This prognostic study evaluates the ability of an electrocardiography-based artificial intelligence risk estimator to predict incident hypertension and stratify risk for hypertension-associated adverse events.
Key Points
Question
Can an electrocardiography (ECG)–based artificial intelligence risk estimator for hypertension (AIRE-HTN) predict incident hypertension and stratify risk for incident hypertension-associated adverse events?
Findings
In this prognostic study including an ECG algorithm trained on 189 539 patients at Beth Israel Deaconess Medical Center and externally validated on 65 610 patients from UK Biobank, AIRE-HTN predicted incident hypertension and stratified risk for cardiovascular death, heart failure, myocardial infarction, ischemic stroke, and chronic kidney disease.
Meaning
Results suggest that AIRE-HTN can predict the development of hypertension and may identify at-risk patients for enhanced surveillance.
Abstract
Importance
Hypertension underpins significant global morbidity and mortality. Early lifestyle intervention and treatment are effective in reducing adverse outcomes. Artificial intelligence–enhanced electrocardiography (AI-ECG) has been shown to identify a broad spectrum of subclinical disease and may be useful for predicting incident hypertension.
Objective
To develop an AI-ECG risk estimator (AIRE) to predict incident hypertension (AIRE-HTN) and stratify risk for hypertension-associated adverse outcomes.
Design, Setting, and Participants
This was a development and external validation prognostic cohort study conducted at Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts, a secondary care setting. External validation was conducted in the UK Biobank (UKB), a UK-based volunteer cohort. AIRE-HTN was trained and tested to predict incident hypertension using routinely collected ECGs from patients at BIDMC between 2014 and 2023. The algorithm was then evaluated to risk stratify patients for hypertension- associated adverse outcomes and externally validated on UKB data between 2014 and 2022 for both incident hypertension and risk stratification
Main Outcomes and Measures
AIRE-HTN, which uses a residual convolutional neural network architecture with a discrete-time survival loss function, was trained to predict incident hypertension.
Results
AIRE-HTN was trained on 1 163 401 ECGs from 189 539 patients (mean [SD] age, 57.7 [18.7] years; 98 747 female [52.1%]) at BIDMC. A total of 19 423 BIDMC patients composed the test set and were evaluated for incident hypertension. From the UKB, AIRE-HTN was tested on 65 610 ECGs from same number of participants (mean [SD] age, 65.4 [7.9] years; 33 785 female [51.5%]). A total of 35 806 UKB patients were evaluated for incident hypertension. AIRE-HTN predicted incident hypertension (BIDMC: n = 6446 [33%] events; C index, 0.70; 95% CI, 0.69-0.71; UKB: n = 1532 [4%] events; C index, 0.70; 95% CI, 0.69-0.71). Performance was maintained in individuals without left ventricular hypertrophy and those with normal ECGs (C indices, 0.67-0.72). AIRE-HTN was significantly additive to existing clinical risk factors in predicting incident hypertension (continuous net reclassification index, BIDMC: 0.44; 95% CI, 0.33-0.53; UKB: 0.32; 95% CI, 0.23-0.37). In adjusted Cox models, AIRE-HTN score was an independent predictor of cardiovascular death (hazard ratio [HR] per standard deviation, 2.24; 95% CI, 1.67-3.00) and stratified risk for heart failure (HR, 2.60; 95% CI, 2.22-3.04), myocardial infarction (HR, 3.13; 95% CI, 2.55-3.83), ischemic stroke (HR, 1.23; 95% CI, 1.11-1.37), and chronic kidney disease (HR, 1.89; 95% CI, 1.68-2.12), beyond traditional risk factors.
Conclusions and Relevance
Results suggest that AIRE-HTN, an AI-ECG model, can predict incident hypertension and identify patients at risk of hypertension-related adverse events, beyond conventional clinical risk factors.
Introduction
Hypertension affects 1 in 3 adults worldwide and accounts for approximately 20% of all US deaths.1,2 As a primary risk factor for cardiovascular disease, hypertension underpins a significant portion of the global disease burden.3
Hypertension-mediated organ damage is an early indicator of inadequate blood pressure control, with multinational guidelines advocating more aggressive treatment strategies when detected.4 Left ventricular hypertrophy (LVH) is a form of hypertension-mediated cardiovascular damage and is an independent predictor for extracardiac damage.5 LVH has been detectable on electrocardiograms (ECGs) since the early 20th century.6 However, ECG LVH criteria have limited diagnostic utility, requiring further imaging to screen for hypertension-mediated cardiac damage.4
Artificial intelligence–enhanced electrocardiography (AI-ECG) offers the potential to detect subtle ECG changes, including features not appreciable to humans, offering insights beyond traditional ECG interpretation.7 A major advantage of AI approaches is the ability to extract features relevant to the specific task, without anchoring on prior beliefs.8,9
We recently described the AI-ECG risk estimation (AIRE) platform, capable of predicting mortality, future atherosclerotic cardiovascular disease, ventricular arrhythmia, and heart failure.8 In this prognostic study, we expand AIRE to identify individuals at risk of incident hypertension and stratify the risk of hypertension-related sequelae. We additionally explore the biological plausibility of the AI-ECG tool and perform explainability analyses in support of clinical application.
