because the AI model learns to identify and ignore noise in waveforms and labels during training. In contrast, noise in the exposure variable dilutes the relation between the exposure variable and outcomes. 18 AI‐VA is readily assessed from a 20‐second, uncalibrated pressure waveform, suggesting that AI‐VA may prove particularly valuable as an indication of aortic stiffness in clinical settings, particularly in low‐resource settings. We hypothesize that higher values for AI‐VA are associated with higher risk of CMD progression, including incident hypertension, diabetes, lipid abnormalities, and obesity, and that CMD progression would be associated with accelerated progression of AI‐VA. We tested these hypotheses in a sample of FHS (Framingham Heart Study) participants who were free of CVD at the baseline visit and had assessments of AI‐VA and CMD status at baseline and up to 2 additional visits spanning approximately 14 years. METHODS The procedure for requesting data from the FHS can be found at https://www.framinghamheartstudy.org/. All protocols were approved by Boston University Medical Campus Institutional Review Board, and all participants provided written informed consent. Study Sample Our study sample (N=7188; median age, 48 years; 55% women) was drawn from Framingham Offspring Study participants who attended the eighth examination cycle (2005–2008, N=2504, mean age 66±8.9 years); Omni‐1 participants who attended the third examination cycle (2007–2008, N=251, mean age 62±8.6 years); and Third Generation, Omni‐2, and New Offspring Spouse participants who attended the first examination cycle (2002–2005, N=4433, mean age 41±10.0 years). Exclusion details for the modeling samples are shown in Figure S1. The design and inclusion criteria of the Framingham Offspring Study, Third Generation, New Offspring Spouse, and racially and ethnically diverse Omni‐1 and Omni‐2 cohorts have been described. 19 , 20 , 21 The FHS tonometry data acquisition methods and training sample for AI‐VA have been described. 15 For Offspring Study participants, Exam 9 (2011–2014) and Exam 10 (2019–2022) were designated as follow‐up exams. For Third Generation participants, Exam 2 (2008–2011) and Exam 3 (2016–2019) were designated as follow‐up exams. For New Offspring Spouse participants, Exam 2 (2008–2011) and Exam 3 (2016–2019) were designated as follow‐up exams. For Omni‐1 participants, Exam 4 (2011–2014) and Exam 5 (2019–2022) were designated as follow‐up exams. For Omni‐2 participants, Exam 2 (2009–2011) and Exam 3 (2016–2019) were designated as follow‐up exams. CMD Condition Definitions Obesity was defined as BMI ≥30 kg/m2. Lipid abnormality was defined as triglycerides ≥150 mg/dL, high‐density lipoprotein cholesterol <50 mg/dL in women or <40 mg/dL in men, or treatment for lipid abnormality. Hypertension was defined as systolic blood pressure ≥140 mm Hg, diastolic blood pressure ≥90 mm Hg, or treatment for hypertension. Diabetes was defined as fasting blood glucose ≥126 mg/dL, random blood glucose ≥200 mg/dL, or treatment for diabetes. Follow‐up data for new‐onset CMD conditions were collected at subsequent examination visits through December 31, 2021. Median hypertension follow‐up interval was 13.6 (25th, 75th percentiles: 6.9–14.4) years. Median diabetes follow‐up interval was 13.7 (7.1–14.4) years. Median lipid abnormality follow‐up interval was 13.5 (6.3–14.3) years. Median obesity follow‐up interval was 13.6 (6.4–14.4) years. Clinical Outcomes Major CVD events were defined as fatal or nonfatal myocardial infarction, unstable angina (prolonged ischemic episode with documented reversible ST segment changes), HF, and ischemic or hemorrhagic stroke that occurred after the baseline examination. Medical records were obtained for all hospitalizations and physician visits related to CVD during follow‐up and were reviewed by a committee of 3 investigators; events were adjudicated following written guidelines. Criteria for these cardiovascular events have been described previously. 22 , 23 Median CVD follow‐up interval was 16.7 (14.8–18.3) years with follow up extending through to December 31, 2021. AI‐Vascular Age Model AI‐VA is a convolutional neural network‐based measure of aortic stiffness trained using tonometry data acquired in 2 Icelandic cohorts (REFINE‐Reykjavik 24 [Risk Evaluation for Infarct Estimates‐Reykjavik] and AGES‐Reykjavik [Age, Gene, Environment, Susceptibility Study‐Reykjavik], 25 N=10 452 participants, N=31 126 waveforms) and validated using tonometry data acquired in FHS (N=7208 participants, N=21 624 waveforms). AI‐VA was trained to predict negative inverse CFPWV from only the shape of normalized arterial pressure waveforms acquired from the carotid, brachial, and radial arteries. Full details of the 8‐layer convolutional neural network model, including hyperparameters, training protocol, data preprocessing, and regression to equivalent age scale have been described in detail. 15 Predictions in the independent FHS validation sample were correlated with measured CFPWV (R2=0.60) and calendar age (R2=0.72) and associated with increased risk for new‐onset CVD events. Statistical Analysis We used Cox proportional hazards regression with follow‐up spanning 2 subsequent exam visits (“visit 2” and “visit 3”) after the baseline arterial tonometry was measured (“visit 1”). We examined relations of baseline AI‐VA with time to CMD progression (incident hypertension, diabetes, lipid abnormalities, and obesity, each considered individually) after confirming the assumption of proportionality was met based on visual inspection of Kaplan–Meier plots. The assumption of proportionality was further tested by visual inspection of Schoenfeld residual plots, which were reasonably flat throughout the follow‐up period. Covariates in Cox models for incident CMD were selected based on components of the Framingham risk score 26 and potential confounders of CMD risk and arterial stiffness. Covariates were measured at the baseline examination for each cohort. Three covariate models were considered. Model 1 adjusted for age, sex, and cohort to represent minimally adjusted hazard ratio estimates. Model 2 added Framingham risk score variables, including seated systolic blood pressure assessed in the research center, use of antihypertensive therapy, total cholesterol/high‐density lipoprotein cholesterol (TC/HDL) ratio, regular use of cigarettes in the prior year, and presence of diabetes, and CMD confounders, including fasting blood glucose and triglyceride concentrations. Model 2 variables were selected in part due to previously reported strong relations with aortic stiffness and stiffness progression. 