Vascular age assessed from an uncalibrated, noninvasive pressure waveform by using a deep learning approach: the AI-VascularAge model

Abstract

Background:

Aortic stiffness, assessed as carotid-femoral pulse wave velocity (CFPWV), provides a measure of vascular age and risk for adverse cardiovascular disease (CVD) outcomes, but is difficult to measure. Shape of arterial pressure waveforms conveys information regarding aortic stiffness; however, best methods to extract and interpret waveform features remain controversial.

Methods:

We trained a convolutional neural network (CNN) with fixed-scale (time and amplitude) brachial, radial, and carotid tonometry waveforms as input, and negative inverse (ni) CFPWV as label. Models were trained with data from 2 community-based Icelandic samples (N=10452 participants with 31126 waveforms) and validated in the community-based Framingham Heart Study (FHS, N=7208 participants, 21624 waveforms). Linear regression rescaled predicted niCFPWV to equivalent artificial intelligence vascular age (AI-VA).

Results:

The AI-VA model predicted niCFPWV with R²=0.64 in a randomly reserved Icelandic test group (N=5061, 16%) and R²=0.60 in FHS. In FHS (up to 18 years of follow-up; 479 CVD, 200 coronary heart disease, and 213 heart failure events), brachial AI-VA was associated with incident CVD adjusted for age and sex (model 1, hazard ratio [HR]=1.79, 95%CI=1.50–2.40/SD, P<0.0001), or adjusted for age, sex, systolic blood pressure, total cholesterol, high-density lipoprotein cholesterol, prevalent diabetes, hypertension treatment, and current smoking (model 2, HR = 1.50, 95%CI=1.24–1.82/SD, P<0.0001). Similar hazard ratios were demonstrated for incident coronary heart disease and heart failure events and for AI-VA values estimated from carotid or radial waveforms.

Conclusions:

Our results demonstrate that CNN-derived AI-VA is a powerful indicator of vascular health and CVD risk in a broad community-based sample.

Graphical Abstract

Studies of the shape and the timing of features of the arterial pulse have intrigued physiologists and physicians since ancient times. Persistent themes include the hypotheses that the volume, frequency, and shape of the pulse waveform convey important prognostic information regarding cardiovascular health. Early work centered on subjective descriptions of pulse characteristics assessed by palpation of various superficial arteries.¹ A more quantitative analysis of the pulse waveform emerged in the mid-19^th century with the development of the sphygmometer by Herisson followed by the kymograph of Ludwig and the sphygmograph of Marey in the late 1800s. Subsequently, Mahomed modified the Marey sphygmograph to include the ability to assess the pressure required to occlude the arterial pulse. Later development of the modern cuff-based sphygmomanometer shifted focus to only the peak and trough of the arterial pressure waveform, reported as conventional systolic and diastolic blood pressure, which are well-known markers of cardiovascular disease (CVD) risk.² Ease of measurement and interpretation of cuff blood pressure led to the loss of interest in pulse waveform characteristics until the mid-20^th century, which saw a resurgence of interest in and a proliferation of methods and devices to assess various potentially informative features of the arterial pressure waveform.¹

Waveform propagation velocity in the aorta, which is proportional to aortic wall stiffness, has recently emerged as a strong indicator of CVD risk. The reference standard measure of aortic stiffness is carotid-femoral pulse wave velocity (CFPWV), which reclassifies CVD risk in community-based studies when evaluated in models that include standard CVD risk factors.^3;4 Additionally, recent studies in adults and adolescents have shown that CFPWV predicts the development and the progression of various CVD risk factors that were once thought to precede and contribute to aortic stiffening, including hypertension, obesity, hyperlipidemia, hyperglycemia, insulin resistance, and diabetes, possibly as a consequence of microvascular damage and dysfunction in various target organs.^5–7 The foregoing observations suggest that aortic stiffness contributes at a very early stage to the pathogenesis of CVD and its risk factors, and also risk of various additional adverse clinical outcomes, including chronic kidney disease, cognitive impairment and dementia.^8;9 Importantly, several studies have shown that aortic stiffness can be modified by various lifestyle and pharmacological interventions,¹⁰ raising the possibility for interventions that target aortic stiffness as a modifiable risk factor. However, the clinical application of aortic stiffness measurements is challenged by the need for specialized equipment and training required for the proper measurement of CFPWV. A method to robustly assess aortic stiffness from an easily accessible pressure waveform would offer the opportunity to leverage the clinical value of aortic stiffness assessment in a broader context, including self-assessment by individuals interested in evaluating the effects of lifestyle modifications.

