Development and Validation of the AI-HeartAge Model in Framingham and UK Biobank

Abstract

Background:

Arterial pressure waveform shape conveys information regarding interactions between the left ventricle (LV) and aorta that could provide an estimate of biological heart age and cardiovascular disease (CVD) risk.

Methods:

Artificial intelligence heart age (AI-HA) was estimated by averaging results from two convolutional neural networks (CNNs) trained to predict mitral annulus tissue Doppler e’ and s’ peak velocities using an uncalibrated arterial tonometry or photoplethysmography waveform as input. Models were developed using Framingham Heart Study (FHS) participant pressure waveforms and echocardiographic measurements (N=6916 participants, 38174 waveforms, 56% women, mean age 61±12). We validated AI-HA using Cox modeling in a FHS holdout set of baseline radial waveforms (N=7018, 54% women, age 50±16 years) and in UK Biobank (UKB) participants (N=67986, 53% women, age 57±8 years).

Results:

In FHS (up to 10 years of follow up, 148 heart failure (HF) and 331 CVD events), using models that adjusted for PREVENT risk factors, AI-HA was associated with incident HF (HR=2.09; CI: 1.64, 2.68; continuous net reclassification [CNR]: 0.22, CI: 0.13, 0.30) and CVD (HR=1.52, CI: 1.28, 1.81; CNR: 0.13, CI: 0.07, 0.20). In UKB (up to 10 years of follow up, 1408 HF and 2709 CVD events), AI-HA was associated with incident HF (HR=1.23; CI: 1.13, 1.33; CNR: 0.09, CI: 0.06, 0.12) and CVD (HR=1.22; CI: 1.15, 1.30; CNR: 0.08, CI: 0.06, 0.09).

Conclusions:

AI-HA is a novel and accessible measure of LV function and HF risk in community-based samples.

Graphical Abstract

Heart failure (HF) and cardiovascular disease (CVD) are major health concerns throughout the world, including low- and middle-income countries.¹ Prior studies examined relations of HF risk with echocardiographic measures of diastolic function of the left ventricle (LV).² However, echocardiography is expensive, time consuming, and may not be widely available in low-resource settings. Recently, we developed an artificial intelligence (AI) model that was trained to predict aortic stiffness from an uncalibrated, noninvasive, peripheral pressure waveform.³ The AI-VascularAge (AI-VA) model was trained in REFINE-Reykjavik and AGES-Reykjavik participants and strongly predicted incident CVD events in Framingham Heart Study (FHS) participants. Since arterial pressure waveforms represent complex interactions between cardiac and vascular function, we inferred that similar AI models could be trained to predict heart function to derive an analogous AI-HeartAge (AI-HA). Therefore, we hypothesized that by using an arterial pressure waveform as input and tissue Doppler measures of LV systolic and diastolic function as labels, we could develop an AI model that predicts a clinically relevant measure of biological heart age (AI-HA). To test this hypothesis, we trained a pair of deep convolutional neural networks (CNNs) to extract waveform features agnostically from uncalibrated pressure waveforms to predict tissue Doppler measures of LV systolic (s’ peak) and diastolic (e’ peak) function. Predictions were standardized, averaged, and converted to equivalent biological age in an internal test sample to derive AI-HA. We used waveform data and long-term follow up from the Framingham Heart Study (FHS) and UK Biobank (UKB) to validate the prognostic implications of AI-HA.

Methods

The procedure for requesting data from the Framingham Heart Study can be found at https://www.framinghamheartstudy.org/. The procedure for requesting access to data from the UK Biobank can be found at https://ams.ukbiobank.ac.uk/ams/. Further details on the data sources are provided in the Supplemental Methods.

Data Sources.