27 , 28 The maximally adjusted Model 3 adds BMI, a major confounder of CMD risk, 29 , 30 and potential confounders of aortic stiffness, including supine mean arterial pressure (MAP) and heart rate assessed during tonometry data acquisition. 31 TC/HDL ratio, triglycerides, fasting blood glucose, and BMI distributions were skewed, so loge‐transformed variables were used in all analyses. CFPWV was negative inverse transformed (niCFPWV=−1000/CFPWV) to normalize the skewed and heteroskedastic distribution, restore directionality, and match the transformation applied during development and validation of the AI‐VA model. 7 , 15 To further estimate the short‐term bidirectional temporal longitudinal relations between aortic stiffness and continuous measures of CMD, we used an autoregressive cross‐lagged panel model (ARCLPM) approach. The ARCLPM method is useful to investigate temporal relations of pairs of potentially interrelated variables while adjusting for autoregression. TC/HDL ratio, triglycerides, MAP, pulse pressure, BMI, fasting blood glucose, and number of CMD conditions were selected as continuous markers of CMD at visits 1 and 2. Panel models were adjusted for age, sex, cohort, height, lipid treatment, hypertension treatment, diabetes status, smoking status, and heart rate from tonometry at visit 1, as well as time between visit 1 and visit 2. Cross‐lagged pathways were adjusted for the independent variable at baseline to accurately estimate the effect on progression. Variables for ARCLPM were selected to represent continuous measures of the CMD conditions investigated in the CMD progression Cox models. Covariates for panel models were selected based on evidence of confounding effects on CVD and CMD. Missing values for variables assessed at visits the participants attended were imputed via 5 hierarchical multiple imputations of repeated measures using the mice R package version 3.18. The most frequent missing covariate was triglyceride concentration (2.4%). Multiple imputation grouping variable was participant, and predictors included age, sex, cohort, time in years between visit 1 and visit 2, diabetes status, smoking status, hypertension status, obesity status, lipid abnormality status, hypertension treatment, lipid abnormality treatment, height, weight, waist circumference, heart rate, fasting blood glucose, triglycerides, TC/HDL ratio, BMI, pulse pressure, MAP, and AI‐VA. In primary analyses, participants who did not attend visit 2 (N=1406) did not have their visit 2 data imputed. Instead, full information maximum likelihood mixed modeling was used to account for missing visits. In secondary analyses, measured CFPWV values were missing for 340 participants who attended visit 1 (4.7%) and 170 participants who attended visit 2 (2.9%). These missing values were additionally imputed to fit ARCLPM models on the full sample (N=7188) with niCFPWV as the aortic stiffness measure instead of AI‐VA. For all analyses, ARCLPM model estimates were pooled across multiply imputed data sets by using the lavaan.mi R package version 0.1‐0. To determine whether the relations between AI‐VA and CMD events or AI‐VA and continuous CMD progression differed in older versus younger participants (dichotomized at the median age), in men versus women, or in participants with or without obesity, we included statistical interaction terms for these variables in separate models that also adjusted for standard vascular risk factors. We performed 2 sensitivity analyses in the ARCLPM analyses to assess confounding effects of hypertension treatment and diabetes. Using a binary treatment indicator covariate to adjust models for the effects of hypertension treatment may be suboptimal. Untreated pressure values are conventionally imputed for patients treated for hypertension by adding 10 and 5 mm Hg to systolic and diastolic pressures, respectively. 32 To assess potential confounding in our cross‐lagged panel models introduced by adjusting for hypertension treatment status, we performed a sensitivity analysis wherein pulse pressure and MAP values were adjusted by adding 5 and 7 mm Hg, respectively, for participants treated for hypertension at the visit in question, which is equivalent to adding 10 mm Hg to systolic blood pressure and 5 mm Hg to diastolic blood pressure while assuming a pressure waveform shape factor of 0.40. 33 To assess potential confounding in fasting blood glucose modeling introduced by including participants with diabetes, we performed a sensitivity analysis excluding all participants with diabetes at visit 1 or visit 2. Unless otherwise noted, all analyses were performed with SAS version 9.4, R version 4.3 or SPSS version 28. All panel modeling was performed using the lavaan R package version 0.6‐19. A 2‐sided P<0.05 was considered statistically significant. We present standardized hazard ratios and regression coefficients to express the difference in a dependent variable or outcome associated with a commonly observed increment (1 SD) of the exposure variable within the sample. Data Availability The procedure for requesting data from the FHS can be found at https://www.framinghamheartstudy.org/. Specimens, study data sets, and other resources can be requested at https://biolincc.nhlbi.nih.gov/studies/fhs/. RESULTS Participants Our study sample was drawn from Offspring participants who attended the eighth examination cycle (2005–2008, N=2504); Omni‐1 participants who attended the third examination cycle (2007–2008, N=251); and Third Generation, Omni‐2, and New Offspring Spouse participants who attended the first examination cycle (2002–2005, N=4433). Baseline characteristics of the sample are summarized in Table 1. We included all participants with valid radial tonometry and covariates and without prevalent CVD at baseline (N=7188; median age, 48 years; 55% women). The sample was middle aged, on average, but with a wide age range (19–90 years). The prevalence of hypertension, lipid abnormalities, and obesity were moderate and diabetes was relatively modest (Table 1). Baseline tonometry measures are summarized in Table 1. Mean radial AI‐VA prediction was 50 (SD 18) years. Supine systolic and pulse pressure were higher and diastolic pressure was lower than seated research center blood pressure (Table 1), which is consistent with the effects of posture. Of the 7188 participants, 5782 participants attended visit 2. In general, participants who attended both visits, as compared with those who attended only visit 1, were modestly younger with a lower cardiometabolic risk factor burden (Table S1). Visits 1 and 2 were separated by an average of 6.0 (median, 6.0 [IQR, 5.6–6.3]) years. Table 1Participant Characteristics at BaselineVariableValueTotal N (% women)7188 (55)Age, y50±15Height, m1.69±0.10Weight, kg78±18Body mass index, kg/m2 26.6 [23.5, 30.2]Seated systolic blood pressure, mm Hg121±16Seated diastolic blood pressure, mm Hg75±10Supine systolic blood pressure, mm Hg129±19Supine diastolic blood pressure, mm Hg68±9Supine mean arterial pressure, mm Hg94±12Supine pulse pressure, mm Hg60±17Supine heart rate, beats/min62±10Supine radial artificial intellgience vascular age, y50±18Total cholesterol, mg/dL189±36HDL cholesterol, mg/dL56±17Total cholesterol/HDL ratio3.4 [2.8, 4.3]Triglycerides, mg/dL96 [67, 138]Fasting blood glucose, mg/dL96 [89, 103]Prevalent hypertension, N (%)2297 (32)Prevalent diabetes, N (%)500 (7)Prevalent obesity, N (%)1866 (26)Prevalent lipid abnormality, N (%)3266 (45)Lipid‐lowering medications, N (%)1402 (20)Hypertension medications, N (%)1664 (23)Diabetes medications, N (%)328 (5)Current smoking, N (%)859 (12)Values represent mean±SD, 50th [25th, 75th] percentile, or N (%).HDL indicates high‐density lipoprotein cholesterol. Participant Characteristics at Baseline Values represent mean±SD, 50th [25th, 75th] percentile, or N (%). HDL indicates high‐density lipoprotein cholesterol. Approximately 11% to 14% of eligible participants without prevalent CVD at baseline were missing CMD follow‐up data, largely because they did not attend either of the 2 follow‐up exams. Among participants with CMD follow‐up data, median CMD follow‐up was 13.7 (25th, 75th percentiles: 7.7–14.5) years. We observed 1155 participants with new‐onset hypertension (27%; 13.5 per 1000 person‐years), 414 with new‐onset diabetes (7%; 4.8 per 1000 person‐years), 1126 with new‐onset lipid abnormality (32%; 13.8 per 1000 person‐years), and 739 participants with new‐onset obesity (16%; 8.8 per 1000 person‐years). We observed new‐onset hypertension at a median age of 60 years, new‐onset diabetes at a median age of 62 years, new‐onset lipid abnormalities at a median age of 59 years, and new‐onset obesity at a median age of 52 years. During a median follow‐up of 16.7 years from baseline exam, we observed 578 incident CVD events (8%; 5.2 per 1000 person‐years) occurring at a median age of 74 years. Continuous CMD Progression Modeling ARCLPM results for AI‐VA and continuous cardiometabolic measures are presented in Figure 1. We observed bidirectional relations of AI‐VA with fasting blood glucose and triglycerides and unidirectional relations of TC/HDL ratio and BMI with AI‐VA progression (Figure 1). ARCLPM results for blood pressure components and number of CMD states are presented in Figure 2. We observed bidirectional relations of AI‐VA with pulse pressure, MAP, and number of CMD conditions (Figure 2). ARCLPM results were similar when AI‐VA was replaced with niCFPWV as the aortic stiffness variable. Autoregressive pathway relations were stronger for niCFPWV. Cross‐lagged relations of baseline fasting blood glucose and triglycerides with niCFPWV progression were similar to relations with AI‐VA progression. However, cross‐lagged relations of BMI with niCFPWV progression were not significant, despite significant corresponding relations in AI‐VA modeling. Cross‐lagged relations of pulse pressure with stiffness progression were significant with niCFPWV as the measure of stiffness, but not significant with AI‐VA as the measure of stiffness. Conversely, cross‐lagged relations of MAP with stiffness progression were significant with AI‐VA as the measure of stiffness, but not with niCFPWV as the measure of stiffness. Additionally, cross‐lagged relations of number of CMD states with stiffness progression were significant with AI‐VA as the measure of stiffness but not with niCFPWV as the measure of stiffness. Relations of baseline aortic stiffness with progression of continuous cardiometabolic measures were similar for both measures of stiffness (Figures S2 and S3). Longitudinal relations between AI‐VA and progression of continuous cardiometabolic measures. Autoregressive cross‐lagged panel model associations of AI‐VA and (A) fasting blood glucose, (B) triglycerides, (C) TC/HDL ratio, and (D) body mass index progression from visit 1 to visit 2 among 7188 Framingham Heart Study participants. Analyses were adjusted for visit 1 covariates age, sex, cohort, hypertension treatment, lipid treatment, diabetes status, smoking status, and heart rate from tonometry, as well as time between visit 1 and visit 2. Beta represents standardized coefficient estimate, or the adjusted change in SDs of dependent variable associated with an increase of 1 SD of independent variable. AI‐VA indicates artificial intelligence vascular age; BG, blood glucose; BMI, body mass index; HDL, high‐density lipoprotein; and TC, total cholesterol. Longitudinal relations between AI‐VA and progression of continuous blood pressure components and CMD states. Autoregressive cross‐lagged panel model associations of AI‐VA and (A) pulse pressure, (B) MAP, and (C) number of CMD states progression from visit 1 to visit 2 among 7188 Framingham Heart Study participants. Number of CMD states defined as sum of hypertension, diabetes, lipid abnormality, and obesity status variables. See Methods for CMD status definitions. Analyses were adjusted for visit 1 covariates age, sex, cohort, hypertension treatment, lipid treatment, diabetes status, smoking