Methods
Ethical Approvals
This study complied with all relevant ethical regulations. The Beth Israel Deaconess Medical Center (BIDMC) cohort ethics review and approval was provided by the BIDMC Committee on Clinical Investigations, and a waiver of consent was granted due to the retrospective nature of the study. The UK Biobank (UKB) has approval from the North West Multi-Centre Research Ethics Committee as a research tissue bank, and informed consent was obtained from all participants. Further details are provided in the eMethods in Supplement 1. This study followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) reporting guidelines.
Cohorts
We studied 2 cohorts: (1) the BIDMC, Boston, Massachusetts, a secondary care cohort consisting of routinely collected health record data and (2) the UKB cohort, a longitudinal study of volunteers between 40 and 69 years of age on joining during 2006 to 2010. A subset of these UKB individuals, who lived near an imaging assessment center, were invited for a second follow-up visit and had an ECG performed. More information about these cohorts and blood pressure measurements can be found in the eMethods in Supplement 1.
In the BIDMC cohort, participants self-identified with the following races and ethnicities: Asian, Black, Hispanic, White, or other, which included multiracial/ethnic and unspecified. In the UKB cohort, participants self-identified with Asian, Black, White, or other, which included multiracial/ethnic and unspecified.
Model Development
We expanded the AIRE platform for prediction of incident hypertension (AIRE-HTN). Hypertension included both essential and secondary hypertension types. Hypertension and outcomes were defined using International Classification of Diseases, Ninth (ICD-9) and Tenth (ICD-10) diagnostic codes in the BIDMC dataset and algorithmically defined diseases (a combination of ICD codes, self-reported diagnoses, primary care records, and causes of death) in UKB.
We used a 2-stage process to develop AIRE-HTN, both stages use an end-to-end neural network. ECG preprocessing and dataset splits (eFigure 1 in Supplement 1) are described in the eMethods in Supplement 1.
We first developed an AI-ECG hypertension classification model using the BIDMC cohort as the derivation dataset. The model architecture was based on convolutional neural networks that incorporate residual blocks, with a final sigmoid layer.10 The model was trained to identify prevalent hypertension as a classification task.
Next, to predict incident hypertension, AIRE-HTN was further developed by adapting the final layer to accommodate a discrete-time survival loss function.11 This approach allows AIRE-HTN to account for both time to outcome (incident hypertension) and censorship (ie, loss to follow-up). The classification model previously described was used for the weight initialization of the AIRE-HTN model. The outputs of the AIRE-HTN model are termed the AIRE-HTN score, which was standardized and normalized. Higher values indicate higher risk of incident hypertension. The model output at 5 years was used for prediction of incident hypertension (irrespective of the duration of follow-up for that particular individual). All ECGs were used for model training, however, as prediction of incident hypertension is most relevant for outpatients, only outpatient ECGs were used in the BIDMC test set. Further details of hyperparameters and model training are in the eMethods in Supplement 1.
Explainability Analysis
To explore the ECG morphologies and standard ECG metrics linked to the AIRE-HTN score, we used 3 approaches.
First, a variational autoencoder (VAE) was trained on median ECG beats. Median beats were derived using the BRAVEHEART ECG analysis software (open source).12 The latent features derived from the VAE were then fed into a linear regression model, aiming to predict the output of AIRE-HTN, a continuous range between 0 and 1 that relates to incident hypertension risk. The VAE was used only for explainability analysis and not to create the AIRE-HTN model (which used an end-to-end neural network as described previously). We identified and visualized the top 3 most significant features, based on their t values, through a process known as latent feature traversal. Further details are available in the eMethods in Supplement 1.
Second, we calculated the median waveform from the 10 000 ECGs with the lowest and highest AIRE-HTN score to qualitatively explore the morphologies associated with risk.
Third, we performed univariable correlation between AIRE-HTN and ECG parameters. Using linear regression, AIRE-HTN was adjusted for age, age squared, height, weight, body surface area (BSA), and waist circumference.
LVH Definitions
LVH was defined in accordance with American Society of Echocardiography guidelines,13 as an anteroseptal or inferolateral wall thickness of greater than or equal to 0.9 cm in females and 1.0 cm in males. LVH ECG criteria were defined using the Sokolow-Lyon criteria, where the S wave in V1 plus the R wave in V5 or V6 were added, and if the sum was greater than 35 mm, LVH was present.14 A continuous sum of maximum voltages for S waves in V1/V2 and R waves in V5/V6 was also calculated. Echocardiographic LV mass was calculated using the linear method as previously described.13 Cardiac magnetic resonance imaging (CMR) is the criterion standard for identification of LVH.15 Therefore, in the UKB, CMR was used to identify individuals with normal LV mass based on previously described age- and sex-specific normal ranges.16 Normal ECG definition is described in the eMethods in the Supplement.