To address this knowledge gap and potentially unmet need, we developed a deep learning model to assess aortic stiffness from a brief 20-second recording of any one of several (carotid, brachial, radial) noninvasively recorded pulse waveforms. Moreover, we sought to examine the implications of the observation that normal aortic PWV in young healthy animals of various species, including humans, is an intrinsic property of the aorta that is similar across an order of magnitude of body size and heart rates in mammals. When the time scale is obscured, pressure waveform morphology is also remarkably similar across a wide range of heart rates, which are inversely related to body size.¹¹ Based on the foregoing, we hypothesized that the intrinsic shape of the arterial pressure waveform is affected by and conveys information regarding aortic stiffness and vascular function that is independent of calibrated pressure amplitude, heart rate, systolic ejection period, and body size. Rather than attempt to extract engineered features from the pressure waveform based on preconceived notions of arterial system characteristics, we trained a deep convolutional neural network (CNN), the AI-VascularAge (AI-VA) model, to extract unbiased waveform features agnostically in order to predict negative inverse CFPWV (niCFPWV) from pressure waveforms with a fixed time and amplitude scale. For this training phase, we used a large, community-based sample of Icelandic participants in the AGES-Reykjavik and REFINE-Reykjavik cohort studies. In a separate sample of participants in the Framingham Heart Study, we used the trained model to predicted niCFPWV, which was converted to an equivalent AI-VA by using linear rescaling, and then assessed the prognostic value of AI-VA in the validation cohort.

Methods

The procedure for requesting data from the Framingham Heart Study can be found at https://framinghamheartstudy.org/. Requests for hemodynamic data derived from evaluations at the AGES Reykjavik and Refine Reykjavik Studies can be directed to the authors.

Waveform acquisition and preprocessing

Details of the waveform acquisition and analysis are presented in the online supplement. Briefly, waveforms were assessed by using a custom tonometer and the Noninvasive Hemodynamic Data Acquisition and Analysis System (NIHem, Cardiovascular Engineering, Inc., Norwood, MA). Carotid-femoral pulse wave velocity was assessed from body surface measurements between carotid and femoral recording sites, adjusted for parallel transmission, and divided by the carotid-femoral transit time.¹² CFPWV values were clipped at 30 m/s and inverted to normalize the distribution, linearize the age relation and limit heteroskedasticity. Values were scaled by a factor of −3.0 to compute niCFPWV = −3.0/CFPWV. The foregoing transformation maps typical CFPWV values, which range from 3 to 30 m/s, to corresponding unitless negative inverse CFPWV (niCFPWV) values, which range from −3.0/3.0 = −1.0 (least stiff aortas) to −3.0/30.0 = −0.1 (stiffest aortas). Signal-averaged carotid, brachial and radial waveforms were then rescaled to a fixed amplitude (range 0–1) and time scale (systolic ejection period 320 points, RR interval 1024 points) by linear rescaling along the time and pressure axes. As a result of rescaling, the CNN model was provided with no information regarding calibrated values for blood pressure, heart rate, systolic ejection period, or the absolute timing of waveform landmarks such as the inflection point marking the return of wave reflections to the proximal aorta. This dual axis standardization of the pressure waveforms forces the model to extract features that depend only on intrinsic waveform shape.