Development and internal test samples were drawn from the Framingham Offspring, New Offspring Spouse, Third Generation, and Omni-1 and Omni-2 cohorts, which have been described.^4–6 Waveforms and echocardiography were assessed at 3 visits spanning a 14-year period (Supplemental Figure S1). Waveforms and echocardiographic measures used to train the AI-HA model were acquired in a development set, consisting of Framingham Offspring participants at examination cycles 9 and 10, Omni-1 participants at examination cycles 4 and 5, and Third Generation, New Offspring Spouse, and Omni-2 participants only at examination cycle 3 because tissue Doppler was not assessed at examination 2. Internal validation was performed in an internal test set in Framingham Offspring participants at examination cycle 8, Omni-1 participants at examination cycle 3, and Third Generation, New Offspring Spouse, and Omni-2 participants at examination cycle 1. Details of data processing and exclusion criteria are provided in Supplemental Methods and Supplemental Figures S1–S3. All protocols were approved by Boston University Medical Campus Institutional Review Board, and all participants provided written informed consent.

The model was further validated by examining relations of AI-HA with incident heart failure and CVD events in an external test sample that was drawn from UK Biobank (UKB). Peripheral pulse waves were acquired using a fingertip photoplethysmographic (PPG) device. We evaluated AI-HA using PPG waveforms in a random 50% subset of participants with PPG waveforms recorded during the first visit (N = 67986). Details of exclusion criteria are provided in Supplemental Methods and Supplemental Figure S4. UKB was approved by the North West Multi-Centre Research Ethics Committee, and all participants provided informed consent.

Waveform acquisition and preprocessing.

Signal averaged carotid, brachial, and radial arterial pressure waveforms were assessed in FHS participants by using a custom tonometer and the Noninvasive Hemodynamic Data Acquisition and Analysis System (NIHem, Cardiovascular Engineering, Inc., Needham, MA). Tonometry waveforms were transformed to PPG-equivalent waveforms by using a generalized transfer function, which was generated from paired fingertip PPG and radial arterial waveforms from a separate cohort (Supplemental Methods).^7,8

Assessment of LV diastolic and systolic function.

We assessed lateral mitral annulus tissue Doppler from the apical 4-chamber view. Peaks (Emax and Smax) of the early diastolic (e’) and systolic (s’) waves were assessed. Values were clipped and normalized for use as labels to train 2 separate AI models as detailed in Supplemental Methods.

CNN model development.

The component AI-HA models are 8-layer CNN models, each analogous to the AI-VA model,³ with an input layer that includes the time- and amplitude-normalized pressure (P) waveforms and their first (dP) and second (d2P) derivatives in separate input channels. Models were trained on waveforms acquired from FHS participants at the second and third visits only (development set, Supplemental Figure S1, S2). After completion of 6 development repetitions, best fit models were selected automatically based on the relation between observed and predicted Emax or Smax (maximum R²) in holdout test sets.

AI-HA was predicted from PPG-transformed radial waveforms recorded at visit 1 in FHS and native PPG waveforms in UKB. Feature maps of the initial convolutional layer were constructed to visualize model sensitivity to specific elements of waveform morphology in PPG-transformed (FHS internal test data) and native PPG (UKB external test data) waveforms.

Clinical outcomes.

Major CVD events in FHS were defined as the first occurrence of a 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 internal test data examination in participants with no prior history of these events. Follow-up evaluations were performed on data acquired through December 31, 2021. Follow-up was censored 10 years after the baseline examination date.

Major CVD events in UKB were defined using the first occurrences data-fields, which extract first occurrences of various 3-digit International Classification of Disease (ICD-10) diagnostic codes across self-report, primary care, hospital inpatient data and death data. Full details of the first occurrences data-fields are provided in Supplemental Methods. Follow-up was censored 10 years after the first examination date.

Statistical analysis.

Characteristics of participants in the development and test sets were summarized separately and tabulated (Table 1). We examined the associations between AI-HA and time to the first major CVD event by using multivariable Cox proportional hazards regression. Covariates for HF and total CVD models were selected based on combined components of the AHA PREVENT-HF and PREVENT-CVD equations.⁹ As a sensitivity analysis, we fitted an extended covariate model that included measured tissue Doppler Emax and Smax, mean arterial pressure, pulse pressure, and augmentation pressure in the internal test set only, based on data availability. Continuous net reclassification and integrated discrimination improvement were assessed by examining risk predictions from a model containing all PREVENT risk factors before and after adding AI-HA.¹⁰ Cumulative probability curves were constructed by using the Kaplan–Meier method, with participants grouped according to quartiles of AI-HA.

Table 1.

Participant characteristics.