Statistical Analysis
Survival Analysis
AIRE-HTN score quartiles were defined using the distributions in the validation set for Kaplan-Meier curves. AIRE-HTN score quartiles were plotted, and statistical significance was assessed using the log-rank test. For incident hypertension analyses, individuals with a hypertension diagnosis at baseline (both cohorts) or within 30 days after the ECG (in BIDMC) were excluded. Individuals without follow-up data (ie, they were censored on the day of the ECG) were also excluded. Cox model analyses were used to predict outcomes from the first ECG per individual only. The AIRE-HTN score variable was standardized, therefore, hazard ratios (HRs) reflect 1-SD changes in AIRE-HTN score. Cox models were fit using the test dataset for BIDMC and UKB datasets. For prediction of incident hypertension, clinical covariates were age, sex, systolic blood pressure (SBP), diastolic blood pressure (DBP), smoking status, prevalent diabetes mellitus (DM), and ethnicity. For UKB analyses, body mass index (BMI) was additionally included. For prediction of adverse events, covariates were age, sex, SBP, DBP, smoking status, DM, hypertension, hyperlipidemia, and ethnicity. UKB analyses additionally included BMI and number of antihypertensives as covariates. Due to the smaller number of patients with BMI data available, BMI was not included in primary BIDMC analyses, but it was included in sensitivity analyses. Medication data were not available in the BIDMC cohort. Individuals were censored at the time of death, last in-person hospital contact (BIDMC), or the UKB national censoring dates. To measure the improvement in prediction performance gained by AIRE-HTN, continuous net reclassification index (NRI) was calculated. Causal mediation analysis was performed using the R package mediation (R Core Team).17 We calculated the proportion mediated using logistic regression for 5-year incident outcomes, adjusted for smoking status, DM, hyperlipidemia, and ethnicity. We investigated the proportion mediated by a diagnosis of hypertension, which could be at baseline or during follow-up (but prior to the adverse event). Statistical analyses were performed with R, version 4.2.0 (R Core Team) or Python, version 3.9 (Python Software Foundation).
Phenome-Wide Association Study
We performed phenome-wide association studies (PheWAS) to better understand the biology underlying AIRE-HTN scores. We used the UKB cohort that contains data from over 3000 phenotypes derived from patient measurements, surveys, and investigations. We also applied this approach to the individuals with echocardiograms within 60 days of the ECG in the BIDMC test set. Using linear regression, the residual of AIRE-HTN score was calculated, after adjusting for age, age squared, sex, height, weight, and BSA. In the UKB, waist circumference was additionally included as a covariate; these data were not available for the BIDMC cohort. Univariate correlation was then performed using the residual to investigate the association between AIRE-HTN scores and phenotypes or cardiac imaging parameters (CMR and echocardiography). Further methods are described in the eMethods in Supplement 1. All P values were 2 sided, and a P value <.05 was considered statistically significant. Study data were analyzed for the BIDMC cohort from 2014 to 2023 and 2014 to 2022 for UKB.
Results
AI-ECG and Prediction of Incident Hypertension
In the BIDMC cohort, 1 163 401 ECGs were available from 189 539 individuals (mean [SD] age, 57.7 [18.7] years; 98 747 female [52.1%]; 90 792 male [47.9%]). Participants self-identified with the following races and ethnicities: 8924 Asian (4.7%), 24 251 Black (12.8%), 10 169 Hispanic (5.4%), and 122 344 White (64.5%), and 23 851 other (12.6%). A total of 34 938 individuals (18.4%) died during follow-up (eTable 1 in Supplement 1). We used the first outpatient ECG in those without hypertension for evaluation of incident hypertension prediction (n = 19 423; mean [SD] follow-up, 6.8 [5.6] years; 6446 events [33.2%]). AIRE-HTN predicted incident hypertension with C index 0.70 (95% CI, 0.69-0.71) with risk quartiles displayed in Figure 1 and eTable 2 in Supplement 1. The high-risk quartile had a 4-fold greater risk of incident hypertension after adjustment for age and sex (HR, 4.02; 95% CI, 3.65-4.43; P <.001). Performance was maintained in males and females and across major ethnic groups (eFigure 2 and eTable 3 in Supplement 1). In the BIDMC cohort, we performed sensitivity analysis in individuals with normal ventricular wall thickness on echocardiography (C index, 0.69; 95% CI, 0.67-0.72), absence of LVH by ECG criteria (C index, 0.70; 95% CI, 0.69-0.71), cardiologist-reported normal ECGs (C index, 0.72; 95% CI, 0.68-0.75), and normal baseline BP (<120/80 mm Hg, C-index, 0.73; 95% CI, 0.72-0.74). In the UKB, AIRE-HTN was evaluated on 65 610 ECGs from same number of participants (mean [SD] age, 65.4 [7.9] years; 33 785 female [51.5%]; 31 825 male [48.5%]). Participants self-identified with the following races and ethnicities: 980 Asian (1.5%), 519 Black (0.8%), 63 151 White (96.3%), and 960 other (1.5%). A total of 35 806 participants without prevalent hypertension were evaluated for incident hypertension. Performance was maintained in the external validation UKB cohort (C index, 0.70; 95% CI, 0.69-0.71; n = 35 806; mean [SD] follow up, 4.0 [1.6] years; 1532 events [4.3%]). UKB sensitivity analysis was performed in individuals with normal LV mass by CMR (C index, 0.70; 95% CI, 0.69-0.72) and normal baseline BP (<120/80 mm Hg) with no antihypertensive medications (C index, 0.71; 95% CI, 0.61-0.80).