CNN model architecture

The AI-VascularAge (AI-VA) model is described in the online supplement. Briefly, the model consists of an input layer that includes the time- and amplitude-normalized pressure waveform and the first (dP) and second (d2P) derivatives of normalized pressure, each in a separate input channel. The input layer is followed by 8 convolution layers, each with 256 convolution filters, a fixed kernel size of 15 points, a rectified linear units (ReLU) activation layer, an average pooling layer, and a dropout layer as described in the online supplemental methods. The average pooling layer reduces the dimensionality of the feature space by a factor of 2 at each layer by averaging pairs of adjacent features. The dropout layer randomly replaces a fraction of feature space points with zeroes in order to prevent over-training. The output of the convolution layers is connected to a single output node with a linear activation function that predicts niCFPWV (the label variable).

CNN models were trained exclusively on waveforms acquired in Icelandic cohorts. Waveforms from various sites (carotid, brachial and radial, total N = 31,153) were combined, shuffled, and split before each training run into training (64%, N=19,840), validation (20%, N=6225) and test (16%, N=5061) sets, with adjustment of the training set size to represent a truncated integer multiple of the batch size (N=128), with the remaining cases redistributed to the test set. Training consisted of 4 repetitions for each model, with the best fit (maximum R²) selected automatically based on the relation between observed and predicted niCFPWV in the test group. See the online supplemental methods for additional details.

The best model was used to predict niCFPWV in the Framingham cohorts. Values for predicted niCFPWV (pniCFPWV) were regressed against age in the Framingham cohorts. The slope (m) and intercept (b) of the relation of pniCFPWV with age were then used to compute estimated AI-VA based on pniCFPWV values predicted by the CNN model as follows: AI-VA = (pniCFPWV – b) / m.

Feature maps and sensitivities.

Waveform feature maps were examined by slicing the AI-VascularAge model at the output of the average pooling layer of each convolution section. We then performed a prediction through to the output of that layer, using as the input the various waveform types from Framingham cases, each analyzed separately. Feature map sensitivities to niCFPWV were assessed by computing the correlation coefficient of the relation between niCFPWV and feature values separately for each feature on the feature map. The feature map sensitivities were converted to 2-dimensional arrays that were then presented as cluster maps. Clustering was performed on the sensitivity map by using the SciPy hierarchical clustering algorithm with clustering by filter axis only.

Clinical outcomes.

Major CVD events were examined in the Framingham sample only and were defined as fatal or nonfatal myocardial infarction, unstable angina (prolonged ischemic episode with documented reversible ST segment changes), heart failure, and ischemic or hemorrhagic stroke. 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.^13;14 Follow-up evaluations were performed on data acquired through December 31, 2019.

Statistical analysis

Characteristics of the training and clinical validation cohorts were summarized separately and tabulated. Relations between the observed and the predicted niCFPWV was assessed with linear regression analysis.

We examined the associations between AI-VA and time to a first major CVD event by using Cox proportional hazards regression, after confirming that the assumption of proportionality was met based on visual inspection of Kaplan-Meier plots. Covariates were selected based on components of the Framingham risk score¹⁵ and included the following at the baseline examination for each cohort: age, sex, clinically assessed seated systolic blood pressure, use of antihypertensive therapy, total and high-density lipoprotein cholesterol concentrations, regular use of cigarettes in the prior year, and presence of diabetes mellitus. Continuous net reclassification analysis was performed by examining risk predicted by the standard risk factor model before and after adding an AI-VA variable to the model.¹⁶ Cumulative probability curves were constructed by using the Kaplan–Meier method, with participant groups segregated according to quartiles of AI-VA.

To determine whether the relations between hemodynamic measures and CVD events differed in older versus younger participants (dichotomized at the median age) or in men versus women, we included statistical interaction terms for these variables in separate models that also adjusted for standard risk factors. All analyses were performed with SAS version 9.4, SPSS version 28 and R version 4.2.3. A 2-sided P<0.05 was considered statistically significant.