status, and heart rate from tonometry, as well as time between visit 1 and visit 2. AI‐VA indicates artificial intelligence vascular age; CMD, cardiometabolic disease; MAP, mean arterial pressure; and PP, pulse pressure. In sensitivity analyses assessing confounding by hypertension treatment, we observed stronger relations of AI‐VA with both pulse pressure and MAP when applying a conservative imputation for untreated pressure values instead of adjusting for binary hypertension treatment status (Figure S4A). In sensitivity analyses assessing confounding by diabetes, bidirectional relations between AI‐VA and fasting blood glucose persisted in a subset of participants free from diabetes (Figure S4B). Continuous CMD Progression Effect Modification There was evidence of effect modification by age, sex, and obese BMI in several of the cross‐lagged pathways (Table S2). Results stratified by age, sex and obesity for all models are presented in Tables S3 through S5. In general, cross‐lagged pathway betas were greater in lower risk groups, including younger participants, women, and participants without obesity (Tables S3 through S5). Relations of baseline AI‐VA with progression of continuous cardiometabolic measures were stronger in younger participants for MAP and number of CMD states and were stronger in older participants for pulse pressure. Conversely, relations of baseline cardiometabolic measures with progression of AI‐VA were stronger in younger participants for MAP and number of CMD states. Notably, the relations of baseline BMI with AI‐VA progression were positive in younger participants and negative in older participants (Tables S2 and S3). Relations of baseline AI‐VA with progression of triglycerides, pulse pressure, and MAP were stronger in women than men (Tables S2 and S4). Relations of baseline AI‐VA with progression of MAP and number of CMD states were stronger in participants without obesity. Conversely, relations of baseline cardiometabolic measures with progression of AI‐VA were stronger in participants without obesity for triglycerides and MAP. There were strong bidirectional relations between CMD and AI‐VA for fasting blood glucose, triglycerides, MAP, and the number of CMD states in participants without obesity (Table S5). Incident CMD States Relations between baseline AI‐VA and new onset of various individual CMD states in participants free of each state at baseline are presented in Table 2. In models that adjusted for age, sex, and cohort, higher AI‐VA was associated with increased risk for incident hypertension, diabetes, lipid abnormalities, and obesity (Table 2, Model 1). Relations of AI‐VA with all 4 outcomes persisted after further adjustment for standard vascular and CMD risk factors (Table 2, Model 2). Associations of AI‐VA with incident lipid abnormality or obesity were attenuated and no longer significant after further adjustment for baseline BMI, MAP, and heart rate (Table 2, Model 3). In contrast, AI‐VA remained associated with incident hypertension and diabetes after including these additional adjustments (Table 2, Model 3). To compare the relations of AI‐VA versus conventional cardiometabolic measures with risk of the various incident CMD states, we examined models that included each of the continuous cardiometabolic metrics as well as AI‐VA. As expected, conventional continuous metrics used to define CMD states were strongly related to the respective CMD state. Importantly, AI‐VA was a consistent predictor of each of the incident CMD states except incident obesity in models that adjusted for all other key cardiometabolic metrics (Table S6). BMI was the only continuous metric significantly associated with incident obesity risk. Table 2Associations of AI Vascular Age With Incident Cardiometabolic OutcomesOutcomeEvents/at riskModel 1Model 2Model 3HR (95% CI) P valueHR (95% CI) P valueHR (95% CI) P valueHypertension1155/43432.00 (1.79, 2.22)<0.00011.44 (1.28, 1.61)<0.00011.20 (1.07, 1.36)0.0025Age≤48 y530/28172.66 (2.28, 3.11)<0.00011.81 (1.52, 2.14)<0.00011.34 (1.11, 1.61)0.0023Age>48 y625/15261.54 (1.33, 1.79)<0.00011.18 (1.01, 1.38)0.041.08 (0.92, 1.26)0.38BMI<30 kg/m2 818/3534* 2.16 (1.89, 2.46)<0.00011.52 (1.32, 1.75)<0.00011.27 (1.10, 1.47)0.001BMI≥30 kg/m2 337/8081.41 (1.16, 1.72)0.0011.19 (0.97, 1.46)0.091.05 (0.85, 1.31)0.64Diabetes414/57511.79 (1.50, 2.13)<0.00011.29 (1.07, 1.56)0.0071.26 (1.03, 1.53)0.02Lipid abnormality1126/34601.33 (1.19, 1.48)<0.00011.13 (1.00, 1.28)0.041.12 (0.98, 1.27)0.09Obesity739/45761.44 (1.25, 1.65)<0.00011.23 (1.06, 1.42)0.0081.06 (0.90, 1.24)0.51Model 1 adjusts for age, sex, and cohort. Model 2 adds research center systolic blood pressure, total cholesterol/high‐density lipoprotein cholesterol ratio, hypertension treatment, prevalent diabetes status, triglycerides, fasting blood glucose, and current smoking to Model 1. Model 3 adds mean arterial pressure, heart rate, and body mass index to Model 2. All HRs expressed per 1 SD higher vascular age.AI indicates artificial intelligence; BMI, body mass index; and HR, hazard ratio.*One case was censored before the earliest event in the stratum. Associations of AI Vascular Age With Incident Cardiometabolic Outcomes Model 1 adjusts for age, sex, and cohort. Model 2 adds research center systolic blood pressure, total cholesterol/high‐density lipoprotein cholesterol ratio, hypertension treatment, prevalent diabetes status, triglycerides, fasting blood glucose, and current smoking to Model 1. Model 3 adds mean arterial pressure, heart rate, and body mass index to Model 2. All HRs expressed per 1 SD higher vascular age. AI indicates artificial intelligence; BMI, body mass index; and HR, hazard ratio. One case was censored before the earliest event in the stratum. Incident CMD Effect Modification In models for incident hypertension, we observed significant interactions of AI‐VA with age (dichotomized at the median) and obese BMI (Table S7). In stratified models, relations between AI‐VA and risk for incident hypertension were stronger in younger participants and in participants without obesity (Table 2). We observed attenuated relations of AI‐VA with incident hypertension in participants with BMI ≥30 kg/m2, suggesting that the vicious cycle of CMD