Figure 1. Artificial Intelligence–Enhanced Electrocardiography Risk Estimation Platform for Prediction of Incident Hypertension (AIRE-HTN).
AIRE-HTN stratified risk of incident hypertension in the Beth Israel Deaconess Medical Center (BIDMC) (A) and UK Biobank (UKB) (B) cohorts. Kaplan-Meier curves show cumulative probabilities of hypertension for the 4 quartiles of risk defined by AIRE-HTN predictions using a single ECG.
In the UKB, we found the AIRE-HTN score was associated with the number of medications prescribed at follow-up (eFigure 3 in Supplement 1). In a subset of individuals with protocolized follow-up and no diagnosis of hypertension or prescription of antihypertensives at baseline, AIRE-HTN predicted future antihypertensive prescription at the time of follow-up (C index, 0.71; 95% CI, 0.64-0.77). We additionally found AIRE-HTN was superior to continuous measures of LVH (including ECG voltage criteria and LV mass) for the prediction of incident hypertension (eResults in Supplement 1).
We then compared AIRE-HTN score with existing clinical characteristics associated with hypertension; age, sex, SBP, DBP, smoking status, prevalent DM, and ethnicity. For UKB analyses, BMI was additionally included. We found AIRE-HTN was significantly additive to existing clinical markers in predicting incident hypertension (Figure 2 and eFigure 4 and eTable 4 in Supplement 1), C index with AIRE-HTN (0.75; 95% CI, 0.73-0.76) vs without AIRE-HTN (0.73; 95% CI, 0.72-0.75; P <.001). AIRE-HTN score provided additive predictive value as measured by the continuous NRI (BIDMC: 0.44; 95% CI, 0.33-0.53; UKB: 0.32; 95% CI, 0.23-0.37).
Figure 2. Cox Models for Prediction of Incident Hypertension.
Hypertension prediction C index results comparing the artificial intelligence–enhanced electrocardiography risk estimation platform for prediction of incident hypertension (AIRE-HTN) with clinical risk prediction methods for the prediction of incident hypertension. AIRE-HTN-Cox includes AIRE-HTN, age, sex, and ECG parameters. Hypertension risk factors include systolic blood pressure (SBP), diastolic blood pressure (DBP), smoking status, prevalent diabetes, and ethnicity.
AIRE-HTN and Adverse Events
We assessed the association of the AIRE-HTN score with hypertension-related adverse outcomes, after adjusting for clinical covariates (listed in the eMethods in Supplement 1). We found that the AIRE-HTN score was an independent predictor of cardiovascular death (HR per SD, 2.24; 95% CI, 1.67-3.00), heart failure (HR, 2.60; 95% CI, 2.22-3.04), myocardial infarction (HR, 3.13; 95% CI, 2.55-3.83), ischemic stroke (HR, 1.23; 95% CI, 1.11-1.37), and chronic kidney disease (HR, 1.89; 95% CI, 1.68-2.12), with all P values <.001 in BIDMC outpatients. These findings were broadly consistent regardless of hypertension status and across cohorts (Figure 3 and eFigure 5 and eTables 5-6 in Supplement 1).
Figure 3. Artificial Intelligence–Enhanced Electrocardiography Risk Estimation Platform for Prediction of Incident Hypertension (AIRE-HTN) Score and Hypertension-Related Adverse Outcomes.
In adjusted Cox models, AIRE-HTN score is an independent predictor of hypertension-related adverse outcomes in individuals without existing cardiovascular/kidney disease (A and C) and without existing cardiovascular/kidney disease but with hypertension (B and D). Covariates for Beth Israel Deaconess Medical Center (BIDMC) analysis: age, sex, systolic blood pressure, diastolic blood pressure, smoking status, prevalent diabetes, prevalent hypertension, prevalent hyperlipidemia, and ethnicity. UK Biobank (UKB) analyses additionally included body mass index and number of antihypertensives as covariates. Hazard ratio (HR) refers to 1-SD increase of AIRE-HTN score.
We subsequently explored if AIRE-HTN score was associated with adverse outcomes mediated through a diagnosis of hypertension. We found AIRE-HTN score was associated with cardiovascular death with partial mediation through a diagnosis of hypertension (proportion mediated, 11% [95% CI, 4%-30%; P <.001] to 35% [95% CI, 17%-167%; P <.001]) depending on cohort and subgroup (eTable 7 in Supplement 1). We found varying degrees of mediation for the other adverse events, from no significant mediation for hemorrhagic stroke to 57% for CKD (95% CI, 31%-110%; P <.001).
Explainable ECG Morphologies and AIRE-HTN Scores
To explore the ECG morphological changes responsible for the AIRE-HTN score, we used a VAE to visualize the features most associated with the AIRE-HTN score. Figure 4A shows the 3 latent features most highly associated with AIRE-HTN scores (t values = 325, 177, and 174, respectively), including intraventricular conduction delay, QRS amplitude, and T wave variation in lead II as most important variation between high- and low-risk ECGs.
Figure 4. Explainability Analyses.
A, A variational autoencoder was used to identify the most important morphological features in artificial intelligence–enhanced electrocardiography risk estimation platform for prediction of incident hypertension (AIRE-HTN) score, each subpanel shows 1 of 3 latent features. B, Mean (SD) (shaded region) electrocardiography (ECG) waveforms for the 10000 highest and lowest AIRE-HTN score from the Beth Israel Deaconess Medical Center (BIDMC) test set.