Results

Participants.

The design and inclusion criteria of the AGES-Reykjavik and the REFINE-Reykjavik cohorts^17;18 and the Framingham Offspring, New Offspring Spouse, Third Generation and racially and ethnically diverse Omni-1 and Omni-2 cohorts^19–21 have been previously described. For the training phase of this current investigation, we included participants who attended the first or second examination cycle of the AGES-Reykjavik study or the first examination cycle of the REFINE-Reykjavik study and had a successful assessment of CFPWV. For the validation phase, we included from the Framingham Study participants those who attended the eighth examination cycle (2005–2008) for the Framingham Offspring, the third examination cycle (2005–2008) for Omni-1, and the first examination cycle for Third Generation, Omni-2 and New Offspring Spouse groups (2002–2005).

For the training cohort (N=12,444), we excluded participants with no tonometry assessment (N=397), no usable tonometry waveform (N=212), or no valid assessment of CFPWV (N=1383), resulting in the final sample (N=10,452 participants contributing 31,126 usable waveforms, Table 1). For the validation cohort (N=7781), we excluded participants who were missing any risk factor data required for the clinical outcome models (N=516) or any of the 3 tonometry waveforms (N=57), resulting in the final sample (N=7208, Table 1). The Icelandic cohort was, on average, a decade older, although both cohorts span a broad age range. Consistent with higher average age, average systolic blood pressure was higher and average diastolic blood pressure was lower in the Icelandic cohort, with greater hypertension medication usage and higher prevalence of cardiovascular disease.

Table 1.

Participant characteristics assessed in the clinic.

Variable	Iceland	Framingham
Total N (% women)	10452 (53.7)	7208 (53.9)
Age, y	62 ± 17	51 ± 16
Height, m	1.71 ± 0.10	1.69 ± 0.10
Weight, kg	79 ± 15	78 ± 18
Body mass index, kg/m2	26.7 ± 4.1	27.1 ± 5.1
Seated systolic blood pressure, mm Hg	130 ± 21	121 ± 16
Seated diastolic blood pressure, mm Hg	71 ± 10	74 ± 10
Total cholesterol, mg/dL	203 ± 41	188 ± 36
HDL cholesterol, mg/dL	60 ± 17	56 ± 17
Total/HDL cholesterol, unitless	3.6 ± 1.1	3.6 ± 1.3
Triglycerides, mg/dL	89 (66, 122)	95 (67, 137)
Fasting blood glucose, mg/dL	100 ± 18	100 ± 21
Prevalent cardiovascular disease, N (%)	1,242 (13)	281 (4)
Lipid-lowering medications, N (%)	1,934 (20)	1,538 (21)
Hypertension medications, N (%)	3,953 (40)	1,796 (25)
Diabetes medications, N (%)	391 (4)	367 (5)
Current smoking, N (%)	1,610 (17)	860 (12)

All values mean ± SD, median (25^th, 75^th percentile), or N (%). HDL, high-density lipoprotein.

Training accuracy.

Following training in the Icelandic train (N=19840, 64%) and validation (N=6225, 20%) waveform sets, the AI-VA model predicted niCFPWV values in a randomly reserved test group of Icelandic waveforms (N=5061, 16%) with R² = 0.64 for observed versus predicted niCFPWV. The corresponding values for observed versus predicted niCFPWV in Framingham cases were R² = 0.60 for carotid, brachial and radial waveforms evaluated separately.

Vascular age.

Standard hemodynamic variables, observed and predicted niCFPWV and AI-VA estimates, are presented in Table 2. The correlations of predicted niCFPWV with age for each waveform type in Framingham data were R² = 0.68, 0.70, and 0.72 for carotid, brachial and radial waveforms, respectively, whereas the correlation of measured niCFPWV with age was somewhat lower (R² = 0.58). Histograms of chronological age and predicted values for AI-VA are presented in supplemental Figure S1. Relations of model-derived AI-VA with chronological age are presented in supplemental Figure S2.