progression was reasonably advanced by the time BMI was in the obese range. At baseline, participants with obesity tended to be older with higher AI‐VA, systolic blood pressure, triglyceride concentrations, and prevalence of hypertension, diabetes, and lipid abnormalities, and lower HDL cholesterol (Table S8). We did not observe significant effect modification by age, sex, or obese BMI for associations of AI‐VA with incidence of the other 3 CMD states. Multistate Transitions The transitions into various CMD states and subsequently into CVD events during the full CVD follow‐up period are summarized in Figure 3. All participants were assumed to be free of CMD at some point before visit 1, with any observed CMD states at visit 1 regarded as the result of an initial transition that antedated or occurred at visit 1. At visit 1, 2708 participants had no prevalent CMD, 2127 had only 1 prevalent CMD condition, and the remaining 2353 had ≥2 CMD conditions. Of the 578 observed CVD events, 404 (70%) occurred in participants who had developed ≥2 CMD conditions before the onset of CVD. Only 53 CVD events (9%) occurred in participants who had not developed any CMD conditions before the CVD event. Flow chart of transitions from no CVD at baseline to new‐onset CMD and CVD during follow up in 7188 FHS participants. Participants enter the analysis initially free of CVD and subsequently transition to various single or multiple (2+) CMD states and CVD events. The thickness of connections is proportional to the number of participants transitioning from left node to right node. Transitions from no CMD or single CMD states to CVD are highlighted with a connection that is the color of the respective node, so their width is approximate. Actual transitions into the 578 observed CVD events are as follows: 404 from 2+ CMD, 61 from isolated hypertension, 53 from no CMD, 48 from isolated lipid abnormality (lipids), 10 from isolated obesity, and 2 from isolated diabetes. See Methods for CMD state definitions. CMD indicates cardiometabolic disease; CVD, cardiovascular disease; FHS, Framingham Heart Study. DISCUSSION In our community‐based sample with no prior CVD, AI‐VA was associated with progression of key continuous measures of CMD risk in the short term (6 years) and with risk for various incident CMD states during up to 17 years of follow up. In the short term (6 years), we observed a unidirectional relation of BMI with progression of AI‐VA that persisted in analyses that excluded participants with obesity. Relations of AI‐VA with risk for incident hypertension and diabetes persisted in models that adjust for standard vascular risk factors, as well as BMI, heart rate, and MAP. Using multistate transition analysis, we demonstrate strong relations between presence of CMD states and CVD risk, with 70% of incident CVD events occurring in participants who had developed multiple CMD conditions before the CVD event. The observed unidirectional relation of BMI with progression of AI‐VA and bidirectional relations between AI‐VA and other CMD risk factors in the short term are consistent with our hypothesis that stiffening of the aorta, manifesting as accelerated vascular age, and related downstream abnormalities in microvascular function, play an integral role in the pathogenesis of CMD. The resulting self‐reinforcing CMD vicious cycle provides a potential basis for the well‐known clustering of CMD risk factors and CMD states, which then progress to CVD if the cycle remains uninterrupted (Figure 4). We have previously demonstrated that AI‐VA, an easily assessed measure of biologic vascular age based solely on a brief noninvasive recording of an uncalibrated peripheral arterial pressure waveform, is associated with risk for new‐onset CVD events. 15 Our present study demonstrates the utility of AI‐VA as a novel tool for assessment of risk for progression to various CMD states traditionally considered as major modifiable risk factors for CVD. CMD progression model. Bidirectional relations between accelerated vascular age and microvascular damage in target organs contribute to progression of blood pressure abnormalities, insulin resistance, and lipid abnormalities. This cycle can progress to cardiometabolic disease states (yellow), which contribute to further vascular aging and convey increased risk for major CVD events (red). The cycle may be initiated in part by increases in body size that trigger aortic remodeling to a larger lumen area, which increases stress on elastic fibers in the aortic wall, leading to accelerated vascular age. BMI indicates body mass index; CMD, cardiometabolic disease; CVD, cardiovascular disease; MAP, mean arterial pressure; and PP, pulse pressure. Prior research by Framingham investigators and others has established that stiffening of the aorta leads to transmission of higher levels of pulsatile power into peripheral arteries and the microvasculature, leading to damage and dysfunction in microvascular beds of various organs. 14 , 17 , 34 , 35 Brisk microvascular flow reactivity in various organ systems plays a critical role in normal processing of a nutrient load. Pancreatic islet blood flow is high in a fasting state and increases markedly after meals because of the incretin effect. 36 , 37 High flow facilitates prompt glucose sensing and insulin release. The released insulin normally triggers vasodilation in peripheral skeletal muscle and adipose tissue to facilitate clearance of glucose and lipids, respectively. Any impairment in the microcirculation that introduces lags in insulin release or peripheral response can have deleterious effects on the metabolic response to meals. Delays in responsiveness in a closed‐loop system result in instability and oscillations in levels of the controlled measure (blood glucose), leading to progressive overshoot (hyperglycemia) following a meal and undershoot (hypoglycemia) in the postprandial period. Periods of hyperglycemia damage tissues, including the aorta, because of nonenzymatic glycation of structural proteins, 38 , 39 , 40 , 41 whereas the periods of hypoglycemia could trigger compensatory snacking, leading to insidious consumption of excess calories and weight gain. 42 The aorta remodels in response to changes in body size in order to maintain normal matching between blood flow and aortic lumen area. 43 Higher BMI is associated with higher aortic diameter and weight loss is associated with a reduction in aortic diameter. 