Variable	FHS Development	FHS Internal Test Data	UKB External Test Data
Total N (% women)	6916 (56)	7018 (54)	67986 (53)
Age, y	63 ± 12	51 ± 15	57 ± 8
Artificial intelligence heart age, y	61 ± 19	51 ± 20	57 ± 15
Height, m	1.67 ± 0.10	1.69 ± 0.10	1.69 ± 0.09
Weight, kg	80 ± 18	78 ± 17	79 ± 16
Body mass index, kg/m2	28.4 ± 5.6	27.1 ± 5.1	27.5 ± 4.8
Systolic blood pressure, mm Hg	124 ± 16	121 ± 16	141 ± 20
Diastolic blood pressure, mm Hg	73 ± 9	74 ± 10	82 ± 11
Total cholesterol, mg/dL	185 ± 38	188 ± 36	220 ± 44
HDL cholesterol, mg/dL	60 ± 19	56 ± 17	56 ± 15
Total/HDL cholesterol ratio	3.3 ± 1.1	3.6 ± 1.3	4.1 ± 1.1
Triglycerides, mg/dL	95 [70, 133]	95 [67, 137]	131 [93, 189]
Fasting blood glucose, mg/dL	102 ± 21	99 ± 21	94 ± 21
eGFR, mL/min/1.73 m2	84 ± 17	97 ± 19	94 ± 13
Prevalent CVD, N (%)	393 (5)	241 (3)	3563 (5)
Prevalent heart failure, N (%)	68 (1)	48 (1)	363 (1)
Diabetes, N (%)	832 (12)	499 (7)	4032 (6)
Lipid-lowering medications, N (%)	2,641 (38)	1467 (21)	12874 (19)
Hypertension medications, N (%)	2,815 (41)	1711 (24)	14798 (22)
Diabetes medications, N (%)	632 (9)	340 (5)	771 (1)
Current smoking, N (%)	389 (6)	845 (12)	6818 (10)
PREVENT 10-year CVD risk, %	7.5 [3.1, 15.7]	2.5 [0.8, 8.1]	6.2 [3.2, 10.2]
PREVENT-HF 10-year HF risk, %	4.5 [1.6, 10.8]	1.1 [0.3, 4.5]	3.0 [1.3, 5.4]

Values represent mean ± SD, 50^th [25^th, 75^th] percentile, or N (%). HDL, high density lipoprotein; eGFR, estimated glomerular filtration rate; CVD, cardiovascular disease; HF, heart failure.

To determine whether the relations between hemodynamic measures and CVD events differed in younger versus older participants (dichotomized at ≥65 years of age), men versus women, or in participants without versus with obese BMI (BMI≥30 kg/m²), we included interaction terms for these variables in models that adjusted for other risk factors. All analyses were performed with SAS version 9.4, SPSS version 28, and R version 4.5.0. A 2-sided P<0.05 was considered statistically significant.

Results

Participants.

Characteristics of the FHS development set (N=6916, 56% women) at the first of up to 2 development visits (visits 2 and 3), the FHS internal test set (N=7018, 54% women) at visit 1, and the UKB external test sample (N=67986, 53% women) are summarized in Table 1. At visit 1, the FHS internal test sample had a broad age distribution (19 to 91 years) centered at midlife (51±15 years). Participants were older at the later development visits. Moderate proportions of participants were treated for hypertension and lipid abnormalities, and a modest proportion was treated for diabetes, with higher proportions of all treatments at the development visits (Table 1). Smoking prevalence was lower at the later development visits. The UKB external test sample comprises middle-aged participants with similar prevalences of hypertension and lipid treatment as the FHS internal test set. Diabetes treatment in UKB represents treatment with insulin only, so the proportion of participants treated for diabetes is lower compared to FHS.

Heart Age.

Measured Smax and Emax in FHS development and internal test sets, respectively, were Smax: 8.8±1.8 and 9.1±1.8 cm/s and Emax: 9.8±2.6 and 11.2±2.9 cm/s. AI-HA predictions are summarized in Table 1. Relations between chronological age and AI-HA in the FHS internal test set and the UKB external test sample are illustrated in Figure 1. While AI-HA and chronological age were highly correlated, there was considerable scatter around the line of identity. The distribution of UKB values around the line of identity resembled the pattern seen in FHS, although with a truncated age range (Figure 1). The correlation between chronological age and AI-HA was lower for UKB, which is consistent with the truncated age distribution in the UKB sample (40–70, range 30 years) as compared to the FHS sample (19–91, range 72 years).