Table 2.

Hemodynamic data recorded in supine participants during the tonometry evaluation in the clinical validation set (N=7208).

Variable	Value
Systolic blood pressure, mm Hg	129 ± 19
Diastolic blood pressure, mm Hg	68 ± 9
Mean pressure, mm Hg	93 ± 12
Pulse pressure, mm Hg	61 ± 17
Heart rate, beats/min	61 ± 10
Augmentation index, %	11 ± 14
Backward wave arrival time, ms	131 ± 29
niCFPWV, unitless	−0.39 ± 0.10
Carotid pniCFPWV, unitless	−0.38 ± 0.09
Brachial pniCFPWV, unitless	−0.37 ± 0.09
Radial pniCFPWV, unitless	−0.38 ± 0.09
Carotid vascular age, years	51 ± 19
Brachial vascular age, years	51 ± 19
Radial vascular age, years	51 ± 18

niCFPWV, negative inverse carotid-femoral pulse wave velocity; pniCFPWV, predicted negative inverse carotid-femoral pulse wave velocity.

Feature maps.

Feature sensitivity maps generated from the output of the first convolution layer for each of the pressure waveform types are presented in Figure 1. Signal-averaged, standardized pressures along with the first and second derivatives of pressure were ensemble averaged by waveform type (Figure 1A). Differences in the averaged waveforms are evident, with an early peak and negative systolic augmentation for radial waveforms, a late peak and positive systolic augmentation for carotid waveforms, and intermediate morphology for brachial waveforms. These complex waveform features are characteristics of the entire arterial system, including the input flow waveform and properties of the aorta, muscular arteries, and microvasculature.

Figure 1.

Signal-averaged pressure waveforms (P), with first (dP) and second (d2P) derivatives (A) and sensitivity heatmaps of features of the carotid (B), brachial (C), and radial (D) waveforms. Sensitivities were computed by assessing the correlation of individual features output by the first layer of the convolutional neural network with niCFPWV labels in the clinical validation set (N=7208). Heatmaps have time along the horizontal axis and filter channels along the vertical axis. Outer blue vertical reference lines are placed at the first and last peaks of the second derivative, corresponding to the foot and dicrotic notch of the waveforms. An intermediate blue vertical line is placed at the approximate mid-systolic inflection point, representing the return of the global reflected pressure wave. See text for details.

Model sensitivities to features created by the first layer of the CNN are presented by the waveform site in Figure 1B–D. Localized temporal regions of marked sensitivity to CFPWV are evident as hotter colors for higher CFPWV and cooler colors for lower CFPWV (Figure 1B–D). The various filters have adapted differing characteristics that extract distinct features of the input waveforms. It is important to note that each of these filters was exposed to the full set of inputs (P, dP, d2P) and waveform types and adapted their frequency response during training to enhance the information of interest for a given filter. For example, filters in the top half of each panel are largely focused on diastolic features that are related to niCFPWV, while those in the bottom half are more focused on systolic features. Brachial and radial waveforms produce areas of marked sensitivity near the dicrotic notch (Figure 1CD, yellow ovals), whereas carotid waveforms have enhanced sensitivity near the foot of the waveform in the region of the dP peak (Figure 1B, green oval).

AI-VA and clinical events.