44 , 45 However, the mass of elastic fibers in the aortic wall is fixed and the gene program required for elastic fiber synthesis is permanently silenced early in life. 46 , 47 , 48 As a result, when the aortic lumen remodels to a larger area in response to weight gain, a fixed number of elastic laminae are placed under higher stress and strain, leading to higher fractional engagement of collagen and initially reversible stiffening of the aortic wall. 49 However, persistently higher stress can progress to elastin fragmentation that cannot be reversed, leading to stiffening of the aortic wall with advancing age as elastic fibers degrade. 46 , 50 The cycle of aortic stiffening and CMD progression potentially could be prevented, and aortic remodeling reversed, if addressed early, before elastic lamina remodeling progresses to fragmentation, making detection of CMD risk and early intervention critical public health priorities. 2 , 3 Our results show especially strong relations of AI‐VA with incident hypertension and diabetes, which are 2 key CVD risk factors that account for a large burden of quality adjusted years of life lost worldwide. 4 , 51 The CVD risks of high blood pressure are linearly related to systolic blood pressure levels as low as 100 mm Hg, highlighting the need to enhance screening in the community, scale up effective treatment of hypertension proportional to hypertension burden, and increase technical capacity and resources allocated to hypertension prevention at all stages of life. 52 Increased aortic stiffness and the associated increase in pulse pressure play a fundamental role in the pathogenesis of predominant systolic hypertension, which is by far the most prevalent form of hypertension. 53 , 54 , 55 Microvascular damage, remodeling, and dysfunction as a result of accelerated aortic stiffness and transmission of excess pulsatile power to the periphery can increase peripheral resistance and thereby contribute to progression of hypertension due to a necessary increase in MAP in order to maintain cardiac output in the face of increased peripheral resistance. Hypertension is often preceded by measurable alterations in the structure and function of the microvascular bed, as vasoconstriction responses are enhanced and vasodilation responses are reduced. 56 ARCLPM results showed bidirectional relations of AI‐VA with MAP and pulse pressure, with notable age, sex, and obese BMI interactions across the various effect pathways. Bidirectional relations of AI‐VA with MAP were stronger in younger participants and participants without obesity, whereas relations with pulse pressure were stronger in older participants. AI‐VA may trigger MAP progression in early adulthood, as pulsatile power damages the microvasculature, and may increase pulse pressure progression only in later adulthood, as elastin stretch gives way to elastin fragmentation. Stratified Cox modeling shows that AI‐VA was a stronger predictor of new‐onset hypertension among participants that were younger or had BMI ≤30 kg/m2, suggesting that they were potentially earlier in the CMD progression cycle. Overall cardiometabolic risk factor burden was lower in these subsamples, suggesting that an early increase in aortic stiffness may provide an informative early measure of risk for development of hypertension and CMD. AI‐VA was not significantly related to risk of incident hypertension among participants with obesity, which was a smaller stratum with higher risk factor burden. Saturation of established risk factors can diminish incremental discrimination of a novel risk factor in a high‐risk group. Additive value of a novel risk factor is inversely proportional to the accuracy of the base model, and an increase in discrimination is more likely to occur in intermediate‐risk groups as opposed to low‐ or high‐risk groups. 57 , 58 Diabetes has been investigated as both a risk factor for and effect of aortic stiffness in previous studies. 59 , 60 , 61 , 62 , 63 Diabetes progression and treatment outcomes are highly heterogeneous, reflecting the wide variety of distinct pathways that contribute to diabetes progression. Individuals with diabetes exhibit widely varying degrees of skeletal muscle insulin resistance, beta cell dysfunction, incretin response, and hepatic glucose regulation that manifest as severe hyperglycemia. 64 , 65 , 66 , 67 Recently, the role that aortic stiffness plays in diabetes progression has received considerable attention. In FHS participants, higher CFPWV and forward wave amplitude, related measures of aortic stiffness, were associated with blunted microvascular reactivity to ischemic stress, which could potentially contribute to insulin resistance. 16 Measures of continuous glucose variability have been reported to be associated with increased CFPWV. 68 A recent longitudinal study found that higher CFPWV in adolescence may be an independent risk factor for insulin resistance in adulthood. 8 Insulin release may be reduced or delayed as a result of microvascular damage and remodeling in the pancreatic islets following prolonged exposure to excess pressure and flow pulsatility. 60 , 69 Thus, microvascular dysfunction associated with aortic stiffness may simultaneously impair insulin release and contribute to insulin resistance, which is consistent with our ARCLPM finding of bidirectional relations between AI‐VA and fasting blood glucose. Hyperglycemia has been reported to increase aortic stiffness in previous studies. 4 , 38 , 39 , 40 , 41 , 70 A previous study of arterial stiffness preceding diabetes reported that brachial‐ankle PWV, a measure of systemic arterial stiffness, was associated with longitudinal increases in fasting blood glucose and incident diabetes, although no significant association of fasting blood glucose with brachial‐ankle WV was observed. 63 A possible explanation for the latter difference is that brachial‐ankle PWV includes a substantial contribution from muscular arteries in the transit path from arm to ankle and therefore represents a hybrid of elastic aorta and muscular artery properties, which have been shown to change in opposite directions with age and exposure to various risk factors. 