Figure 1.

Scatterplot and density plots of chronological age and artificial intelligence heart age (AI-HA). FHS symbols and density plots are blue and UKB are orange. UKB points were randomly decimated by 10-fold to ensure that the density of FHS and UKB points was comparable. FHS, Framingham Heart Study; UKB, UK Biobank.

Feature maps.

Feature sensitivity maps for transformed radial (FHS, Figure 2A) or native PPG (UKB, Figure 2B) waveforms are presented in Figure 2. Differences in the ensemble averaged waveforms are evident, with a discrete early systolic peak in the youngest AI-HA groups (I) as compared to a late systolic peak in the oldest AI-HA groups (IV) in both cohorts (Figure 2). Correlations of predicted Emax and Smax with model features created by the first layer of the corresponding CNNs (Emax, Smax) are presented as heatmaps in the lower panels of Figure 2A (FHS) and Figure 2B (UKB). The feature maps share considerable similarities, although with modest variation in intensity and numbers of filters in the various clusters (Figure 2).

Figure 2.

AI-HA model component sensitivity maps derived from (A) PPG-transformed radial tonometry waveforms in FHS or (B) native PPG waveforms in UKB. Signal-averaged radial artery pressure waveforms (P), with first (dP) and second (d2P) derivatives, ensemble averaged by predicted AI-HA quartile groups, are in the top panel and sensitivity heatmaps of layer 1 features produced by the Emax and Smax convolutional neural network models are in the second and third panels. Sensitivities were computed by assessing the correlation of individual features output by the first layer of the convolutional neural networks (Emax or Smax) with the respective predicted values. Heatmaps have relative time along the horizontal axis and filter channels along the vertical axis. Vertical reference lines are placed at the 5 characteristic peaks and troughs (labeled a-e) of the averaged d2P waveforms. Peaks “a” and “e” represent the foot and dicrotic notch, respectively, whereas peak “c” is at the approximate mid-systolic inflection point, which represents return of the global reflected pressure wave. Localized temporal regions of marked sensitivity to Emax and Smax are evident as brighter warm colors for positive correlation (better predicted LV function) with predicted Emax or Smax and brighter cool colors for negative correlation (worse function). AI-HA, artificial intelligence heart age, PPG, photoplethysmography; FHS, Framingham Heart Study; UKB, UK Biobank;

AI-HA and clinical events.

Association of AI-HA with incident clinical events in the FHS internal test set and the UKB external test set are presented in Table 2. Effect sizes are presented per 1SD of AI-HA (20 years) in the FHS internal test sample (Table 1). In FHS participants, there were 148 HF (2.2 per 1000 person-years) and 331 CVD (5.1 per 1000 person-years) events during up to 10 years of follow up. In models adjusted for PREVENT risk factors, AI-HA was associated with incident HF (HR=2.09, CI: 1.64, 2.68) and incident CVD (HR=1.52, CI: 1.28, 1.81) (Table 2). In UKB external test data participants, there were 1408 HF (2.1 per 1000 person-years), and 2709 CVD (4.3 per 1000 person-years) events during up to 10 years of follow-up. In PREVENT risk factor-adjusted models, AI-HA was associated with incident HF (HR=1.23, CI: 1.13, 1.33) and CVD (HR=1.22, CI: 1.15, 1.30) events (Table 2). Addition of AI-HA to models that included calendar age and sex (and cohort for FHS) (Model 1A) as well as individual PREVENT risk factors (Model 2A) improved model fit (Supplemental Table S1) and was associated with significant reclassification of risk for incident HF (continuous net reclassification in FHS: 0.24, CI: 0.12, 0.33; UKB: 0.09, CI 0.06, 0.12) and incident CVD (FHS: 0.13, CI: 0.07, 0.20; UKB: 0.08, CI 0.06, 0.09) (Supplemental Table S2). When AI-HA was added to models that included the individual log hazards from the PREVENT risk calculator, model fit (Supplemental Table S1), discrimination, and risk reclassification (Supplemental Table S3) were improved in both cohorts for CVD events. We observed significant discrimination improvement in the FHS internal test set for HF events, and significant reclassification in both cohorts.