Next using median beats, we plotted the mean representations of the 10000 ECGs with the highest and lowest AIRE-HTN scores (Figure 4B). We found delayed R to S transition from V3 in low-risk groups to V5 in high-risk groups, combined with much greater heterogeneity in QRS complex and T waves in high-risk groups.
Finally, we assessed the Pearson correlation of established ECG parameters with AIRE-HTN scores (eFigure 6 in Supplement 1). These 3 methods highlight QRS voltage (r = 0.22; P <.001), QRS duration (r = 0.05; P <.001), R wave progression and axis (r = 0.13; P <.001), PR interval (r = 0.15; P <.001), and T wave morphology and QT interval (r = 0.08; P <.001) as important factors in the derivation of AIRE-HTN scores.
Biological Exploration: Phenotypic Associations of the AIRE-HTN Score
To investigate the biological associations with AIRE-HTN score and support model credibility, we performed a PheWAS in the UKB. First, we examined the correlations of AIRE-HTN score with age, sex, SBP, and DBP and, in particular, found a modest correlation with age (r = 0.40; P <.001) (eResults in Supplement 1). We then correlated AIRE-HTN score with over 3000 biological and clinical variables in the UKB, with covariates including age and age squared. We investigated associations with echocardiographic parameters in the BIDMC cohort. A Manhattan plot depicts the significant associations in UKB (eFigure 7 in Supplement 1). In particular, echocardiographic and CMR parameters included independent associations with LV wall thickness and mass (r = 0.16; P <.001), measures of diastolic function and filling pressure (r = 0.19; P <.001), and aortic dimensions (r = 0.13; P <.001) (Figure 5). Other significant positive correlations included carotid intima-media thickness (r = 0.07; P <.001), measured blood pressure (r = 0.21; P <.001), arterial stiffness (r = 0.06; P <.001), and measures of adiposity (r = 0.07; P <.001). Significant negative correlations included peak heart rate (r = 0.07; P <.001) on exercise and birth weight (r = 0.07; P <.001).
Figure 5. Cardiac Magnetic Resonance (CMR) Imaging Associations of Artificial Intelligence–Enhanced Electrocardiography Risk Estimation Platform for Prediction of Incident Hypertension (AIRE-HTN) Score.
Univariate correlation between AIRE-HTN score and cardiac magnetic resonance imaging (MRI), UK Biobank (UKB) cohort (A) and echocardiographic, Beth Israel Deaconess Medical Center (BIDMC) cohort (B) parameters was performed. Variables with multiple points indicate results of multiple measurements at varying anatomical locations. Comparisons meeting significance after Bonferroni correction are shown. AV indicates aortic valve; E, E wave; E′, E wave prime; E/A, E wave to A wave ratio; echo, echocardiography; E/e′, E wave over e prime; circum, circumferential; inf, inferior; LA, left atrium; lat, lateral; LV, left ventricle; LVEDD, left ventricular end diastolic dimension; LVEF, left ventricular ejection fraction; LVESD, left ventricular end systolic dimension; LVOT, left ventricular outflow tract; MV peak A, mitral valve inflow velocity in late diastole due to atrial contraction; RA, right atrium; RV, right ventricle; TR, tricuspid regurgitation.
Discussion
In this article, we introduce AIRE-HTN, a biologically credible AI model for ECG-based prediction of incident hypertension and risk stratification for adverse events.
The AI-ECG signature for hypertension has been previously explored, focusing on estimating blood pressure from ECG morphology,18 diagnosing LVH19 and prevalent hypertension.20,21,22,23 Unlike these existing AI-ECG hypertension studies, AIRE-HTN leveraged large, unselected cohorts, was externally validated, and focused on prediction of incident hypertension and risk stratification. AIRE-HTN retained its predictive accuracy in the absence of electrical or structural LVH and in cardiologist-reported normal ECGs.
The ability of AIRE-HTN to predict incident hypertension was additive to traditional markers for both cohorts, whereas previous statistical models have shown LVH as the only ECG criterion predictive for incident hypertension.24 Importantly, AIRE-HTN was an independent predictor of cardiovascular death, heart failure, myocardial infarction, ischemic stroke, and chronic kidney disease in both populations. Using this model to predict incident hypertension and its link to cardiovascular outcomes could help identify at-risk patients for whom more active surveillance and lifestyle interventions25,26,27,28 might be recommended at an earlier stage in the life course.
Although the current approach to hypertension treatment relied on population-derived thresholds of BP measurement, alternative biomarkers such as AIRE-HTN risk scores might increase awareness of the continuous pathogenic vascular syndrome that is labeled hypertension. The consistent model performance in a group of individuals with a normal baseline BP and no LVH suggests that AIRE-HTN may be identifying the latent biology of hypertension pathogenesis and not just simply identifying individuals with milder hypertension at baseline.