71 Our measure of AI‐VA was trained to predict CFPWV, which is a more specific measure of aortic stiffness. Though not as definitive as hypertension and diabetes, our results are consistent with bidirectional relations of AI‐VA with lipid abnormalities and obesity. ARCLPM demonstrated strong bidirectional relations of AI‐VA with triglycerides but did not reveal associations of AI‐VA with progression of continuous measures of BMI or TC/HDL ratio. In Cox models, higher AI‐VA was associated with increased risk for incident lipid abnormalities and obesity, but these relations were attenuated after adding adjustments for BMI, MAP, and heart rate. Obesity is often a reflection of an advanced state of cardiometabolic disease, such that various consequences of aortic stiffness may compete for a component of the AI‐VA variance as seen in the BMI‐adjusted Model 3. We included analyses of effect modification by obesity to assess this phenomenon in our sample. The use of BMI as a single measure of obesity has been questioned recently, in part due to the inability of BMI to differentiate lean and fat mass. However, BMI ≥30 kg/m2 remains the primary definition of obesity used clinically, and we have used it in our obesity event definitions accordingly. BMI as a measure of mass per unit height is an acceptable variable for the purposes of our study, because both lean and fat mass correlate with higher cardiac output in both men and women. 72 Previous longitudinal studies have reported strong bidirectional associations of CFPWV with measures of lean and fat mass, which differs from our unidirectional association with BMI, as well as moderate associations of CFPWV with lipid progression. 8 , 9 These studies focused on the transition from late adolescence to young adulthood in a sample of relatively lean participants and used sophisticated dual energy X‐ray absorptiometry to assess body composition, which may have affected observed relations with aortic stiffness. The burden of high BMI is greater than any other risk factor in regions with relatively low social deprivation such as the United States. 4 Strong interactions with metabolic risk factors including high blood pressure and high fasting blood glucose, as well as behavioral risk factors including low physical activity and diet high in sugar‐sweetened beverages further indicate an urgent need for interventions focused on obesity and CMD. 73 , 74 We observed important sex differences in relations of AI‐VA with continuous cardiometabolic measures. Sex and genetic determinants of body size may contribute to the ultimate capacity of the aorta to remodel in response to changes in body size across the lifespan. Women are known to have a smaller aortic lumen area than men regardless of the approach used to adjust for differences in body size. 43 , 75 , 76 , 77 A recent study suggests that young women may also experience more severe metabolic‐induced cardiovascular alterations than young men. 78 Additional genetic or early life factors may affect production of aortic elastic lamina, resulting in variation in the capacity for the aorta to remodel to a larger lumen area in order to keep pressure pulsatility low despite increasing aortic wall stiffness. 79 In stratified ARCLPM analysis, we observed stronger positive associations between AI‐VA and progression of triglycerides, pulse pressure, and MAP among women compared with men, and we observed a significant association of BMI with AI‐VA progression in women but not in men. These results suggest that women may be predisposed to greater aortic stiffening, acceleration of AI‐VA and progression of cardiometabolic measures as the aorta remodels in response to changes in body size. Some aortic stiffening with advancing age is inevitable, but there are opportunities for intervention to minimize damage to the aorta, microvasculature, and metabolic systems and maximize healthspan. Unhealthy lifestyle factors, such as suboptimal diet and sedentary behavior, contribute to the burden of CMD, including diabetes and CVD, with effects mediated primarily by unhealthy weight gain and obesity. 73 Recent studies in a sample of English youth suggest sedentary behavior is a primary risk factor for aortic stiffness progression in adolescence 80 and that stiffness temporally precedes metabolic syndrome. 81 Personalized dietary interventions have shown some potential to improve cardiometabolic health by favorably affecting triglycerides, glycated hemoglobin, gut microbiome diversity, and body weight with strong adherence to the intervention. 82 , 83 Early detection of aortic stiffness, before signs of CMD that would traditionally indicate the need for intervention, may be an optimal strategy to trigger an early alert for high‐risk individuals. Furthermore, aortic stiffness monitoring may be a valuable tool alongside dietary and lifestyle interventions for curbing progression of and ideally reversing the CMD cycle. CFPWV is an established risk factor for CMD and CVD, 84 but CFPWV data acquisition requires expensive specialized equipment and skill sets that are generally unavailable clinically, especially in low‐resource settings. 17 Given that the AI‐VA model requires only an uncalibrated pressure waveform, the approach could potentially be applied to a relatively high fidelity arterial pressure waveform from any source, including photoplethysmographic waveforms from a smartphone camera or wearable fitness tracker. 85 , 86 The ease of use and relative affordability of AI‐VA position it as a potentially valuable tool for CMD risk assessment and cardiometabolic health monitoring alongside blood pressure monitoring, comprehensive metabolic panels, and body weight monitoring. There are several strengths and possible limitations that should be considered along with the results of our study. One major limitation is the racial and ethnic characteristics of the sample, which comprised mostly White individuals of European ancestry. This limitation is also present in the Icelandic cohort that provided the training and validation samples for the AI‐VA model. 