We examined the association of AI-VA with incident clinical events for each of the pulse waveforms separately (Table 3). In models that adjusted for chronologic age and sex (Model 1), AI-VA predicted by each of the pulse waveforms was strongly associated with incident CVD, coronary heart disease, and heart failure events. These associations persisted in models that additionally adjusted for standard CVD risk factors (Model 2). Continuous net reclassification analysis demonstrated significant reclassification for each event and AI-VA model combination except for radial AI-VA in the coronary heart disease model (Supplemental Table S2). We observed no evidence of effect modification by median age (50 years) for any endpoint or waveform type (interaction P>0.24). We found significant statistical interactions between AI-VA and sex for the outcome of CVD only (P≤0.013 for each waveform type), where hazard ratios were modestly higher in men for estimates based on carotid waveforms and in women for brachial and radial waveforms (Supplemental Table S3). The statistical significance of the sex interaction term was attenuated after the addition of a separate sex interaction term for current smoking status to the pooled sex multivariable model. In a secondary analysis, relations of CFPWV and vascular age with incident stroke were examined (Supplemental Table S4). Associations were less robust than those with coronary heart disease and congestive heart failure events (Table 3), with similar hazard ratios but wider confidence intervals for this endpoint with only 161 events.

Table 3.

Association of vascular age with the incidence of cardiovascular disease events.

	Model 1			Model 2
Site	HR	(95% CI)	P value	HR	(95% CI)	P value
Cardiovascular disease (479 / 6927)
CFPWV	1.56	(1.35, 1.80)	<0.0001	1.32	(1.13, 1.54)	0.0005
C	1.59	(1.36, 1.86)	<0.0001	1.37	(1.17, 1.62)	0.0001
B	1.79	(1.50, 2.14)	<0.0001	1.50	(1.24, 1.82)	<0.0001
R	1.67	(1.40, 1.98)	<0.0001	1.44	(1.20, 1.73)	<0.0001
Coronary heart disease (200 / 7025)
CFPWV	1.88	(1.50, 2.34)	<0.0001	1.46	(1.14, 1.87)	0.0026
C	1.74	(1.37, 2.20)	<0.0001	1.39	(1.09, 1.79)	0.008
B	2.16	(1.66, 2.81)	<0.0001	1.64	(1.24, 2.19)	0.0006
R	1.79	(1.39, 2.32)	<0.0001	1.43	(1.08, 1.88)	0.012
Congestive heart failure (213 / 7137)
CFPWV	1.43	(1.16, 1.76)	0.0009	1.20	(0.95, 1.50)	0.12
C	1.97	(1.54, 2.51)	<0.0001	1.67	(1.29, 2.15)	<0.0001
B	1.93	(1.45, 2.56)	<0.0001	1.65	(1.22, 2.23)	0.0013
R	2.09	(1.59, 2.75)	<0.0001	1.86	(1.39, 2.48)	<0.0001

Model 1 adjusts for age, sex, and cohort. Model 2 adds clinic systolic blood pressure, total cholesterol, high-density lipoprotein cholesterol, prevalent diabetes, hypertension treatment, and current smoking to Model 1. All hazard ratios (HR) expressed per 1 SD higher CFPWV or vascular age. Pulse sites are: C, carotid; B, brachial; R, radial. CFPWV, carotid-femoral pulse wave velocity.

Cumulative incidences of CVD, coronary heart disease, and heart failure by quartiles of AI-VA estimated from brachial waveforms are plotted in Figure 2. Incident events were very low in cases in the lowest quartile of AI-VA and were markedly higher in those in the highest quartile of AI-VA with intermediate values in the middle quartiles.

Figure 2.

Cumulative incidence plots for cardiovascular disease (A), coronary heart disease (B), and heart failure (C) for groups based on quartiles of vascular age determined by using brachial tonometry waveforms.

Discussion

To our knowledge, we have developed and validated the first artificial intelligence model that can predict a measure of biological age that is strongly associated with risk for adverse CVD events using as input only a brief 20-second noninvasive recording of either a carotid, brachial or radial pressure waveform. Waveforms were signal-averaged and rescaled to create a single representative beat with a fixed amplitude (0–1) and time scale (1024 points per beat, 320 points in systole), thereby effectively removing the time and pressure scale from the model. The single beat pressure waveform and its first and second derivatives were standardized (Z score by signal channel and case) and used to train a deep CNN regression model using niCFPWV as the label. The model was trained exclusively in a white cohort of European ancestry from Iceland and validated clinically in a separate US-based cohort from Framingham that included non-white participants (Omni-1 and Omni-2). Predicted values for niCFPWV were converted to an equivalent AI-VA, which was shown to predict risk for incident CVD events in models that adjusted for standard CVD risk factors, including chronologic age and all components of the Framingham risk score.