Our explainability and biological plausibility analyses, designed to support the credibility of AIRE-HTN, highlighted VAE ECG features associated with higher AIRE-HTN scores and are consistent with the literature for human-derived features of hypertension, including LVH ECG criteria,29 resting heart rate,30 ECG strain,31 and QRS duration and QT interval.32 Importantly, correlations with ECG LVH criteria were relatively weak, demonstrating that AIRE-HTN predictions are not dependent on them. Associations with CMR and echocardiographic parameters align with expectations typical of hypertensive remodeling, including increased left ventricular mass, larger aortic dimensions, diminished LV strain, lower ejection fractions, and impaired diastolic function.
PheWAS identified biologically plausible associations, reinforcing the reliability of AIRE-HTN. Associations included established hypertension correlates, such as measures of adiposity, smoking status, and increased carotid intimal thickness.33 Negative associations with AIRE-HTN risk scores included birth weight34 and peak heart rate during exercise.
Limitations
The inherent limitations to epidemiological studies include the accuracy of coded outcomes within the clinical datasets. Hypertension within BIDMC was defined using ICD codes, which lack granularity and may not match contemporary guidelines. We were unable to validate our findings against the criterion standard of ambulatory monitoring. However, sensitivity analysis with baseline BP recordings and medication use, and external validation with algorithmically defined diseases, demonstrated consistent findings. Although AIRE-HTN has been internally and externally validated in a large number of individuals, its performance across diverse populations and clinical settings remains to be explored. The incident hypertension model shows only modest discriminative capability; however, it was significantly additive to traditional risk stratification methods and superior to commonly used clinical risk tools in other clinical scenarios.35 The UKB PheWAS results were drawn from a population of European ancestry and require validation in other populations. The UKB cohort is known to consist of a relatively healthy population36 and may not reflect the general population; in contrast, BIDMC is a relatively unhealthy hospital-based population. Importantly, AIRE-HTN was able to discriminate risk of hypertension in both cohorts.
Conclusions
In this prognostic study, we have developed AIRE-HTN, an AI-ECG model for prediction of incident hypertension and for risk stratification for hypertension-associated adverse events. AIRE-HTN was significantly additive to known clinical predictors of hypertension. Results of exploratory and phenotypic analyses suggest the biological plausibility of these findings. Enhanced predictability could influence surveillance programs and primordial prevention.
eMethods.
eResults.
eFigure 1. Schematic Showing Dataset Splits for Model Development and Validation
eFigure 2. AIRE-HTN, Incident Hypertension Prediction Performance Stratified by Gender and Ethnicity
eFigure 3. Violin Plot Showing the Association Between AIRE-HTN Score at Baseline and the Number of Prescribed Antihypertensives at the Follow-Up Visit in the UKB Cohort
eFigure 4. Cox Models for Prediction of Incident Hypertension, Including Blood Results
eFigure 5. Sensitivity Analysis for Figure 2 Now Including BMI as a Covariate in BIDMC Cohort
eFigure 6. AIRE-HTN Score Correlations With ECG Parameters
eFigure 7. UK Biobank Phenome-Wide Association Study
eTable 1. Dataset Demographics
eTable 2. Cohort Demographics
eTable 3. Summary Table of Sex- and Ethnicity-Specific AIRE-HTN Performance for Hypertension Prediction in BIDMC Test Set
eTable 4. Summary Table of Cox Model Results Comparing AIRE-HTN to Clinical Risk Prediction Models in the BIDMC Test Set
eTable 5. Continuous Net Reclassification Index for the Addition of AIRE-HTN Score to a Baseline Model of Age, Sex, SBP, DBP, Smoking Status, Prevalent DM, Prevalent Hypertension, Prevalent Hyperlipidemia, and Ethnicity
eTable 6. Hypertension-Related Adverse Outcomes, Number of Events in Each Analysis
eTable 7. Mediation Analysis
eReferences
Data Sharing Statement.