Table 2.

Associations of AI-HeartAge with the incidence of cardiovascular disease events.

		Model 1		Model 2
	Events / At risk	HR (95% CI)	P value	HR (95% CI)	P value
Heart failure
FHS	148 / 6970	2.15 (1.69, 2.73)	<0.0001	2.09 (1.64, 2.68)	<0.0001
UKB	1408 / 67623	1.46 (1.35, 1.58)	<0.0001	1.23 (1.13, 1.33)	<0.0001
Cardiovascular Disease
FHS	331 / 6777	1.60 (1.36, 1.88)	<0.0001	1.52 (1.28, 1.81)	<0.0001
UKB	2709 / 64423	1.40 (1.32, 1.48)	<0.0001	1.22 (1.15, 1.30)	<0.0001

Model 1 adjusts for age and sex (and cohort in FHS only). Model 2 adds PREVENT risk factors to Model 1 (total cholesterol, high-density lipoprotein cholesterol, systolic blood pressure, body mass index, estimated glomerular filtration rate, current use of cigarettes, presence of diabetes mellitus, hypertension treatment, and lipid treatment). All hazard ratios (HR) expressed per 20 years higher heart age.

A summary of tests for effect modification by age, sex, and obesity is presented in Supplemental Figures S5 and S6. There were significant interactions for age (<65 vs. ≥65 years) in UKB for both HF and CVD event models (Supplemental Figure S6). AI-HA associations with incident HF and CVD events in UKB were stronger in younger participants (Supplemental Figure S6). Tests for effect modification by development set membership are presented in Supplemental Table S4. There was no evidence that development set membership modified the relation between AI-HA and HF or CVD events. In addition, in models that included only participants who were not included in the development set, AI-HA continued to predict events (Supplemental Table S5).

Cumulative incidences of HF and CVD by quartile groups of AI-HA are presented in Figure 3. The incidence of events was low in participants in the lowest AI-HA groups and was markedly higher in those in the highest AI-HA groups, with intermediate rates in the middle groups. Separation between groups was greater in the FHS internal test sample, which may be related to the much wider age range in FHS (19–91 years) as compared to UKB (40–70 years). Secondary cumulative incidence plots that include only FHS participants within the UKB age range (40–70 years) show similar separation between groups in FHS (Supplemental Figure S7) and UKB (Figure 3B). When Cox model 2 was fitted in this truncated age subset, AI-HA continued to predict incident HF (HR=2.41, CI: 1.59, 3.64, P<0.001) and CVD (HR=1.58, CI: 1.23, 2.02, P<0.001) events.

Figure 3.

Cumulative incidence plots for heart failure and cardiovascular disease quartiles of AI-HA determined using (A) PPG-transformed radial tonometry waveforms in the FHS internal test set or (B) native PPG waveforms in the UKB external test data set. AI-HA, artificial intelligence heart age; PPG, photoplethysmography; FHS, Framingham Heart Study; UKB, UK Biobank.

Sensitivity analyses.

To assess the effect on model predictions of transformation of the tonometry waveform to a PPG equivalent, we compared AI-HA predictions derived using each of the waveform types in FHS and found a high correlation (R²=0.86).

As an additional sensitivity analysis, we computed an echocardiographic heart age (Echo-HeartAge) using measured tissue Doppler Emax and Smax rather than predicted values and fitted Cox models using Echo-HeartAge entered separately and then together with AI-HA in the same model (Supplemental Table S6). In a model that adjusts for PREVENT risk factors, Echo-HeartAge was associated with risk of new-onset HF but not composite CVD (Supplemental Table S6). When AI-HA was added to the model, both Echo-HeartAge and AI-HA were associated with incident HF whereas only AI-HA was associated with incident CVD (Supplemental Table S6).

In an extended covariate model that adds measured Emax and Smax, mean arterial pressure, pulse pressure, and augmentation pressure to the PREVENT covariate model, AI-HA remains a predictor of new-onset HF and CVD events in the internal test set (Supplemental Table S7).

