206 Artificial intelligence–enabled electrocardiography for hypertension diagnosis

Joseph Barker, Arunashis Sau, Libor Pastika, Ewa Sieliwończyk, Konstantinos Patlatzoglou, Kathryn A. McGurk, Antônio H. Ribeiro, Nicholas S. Peters, Declan P. O’Regan, James S. Ware, Daniel B. Kramer, Jonathan W. Waks, Fu Siong Ng2024 · DOI

Introduction

Hypertension remains a primary driver of global morbidity and mortality, often termed the "silent killer" due to its asymptomatic progression. Clinical guidelines emphasize early intervention to prevent hypertension-mediated organ damage (HMOD), yet current diagnostic tools often rely on late-stage indicators. In standard clinical practice, the 12-lead electrocardiogram (ECG) is used to detect hypertension primarily through markers of Left Ventricular Hypertrophy (LVH). However, by the time LVH is detectable by human interpretation, significant cardiovascular damage has often already occurred.

This research introduces AIRE-HTN (AI Risk Estimator for hypertension), a deep-learning model designed to identify the subtle morphological signatures of hypertension from standard 12-lead ECGs, even when those signals appear "normal" to the human eye.

Methodology

The researchers developed a deep neural network based on a residual network (ResNet) architecture. Unlike traditional ECG analysis, which focuses on predefined intervals and amplitudes (such as the Sokolow-Lyon criteria), AIRE-HTN is feature-agnostic. It processes the raw 12-lead signal to identify complex, non-linear patterns associated with hypertensive states.

Data and Training

The model was derived from a massive dataset of 1,163,401 ECGs sourced from a secondary care setting in the United States.

  • Development/Internal Validation: 113,723 unique patients were used for training and testing, while 75,815 were reserved for internal validation.
  • External Validation: To ensure generalizability, the model was tested on the UK Biobank (UKB), a cohort of 65,610 participants. This external validation is crucial as the UKB represents a healthier "volunteer" population compared to the clinical population used for training.

Key Findings

State-of-the-Art Performance

AIRE-HTN demonstrated robust diagnostic accuracy across diverse populations. The model achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.77 in the US cohort and 0.71 in the UK Biobank.

Seeing Beyond Human Limits

The most significant breakthrough of AIRE-HTN is its ability to detect hypertension in the absence of traditional markers:

  • Absence of LVH: The model maintained an AUROC of 0.78 in patients without structural LVH (confirmed via echocardiography) and without electrophysiological LVH (Sokolow <3.5 mV).
  • "Normal" ECGs: In cases where cardiologists explicitly labeled the ECG as "normal," the model still achieved an AUROC of 0.78.

These results suggest that hypertension induces a distinct "morphological phenotype" in the heart's electrical activity that precedes visible structural changes like muscle thickening. This phenotype appears to be beyond the conventional human understanding of structural adaptation to high blood pressure.

Discussion and Future Implications

The success of AIRE-HTN challenges the conventional understanding of hypertension-mediated organ damage. It suggests that the heart undergoes subtle electrical remodeling long before traditional diagnostic thresholds are met.

Limitations and Future Work

While the model is highly effective at identifying existing (prevalent) hypertension, the study identifies two primary areas for future development:

  1. Incident Prediction: Future iterations will focus on whether the model can predict the future onset of hypertension in currently normotensive patients.
  2. Interpretability: As with many deep-learning models, the specific features AIRE-HTN uses to make its predictions remain a "black box." Further research is required to map these AI-identified features back to known physiological processes.

Conclusion

AIRE-HTN represents the first large-scale AI-ECG model optimized for hypertension diagnosis. By identifying hypertensive signatures that are invisible to human clinicians and independent of LVH, this tool offers a potential pathway for earlier intervention, remote disease monitoring, and more personalized cardiovascular risk assessment.

Summary updated

Abstract

<h3>Background</h3> Hypertension underpins significant global morbidity and mortality. Hypertension-mediated organ damage is an early indicator of serious cardiovascular sequelae, with multinational guidelines advocating aggressive treatment where detected. Artificial intelligence-enabled electrocardiogram (AI-ECG) offers novel strategies to facilitate personalised hypertensive risk assessments, early intervention, and remote disease monitoring in hypertension care. <h3>Purpose</h3> We aimed to develop and explore an AI-ECG model designed to diagnose hypertension from 12 lead ECGs. <h3>Methods</h3> We developed a deep neural network AI-ECG model, AI Risk Estimator for hypertension diagnosis (AIRE-HTN) from 12 lead ECGs. The model uses a residual network with sigmoid classification layer for prevalent hypertension diagnosis. It was derived from 1,163,401 ECGs from a real-world secondary care setting in the USA. Model development took 60% of ECGs (50/10% train/test split, n = 113,723 unique patients) and internally validated on 40% (n = 75,815 unique patients). The results were externally validated in the UK Biobank (UKB) volunteer cohort (n = 65,610 participants). <h3>Results</h3> AIRE-HTN demonstrates state-of-the-art performance in identifying existing hypertension (area under receiver operating characteristic curve (AUROC): USA cohort 0.77 (0.77–0.78), UKB: 0.71 (0.70–0.71)). AIRE-HTN works in the absence of both structural left ventricular hypertrophy (LVH) by echocardiography (AUROC 0.78 (0.77–0.80)), electrophysiological LVH (Sokolow &lt;3.5 mV, AUROC 0.78 (0.77–0.80)) and in those with cardiologist-reported normal ECGs (AUROC 0.78 (0.77–0.80). <h3>Conclusion</h3> We have developed AIRE-HTN, the first large scale AI-ECG model for hypertension diagnosis from standard 12 Lead ECGs. It operates effectively even in the absence of both structural and electrophysiological LVH, as well as in cases with cardiologist-reported normal ECGs. This suggests the morphological ECG phenotype the model derives as related to hypertension is beyond conventional human understanding of structural adaptation to hypertension. These findings pose an opportunity to rethink cardiovascular hypertension-mediated organ damage. Further work is planned for hypertension diagnosis prediction and further in-depth AIRE-HTN model exploration. <h3>Conflict of Interest</h3> None
BibTeX
@misc{barker2024artificial,
  author = {Joseph Barker and Arunashis Sau and Libor Pastika and Ewa Sieliwończyk and Konstantinos Patlatzoglou and Kathryn A. McGurk and Antônio H. Ribeiro and Antônio Luiz Pinho Ribeiro and Nicholas S. Peters and Declan P. O’Regan and James S. Ware and Daniel B. Kramer and Jonathan W. Waks and Fu Siong Ng},
  title = {{206 Artificial intelligence–enabled electrocardiography for hypertension diagnosis}},
  year = {2024},
  doi = {10.1136/heartjnl-2024-bcs.201},
}
Figure