References
- 1.Centers for Disease Control and Prevention . National Center for Health Statistics mortality data on CDC WONDER. Accessed February 2, 2023. https://wonder.cdc.gov/mcd.html
- 2.World Health Organization . Global report on hypertension: the race against a silent killer. Accessed February 2, 2023. https://www.who.int/publications/i/item/9789240081062
- 3.Carey RM, Muntner P, Bosworth HB, Whelton PK. Prevention and control of hypertension: JACC health promotion series. J Am Coll Cardiol. 2018;72(11):1278-1293. doi: 10.1016/j.jacc.2018.07.008 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Williams B, Mancia G, Spiering W, et al. ; ESC Scientific Document Group . 2018 ESC/ESH Guidelines for the management of arterial hypertension. Eur Heart J. 2018;39(33):3021-3104. doi: 10.1093/eurheartj/ehy339 [DOI] [PubMed] [Google Scholar]
- 5.Verdecchia P, Schillaci G, Borgioni C, et al. Prognostic significance of serial changes in left ventricular mass in essential hypertension. Circulation. 1998;97:48-54. doi: 10.1161/01.cir.97.1.48 [DOI] [PubMed] [Google Scholar]
- 6.Einthoven W. Galvanometrische registratie van het menschelijk electrocardiogram. In: Herinneringsbudendedl Professor S.S. Rosenstein. Eduard Ijdo; 1902:101-107. [Google Scholar]
- 7.Barker J, Li X, Kotb A, et al. Artificial intelligence for ventricular arrhythmia capability using ambulatory electrocardiograms. Eur Heart J Digit Health. 2024;5(3):384-388. doi: 10.1093/ehjdh/ztae004 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Sau A, Pastika L, Sieliwonczyk E, et al. Artificial intelligence–enabled electrocardiogram for mortality and cardiovascular risk estimation: a model development and validity study. Lancet Digit Health. 2024;6(11):e791-e802. doi: 10.1016/S2589-7500(24)00172-9 [DOI] [PubMed] [Google Scholar]
- 9.Sau A, Ribeiro AH, McGurk KA, et al. Prognostic significance and associations of neural network–derived electrocardiographic features. Circ Cardiovasc Qual Outcomes. Published online November 14, 2024. doi: 10.1161/CIRCOUTCOMES.123.010602 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Ribeiro AH, Ribeiro MH, Paixão GMM, et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat Commun. 2020;11(1):1760. doi: 10.1038/s41467-020-15432-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Gensheimer MF, Narasimhan B. A scalable discrete-time survival model for neural networks. PeerJ. 2019;7:e6257. doi: 10.7717/peerj.6257 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Stabenau HF, Waks JW. BRAVEHEART: open-source software for automated electrocardiographic and vectorcardiographic analysis. Comput Methods Programs Biomed. 2023;242:107798. doi: 10.1016/j.cmpb.2023.107798 [DOI] [PubMed] [Google Scholar]
- 13.Lang RM, Badano LP, Mor-Avi V, et al. Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. Eur Heart J Cardiovasc Imaging. 2015;16(3):233-270. doi: 10.1093/ehjci/jev014 [DOI] [PubMed] [Google Scholar]
- 14.Sokolow M, Lyon TP. The ventricular complex in left ventricular hypertrophy as obtained by unipolar precordial and limb leads. Am Heart J. 1949;37(2):161-186. doi: 10.1016/0002-8703(49)90562-1 [DOI] [PubMed] [Google Scholar]
- 15.Lenstrup M, Kjaergaard J, Petersen CL, Kjaer A, Hassager C. Evaluation of left ventricular mass measured by 3-D echocardiography using magnetic resonance imaging as gold standard. Scand J Clin Lab Invest. 2006;66(8):647-657. doi: 10.1080/00365510600892233 [DOI] [PubMed] [Google Scholar]
- 16.Maceira AM, Prasad SK, Khan M, Pennell DJ. Normalized left ventricular systolic and diastolic function by steady state free precession cardiovascular magnetic resonance. J Cardiovasc Magn Reson. 2006;8(3):417-426. doi: 10.1080/10976640600572889 [DOI] [PubMed] [Google Scholar]
- 17.Tingley D, Yamamoto T, Hirose K, Imai K, Keele L. Mediation: R package for causal mediation analysis. J Stat Softw. 2014;59(5):1-38. doi: 10.18637/jss.v059.i0526917999 [DOI] [Google Scholar]
- 18.Nour M, Kandaz D, Ucar MK, Polat K, Alhudhaif A. Machine learning and electrocardiography signal-based minimum calculation time detection for blood pressure detection. Comput Math Methods Med. 2022;2022:5714454. doi: 10.1155/2022/5714454 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Haimovich JS, Diamant N, Khurshid S, et al. Artificial intelligence-enabled classification of hypertrophic heart diseases using electrocardiograms. Cardiovasc Digit Health J. 2023;4(2):48-59. doi: 10.1016/j.cvdhj.2023.03.001 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Chen C, Zhao HY, Zheng SH, et al. Interpretable hybrid model for an automated patientwise categorization of hypertensive and normotensive electrocardiogram signals. Comput Methods Programs Biomed Update. 2023;3:100097. doi: 10.1016/j.cmpbup.2023.100097 [DOI] [Google Scholar]
- 21.Lu L, Zhu T, Ribeiro AH, et al. Decoding 2.3 million ECGs: interpretable deep learning for advancing cardiovascular diagnosis and mortality risk stratification. Eur Heart J Digit Health. 2024;5(3):247-259. doi: 10.1093/ehjdh/ztae014 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Soh DCK, Ng EYK, Jahmunah V, Oh SL, San TR, Acharya UR. A computational intelligence tool for the detection of hypertension using empirical mode decomposition. Comput Biol Med. 2020;118:103630. doi: 10.1016/j.compbiomed.2020.103630 [DOI] [PubMed] [Google Scholar]