We chose niCFPWV as the training label for our model because it has been shown to relate to abnormalities in several key CVD risk factors, including blood pressure progression and incident hypertension, obesity, hyperglycemia, insulin resistance, incident diabetes, and lipid abnormalities.^5–7;22 In addition, niCFPWV is a strong predictor of CVD events that has been shown to reclassify risk in community-based samples, especially in younger individuals.^3;4 Our investigation is the first to demonstrate that a holistic, agnostic examination of the intrinsic shape of the arterial pressure waveform, devoid of pressure or time calibration, can provide critical insights into cardiovascular health. Additionally, the findings of our investigation, including the observation that waveform calibration is not required, suggest that AI-VA offers an opportunity to assess and monitor long-term risk for the development of adverse levels of various key risk factors and adverse clinical outcomes using relatively simple point-of-care and wearable devices. If replicated in larger multi-ethnic samples, AI-VA might provide an ideal opportunity for early detection of risk and targeted, primordial prevention of cardiometabolic disorders and CVD.

Importantly, whereas the model was trained using niCFPWV as a label, the resulting model-predicted AI-VA was in some cases more closely related to events than the value for measured niCFPWV (see heart failure events, Table 3). This improvement reflects the ability of AI models to learn prevailing features of the data while ignoring inconsistent cases or labels, provided the model is trained with sufficient quantities of correctly labeled data. Noise (mislabeling in this case) in an exposure variable produces regression dilution bias, which reduces the regression slope when predicted values are compared to labels or the hazard ratio when labels are directly used as the exposure variable in an outcomes model.²³ Prior work has shown that AI models that include adequate measures to prevent overfitting, such as the dropout layers and early stopping employed in our model, are remarkably robust to mislabeling.^24;25 As a result, the AI-VA model offers a novel approach to the assessment of aortic stiffness that is easier to perform and potentially more robust than directly assessed CFPWV, a premise that warrants further replication in larger multi-ethnic samples.

The observation that our model is based on an arterial waveform that has been signal averaged and standardized with respect to time, systolic ejection period, and amplitude demonstrates that such a pressure waveform represents a complex interaction of an optimal universal pulse wave velocity with highly variable but tightly coupled body length, heart rate and standardized arterial compliance and peripheral resistance, resulting in the consistent relative timing of wave reflections and systolic and diastolic pressure time constants.¹¹ While prior studies have examined many of the foregoing components of arterial function in isolation, the AI-VA model integrates useful information from all aspects of arterial function that impact the shape of the arterial pressure waveform. Our study included observations from individuals with a wide range of heart rates (31–109 beats/min), cardiac cycle durations (358–2002 ms), and torso lengths (410 to 700 mm). However, additional work in children and various other species will be required to determine whether the model can accurately predict aortic pulse wave velocity across a wider range of heart rates and body lengths using only a fixed time- and amplitude-scaled pressure waveform.

To better understand the features of the waveform that were most interesting to the model, we examined sensitivities to niCFPWV of features that were identified by the first convolution layer of the model. The initial layers of a CNN are often thought to examine only very localized features of the waveform. While it is true that the convolution kernel has a limited temporal extent (approximately 15 ms depending on heart rate), it is important to recall that each channel of the CNN represents a finite impulse response filter with a tunable frequency response that can enhance features of the waveform that span a considerably longer relative temporal extent than the filter kernel per se. Figure 1 illustrates this point by showing that individual filters were able to identify various distinct and, in some cases, relatively long temporal regions that were closely related to niCFPWV. Deeper layers of the model will have access to increasingly longer temporal extents by the fixed kernel length because of the 2:1 pooling after each convolution layer. In light of the improved performance with increased numbers of layers up to the final model including 8 layers, this broadening of the effective temporal extent of the convolution kernel presumably is required to integrate relative timing and relations of various features across the full cardiac cycle and predict an accurate niCFPWV.