Discussion

We used a feature-agnostic AI model to estimate biological age of the heart (AI-HA) from time- and amplitude-normalized arterial pressure and PPG waveforms. Models were trained using a combination of arterial pressure and PPG-transformed waveforms and echocardiographic data from FHS participants. Predicted AI-HA was strongly related to risk for incident HF in FHS participants. Results were replicated in a random subset of UKB participants in whom peripheral PPG waveforms were acquired. In models that adjusted for continuous values of various well-known risk factors for HF and CVD, each 1SD increase in AI-HA was associated with a doubling of HF risk in FHS participants. Addition of AI-HA to models that included PREVENT risk factors substantially reclassified risk in FHS and UKB. Given the widespread availability of PPG sensors in pulse oximeters and various wearable devices, our findings suggest that AI-HA could provide an accessible and valuable tool for HF risk assessment.

To assess AI-HA, two component CNN models were trained with key measures of left ventricular systolic (Smax) and diastolic (Emax) function assessed by using expensive and time-consuming echocardiography. Results from each model were standardized based on distribution in the internal test set, averaged, and scaled to derive a composite estimate of biological age of the heart (AI-HA). Models were trained using uncalibrated arterial pressure waveforms, a readily assessed physical manifestation of ventricular-vascular interaction. Heatmaps of model sensitivity to first layer features extracted by the models demonstrate sensitivity to regions of the pressure waveform known to be related to ventricular-vascular interaction, such as the peak of the forward pressure wave, the inflection point, subsequent secondary pressure rise, and early diastolic pressure bounce associated with wave reflection, and pressure decay in mid-late diastole. Derivation of the models using echocardiographic measures that assess mechanical coupling between LV and aorta, and strong relations of the resulting AI-HA with clinical outcomes, underscores the importance of LV long axis function and related mechanical coupling between the aorta and LV.^11–13 Considering the limited input requirement (uncalibrated pressure waveform), AI-HA potentially provides a simple point-of-care test that could be utilized in the clinic, at home, and in low-resource settings to identify high-risk individuals and potentially monitor effects of existing and novel treatments thought to improve aortic stiffness and LV function.

The concept of presenting standard clinical risk factors as equivalent “biological heart age” to facilitate interpretability has been examined previously.^14,15 Alternative measures of biological age derived from electrocardiographic (ECG) recordings have been examined by numerous investigators—a trend accelerated by introduction of novel AI methods in recent years.^16–21 Additional measures of biological age have been based on plasma proteome, genome, or epigenome.^22–24 Our approach requires only a peripheral pressure waveform, which is sensitive to interactions between heart and aorta and easily obtained at low cost with minimal technical requirements. Since we trained the models to detect differences in LV systolic (Smax) and diastolic (Emax) function, which are known to change with advancing age and various other exposures, the AI-HA model captures a spectrum of ventricular-vascular interaction that accompanies chronological aging. AI-HA is strongly related to chronological age but provides considerable additional predictive power for CVD risk. Our approach could be implemented on a wide scale and is amenable to long-term monitoring in clinical and nonclinical settings.

The fundamental goal of proper coupling between the heart and aorta is to deliver an adequate mean flow to organs with limited pulsatile overhead for the heart and periphery. The aorta couples the pulsatile LV pump with the periphery and transforms highly pulsatile LV outflow into a much less pulsatile flow waveform in the periphery. Elastic fibers in the aorta are key to this transformation. These elastic fibers are created once, during early development through toddler age, and must last a lifetime.²⁵ As a result, the aorta represents a true internal biological clock that is effectively “wound” once in early life. The rate at which this clock ticks, meaning variability in the rate of elastic fiber degradation and aortic stiffening, depends on various fixed and modifiable factors, including genetics,²⁶ early life environment,²⁷ and lifestyle,²⁸ resulting in a broad spectrum of biological age at a given chronological age.