15 Additional research examining AI‐VA and CMD relations in more diverse cohorts will be required to evaluate relevance of our findings in other racial and ethnic groups. In addition, our sample included adults only, albeit including participants from a wide age range. Additional work will be required to demonstrate the value of early detection of aortic stiffness using AI‐VA during childhood and adolescence. Finally, despite showing that aortic stiffness and accelerated AI‐VA can precede and potentially contribute to progression of cardiometabolic measures, our CMD Cox models and ARCLPM analysis are not randomized controlled trials and therefore should be interpreted as associations rather than causality. We cannot exclude the possibility of residual confounding by important cardiometabolic factors not included in our statistical models, such as diet, gut microbiome diversity, exercise habits, genetics, and environmental hazards. 73 , 74 , 87 In particular, the ARCLPM analyses are useful for exploring temporal associations, but causal inference may be limited due to the strict underlying assumptions of the model, including stationarity and linearity of effects, equal time intervals, and no unmeasured confounding. We visually inspected scatterplots for all pathways measured in ARCLPM analysis, as well as age/time relations, and all associations displayed reasonable linearity over the time interval from visit 1 to visit 2. Pulse pressure and MAP have nonlinear relations with age that may or may not be recapitulated in analyses of change over time. 88 We have modeled linear relations of change in values over a relatively short and consistent interval to minimize confounding by nonlinearity, nonrandomized treatment effects, and dropout. Despite some limitations, there are many strengths of our study including our sample of deeply phenotyped individuals spanning a broad age range, multiple detailed assessments of hemodynamics and risk factors over a 14‐year period, and long‐term follow up with careful adjudication of incident CVD events. Our results reinforce a growing body of evidence that implicates elevated aortic stiffness as preceding and contributing to a cycle of pathophysiology that accelerates CMD progression and CVD events. Conclusions In summary, we have evaluated relations of AI‐VA, a measure of aortic stiffness derived from an AI model that was trained to predict CFPWV, with development of new‐onset CMD. We observed relations between AI‐VA and risk for new‐onset hypertension, diabetes, lipid abnormalities, and obesity across a broad age range in a community‐based cohort free of prevalent CVD at the initial visit. We provide one potential unifying pathophysiological basis for the clustering of cardiometabolic risk factors. Our results are consistent with a cyclical model of progressive aortic stiffening, microvascular damage, metabolic dysfunction, and CMD progression defined by bidirectional relations between AI‐VA and continuous CMD variables. Over time, excess pulsatile power conveyed by the stiff aorta may contribute to microvascular damage and remodeling that has a pleiotropic adverse effect on the ability of the microvasculature to rapidly modulate flow in response to diverse physiologic demands in organs throughout the body. 14 , 16 The cycle of progressive CMD can be difficult to interrupt once it is established, underscoring the need for early identification and intervention. We highlight the potential utility of AI‐VA as an early indicator of subclinical CMD and subsequent major adverse clinical events. Sources of Funding This study was supported by National Heart, Lung, and Blood Institute contracts N01‐HC‐25195, HHSN268201500001l, and 75N92019D00031 (Ramachandran S. Vasan) and by DK080739 (Ramachandran S. Vasan), HL107385, HL126136, HL93328, HL142983, HL143227, HL131532 and AG079390 (Ramachandran S. Vasan, Gary F. Mitchell), and HL70100, HL092577, 2U54HL120163, and AG066010 (Emelia J. Benjamin). Ramachandran S. Vasan was supported in part by the Evans Medical Foundation and the Jay and Louis Coffman Endowment from the Department of Medicine, Boston University Chobanian and Avedisian School of Medicine. Naomi M. Hamburg was funded by U54HL120163, American Heart Association 20SRFRN35120118, R01HL115391, and R01HL168889. Leroy L. Cooper was funded by National Heart, Lung, and Blood Institute (K01HL161494). Disclosures Gary F. Mitchell is the owner of Cardiovascular Engineering, Inc., a company that designs and manufactures devices that measure vascular stiffness. The company uses these devices in clinical trials that evaluate the effects of diseases and interventions on vascular stiffness. Gary F. Mitchell also serves as a consultant to and receives grants and honoraria from Novartis, Merck, Bayer, Servier, Philips, and deCODE genetics. Timothy J. Korzinski and David J. Hamel‐Sellman are employees of Cardiovascular Engineering, Inc. Gary F. Mitchell, Timothy J. Korzinski, and David J. Hamel‐Sellman are inventors on a pending patent application that discloses methods for predicting various measures of biological age using pressure waveforms. Gary F. Mitchell is a co‐inventor on a pending patent application that discloses a method for estimating carotid‐femoral pulse wave velocity and vascular age by using a convolutional neural network. The remaining authors report no conflicts. Supporting information Tables S1–S8 Figures S1–S4 Acknowledgments From the Framingham Heart Study of the National Heart, Lung, and Blood Institute of the National Institutes of Health and Boston University School of Medicine. This article was sent to So Mi Cho, PhD, Assistant Editor, for review by expert referees, editorial decision, and final disposition. Supplemental Material is available at https://www.ahajournals.org/doi/suppl/10.1161/JAHA.125.044872 For Sources of Funding and Disclosures, see page 12. References Associated Data Supplementary Materials Tables S1–S8 Figures S1–S4 Data Availability Statement The procedure for requesting data from the FHS can be found at https://www.framinghamheartstudy.org/. Specimens, study data sets, and other resources can be requested at https://biolincc.nhlbi.nih.gov/studies/fhs/.