Prior work at Framingham and by others has shown that CFPWV is strongly related to the progression of various risk factors and increased risk of several adverse clinical outcomes, underscoring the critical importance of aortic function.^3;4 The observation that only a standardized arterial pulse waveform is needed, with no requirement for pressure or time calibration, raises the possibility for incorporation of the algorithm into standard wearable devices, which could facilitate self-assessment and long-term evaluations to monitor the success of various potential lifestyle modifications that could favorably impact CFPWV and promote long-term vascular health and successful aging. Therefore, the AI-VA model might offer the potential to examine risk for the future development of abnormal risk factors and adverse outcomes at a very early stage, when primordial prevention through modest lifestyle interventions may be highly effective.

There are several strengths and potential limitations of our investigation that should be considered. Limitations of our investigation include the race/ethnic composition of our training and clinical validation samples, which were mostly white individuals of European ancestry. Additional work will be required to demonstrate utility in other racial and ethnic groups. In addition, we examined adults only, albeit across a very broad age range. Additional studies will be required to assess the accuracy and utility of our AI-VA algorithm in children and adolescents. Our AI-VA model uses a deep CNN with more than 6 million trainable parameters, resulting in a considerable capacity to memorize input data. To prevent over-training of the model, we implemented an aggressive dropout rate (0.30) on each convolution layer and utilized early stopping based on model loss in a validation sample. In addition, we used a separate holdout test sample to assess the accuracy and determine the best model at the end of a series of four training runs. Finally, we validated model utility in a distinct and separate cohort with no participant overlap, thereby preventing predictions based on memorized waveform characteristics.

Balancing these limitations, there are many strengths of our study. We were able to train and validate the model with large quantities of uniformly ascertained, high-quality arterial pressure waveforms that were routinely assessed as a part of two large, ongoing epidemiological studies involving a total of more than 17,000 deeply-phenotyped participants who spanned a broad age range. In addition, our clinical validation cohort had a comprehensive standardized assessment of routine CVD risk factors at baseline and up to 20 years of follow-up to detect centrally adjudicated endpoints that were blinded to arterial waveform measures.

Perspectives

In summary, we have developed an AI-VA model that can predict a clinically-relevant vascular age that is strongly associated with incident CVD events in a community-based sample with a broad range of chronological age and baseline risk based on standard CVD risk factors. The AI-VA model offers a novel approach to personalized medicine that allows for assessment and long-term monitoring of aortic stiffening in a broad segment of the population. Considering known associations of niCFPWV with risk for microvascular damage and dysfunction in various target organs, a noninvasive method to detect and monitor aortic stiffness and vascular aging should facilitate successful intervention at the very earliest stages of the pathogenesis of various diseases, including diabetes, hypertension, coronary artery disease, dementia, chronic kidney disease, and heart failure.

Novelty and Relevance.

What is New?

The artificial intelligence vascular age (AI-VA) model can predict a prognostically important measure of vascular age from a single 20-second noninvasive recording of an amplitude- and time-normalized carotid, brachial or radial pressure waveform.

What is Relevant?

AI-VA is strongly related to risk for various adverse cardiovascular disease outcomes.

Clinical/Pathophysiological Implications?

The AI-VA model provides a simple tool for rapid point-of-care assessment of vascular age and cardiovascular disease risk, providing potential opportunity for primordial prevention of cardiometabolic abnormalities that are associated with accelerated vascular aging.

Abbreviations:

AI-VA

Artificial intelligence vascular age

CFPWV

Carotid-femoral pulse wave velocity

CVD

Cardiovascular disease

CNN

Convolutional neural network

FHS

Framingham Heart Study