Mechanical coupling between the aorta and LV plays a key role in LV diastolic function. The aortic arch and apex of the LV move very little during systole; hence, apical displacement of the atrioventricular plane, due to LV long axis shortening, stretches the ascending aorta and stores energy in the elastic fibers.^11,29 That energy can be recovered as early diastolic recoil and higher Emax, which facilitates early diastolic filling.¹² Degradation of elastic fibers in the aorta increases stiffness, which increases the mechanical load on the LV long axis and impairs longitudinal shortening.¹³ The resulting loss of aortic stretch is associated with impaired early diastolic filling of the LV.³⁰ Strong relations of AI-HA, which is trained on measures of LV long axis systolic and diastolic function, with HF events supports the important contribution of LV-aorta mechanical coupling to LV structure and function.

The relative ease of measurement of AI-HA could enable long-term monitoring using PPG sensors in wearable devices to motivate and monitor lifestyle modifications. Additionally, individuals with a higher risk for CVD, despite low or intermediate risk factor burden, could be identified. Further work is needed to determine whether targeted lifestyle or other interventions in individuals with increased AI-HA will result in lower event rates.

Several strengths and limitations of our investigation should be considered. Limitations include racial and ethnic composition of our development, internal test, and external test samples largely comprised of white European ancestry. Additional work will be required to demonstrate utility in other racial and ethnic groups. Our AI-HA model uses deep CNNs with more than 6 million trainable parameters, resulting in considerable capacity to memorize input data. To prevent over-training, we implemented aggressive dropout on each convolution layer and utilized early stopping based on model loss in a validation split. In addition, we used separate random train-validation-test set splits during model development to assess accuracy of each run and determine the best model at the end of four training runs. We assessed clinical utility in FHS at a distinct internal test set of examination cycles from those used for development and externally tested results in UKB, thereby preventing predictions based on memorized, nonphysiologic waveform characteristics. Supplemental sensitivity analyses showed no interaction between AI-HA and participant inclusion in development visits. In addition, FHS is a family-based cohort, meaning that related individuals were included in the FHS development and internal test sets, which could artificially improve model performance in internal testing. To address this limitation, we performed pedigree-based cross-validation to ensure that related individuals were not included in train and test sets. In addition, we externally validated our findings in a European sample with minimal close familial relations to FHS participants. We cannot exclude the possibility of residual confounding by factors not included in our statistical models, though we have adjusted for confounders included in the American Heart Association PREVENT equations. Balancing these limitations are many strengths. We trained and validated models using non-invasive tonometry and PPG-transformed waveforms and echocardiographic measures of LV function that were routinely acquired over a series of repeated visits in a large, community-based sample that spanned a broad age range. We performed external model evaluation in a large, independent, community-based UKB sample using fingertip PPG waveforms. The results demonstrate that AI-HA is a potentially widely applicable predictor of risk for incident HF events, and that it provides additional prognostic information beyond the PREVENT equations.

Perspectives

Our study proposes a novel AI-based measure of left ventricular systolic and diastolic function that accurately predicts HF and CVD risk in community-based samples. Internal and external model evaluations suggest that AI-HA adds to discrimination and reclassification of risk beyond AHA PREVENT risk factors alone using only a brief, uncalibrated peripheral pressure waveform as input. AI-HA could potentially be monitored using PPG sensors in consumer wearables and medical devices to provide valuable insights to clinicians and patients, especially in low resource settings.

Supplementary Material

Supplemental Methods

Tables S1–S7

Figures S1–S5

Supplemental References^31–37

Novelty and Relevance.

What Is New?

We present a novel measure of heart and vascular function using as input only a brief peripheral pressure waveform from a PPG sensor.

What is Relevant?

In this observational cohort study including deep learning models trained on 6916 Framingham Heart Study participants, validated in 7018 Framingham participants, and replicated in 67986 UK Biobank participants, higher AI-HA was associated with higher 10-year risk for incident heart failure and total cardiovascular disease.

Clinical/Pathophysiological Implications?

AI-HA provides an easily assessed measure of heart and vascular health that improves discrimination and reclassifies risk for heart failure and cardiovascular disease events.

Abbreviations:

AI

Artificial intelligence

AI-HA

Artificial intelligence heart age

CVD

Cardiovascular disease

CNN

Convolutional neural network

FHS

Framingham Heart Study

UKB

UK Biobank

HF

Heart failure

LV

Left ventricle