- 23.Angelaki E, Barmparis GD, Kochiadakis G, et al. Artificial intelligence-based opportunistic screening for the detection of arterial hypertension through ECG signals. J Hypertens. 2022;40(12):2494-2501. doi: 10.1097/HJH.0000000000003286 [DOI] [PubMed] [Google Scholar]
- 24.Takase H, Sugiura T, Murai S, Yamashita S, Ohte N, Dohi Y. Use of electrocardiography to predict future development of hypertension in the general population. Medicine (Baltimore). 2016;95(17):e3483. doi: 10.1097/MD.0000000000003483 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Blood Pressure-Lowering Treatment Trialists’ Collaboration . Blood pressure-lowering treatment based on cardiovascular risk: a meta-analysis of individual patient data. Lancet. 2014;384(9943):591-598. doi: 10.1016/S0140-6736(14)61212-5 [DOI] [PubMed] [Google Scholar]
- 26.Uhlig K, Patel K, Ip S, Kitsios GD, Balk EM. Self-measured blood pressure monitoring in the management of hypertension: a systematic review and meta-analysis. Ann Intern Med. 2013;159(3):185-194. doi: 10.7326/0003-4819-159-3-201308060-00008 [DOI] [PubMed] [Google Scholar]
- 27.He FJ, Li J, Macgregor GA. Effect of longer term modest salt reduction on blood pressure: Cochrane systematic review and meta-analysis of randomized trials. BMJ. 2013;346:f1325. doi: 10.1136/bmj.f1325 [DOI] [PubMed] [Google Scholar]
- 28.Whelton SP, Chin A, Xin X, He J. Effect of aerobic exercise on blood pressure: a meta-analysis of randomized, controlled trials. Ann Intern Med. 2002;136(7):493-503. doi: 10.7326/0003-4819-136-7-200204020-00006 [DOI] [PubMed] [Google Scholar]
- 29.Brown DW, Giles WH, Croft JB. Left ventricular hypertrophy as a predictor of coronary heart disease mortality and the effect of hypertension. Am Heart J. 2000;140(6):848-856. doi: 10.1067/mhj.2000.111112 [DOI] [PubMed] [Google Scholar]
- 30.Palatini P, Thijs L, Staessen JA, et al. ; Systolic Hypertension in Europe (Syst-Eur) Trial Investigators . Predictive value of clinic and ambulatory heart rate for mortality in elderly subjects with systolic hypertension. Arch Intern Med. 2002;162(20):2313-2321. doi: 10.1001/archinte.162.20.2313 [DOI] [PubMed] [Google Scholar]
- 31.Okin PM, Devereux RB, Nieminen MS, et al. ; LIFE Study Investigators . Electrocardiographic strain pattern and prediction of cardiovascular morbidity and mortality in hypertensive patients. Hypertension. 2004;44(1):48-54. doi: 10.1161/01.HYP.0000132556.91792.6a [DOI] [PubMed] [Google Scholar]
- 32.Oikarinen L, Nieminen MS, Viitasalo M, et al. ; LIFE Study Investigators . QRS duration and QT interval predict mortality in hypertensive patients with left ventricular hypertrophy: the Losartan Intervention for Endpoint Reduction in Hypertension Study. Hypertension. 2004;43(5):1029-1034. doi: 10.1161/01.HYP.0000125230.46080.c6 [DOI] [PubMed] [Google Scholar]
- 33.Khutan H, Aggarwal S, Kajal KS, Garg R, Kaur R, Kaur A. Study of carotid intimal medial thickness in essential hypertension with or without left ventricular hypertrophy. Ann Afr Med. 2017;16(4):192-195. doi: 10.4103/aam.aam_9_17 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Bruno RM, Faconti L, Taddei S, Ghiadoni L. Birth weight and arterial hypertension. Curr Opin Cardiol. 2015;30(4):398-402. doi: 10.1097/HCO.0000000000000180 [DOI] [PubMed] [Google Scholar]
- 35.Lip GY, Nieuwlaat R, Pisters R, Lane DA, Crijns HJ. Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: the Euro Heart Survey on atrial fibrillation. Chest. 2010;137(2):263-272. doi: 10.1378/chest.09-1584 [DOI] [PubMed] [Google Scholar]
- 36.Fry A, Littlejohns TJ, Sudlow C, et al. Comparison of sociodemographic and health-related characteristics of UK Biobank participants with those of the general population. Am J Epidemiol. 2017;186(9):1026-1034. doi: 10.1093/aje/kwx246 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
eMethods.
eResults.
eFigure 1. Schematic Showing Dataset Splits for Model Development and Validation
eFigure 2. AIRE-HTN, Incident Hypertension Prediction Performance Stratified by Gender and Ethnicity
eFigure 3. Violin Plot Showing the Association Between AIRE-HTN Score at Baseline and the Number of Prescribed Antihypertensives at the Follow-Up Visit in the UKB Cohort
eFigure 4. Cox Models for Prediction of Incident Hypertension, Including Blood Results
eFigure 5. Sensitivity Analysis for Figure 2 Now Including BMI as a Covariate in BIDMC Cohort
eFigure 6. AIRE-HTN Score Correlations With ECG Parameters
eFigure 7. UK Biobank Phenome-Wide Association Study
eTable 1. Dataset Demographics
eTable 2. Cohort Demographics
eTable 3. Summary Table of Sex- and Ethnicity-Specific AIRE-HTN Performance for Hypertension Prediction in BIDMC Test Set
eTable 4. Summary Table of Cox Model Results Comparing AIRE-HTN to Clinical Risk Prediction Models in the BIDMC Test Set
eTable 5. Continuous Net Reclassification Index for the Addition of AIRE-HTN Score to a Baseline Model of Age, Sex, SBP, DBP, Smoking Status, Prevalent DM, Prevalent Hypertension, Prevalent Hyperlipidemia, and Ethnicity
eTable 6. Hypertension-Related Adverse Outcomes, Number of Events in Each Analysis
eTable 7. Mediation Analysis
eReferences
Data Sharing Statement.