Hypertension Detection: HTN-AI Model for Cardiovascular Risk Stratification

Could a simple ECG improve hypertension detection? A recent article in Nature presented a hypertension detection model, HTN-AI, designed to identify hypertension and stratify the risk of cardiovascular disease (CVD) using 12-lead electrocardiogram (ECG) waveforms. The model was trained on over 750,000 ECGs from more than 100,000 patients and validated in a separate sample of over 56,000 patients. HTN-AI accurately identifies prevalent hypertension, predicts the short-term incidence of hypertension, and shows associations with elevated 24-hour ambulatory blood pressure. It also stratifies the risk of CVD outcomes such as mortality, heart failure, myocardial infarction, stroke, and aortic dissection.

Highlights from the Study

HTN-AI can identify hypertension using only 12-lead ECG waveforms, achieving an area under the receiver operating characteristic curve (AUROC) of 0.803 and 0.771 in internal and external validation samples, respectively. The model stratifies the risk of hypertension-associated CVD, including mortality, heart failure, myocardial infarction, stroke, and aortic dissection, regardless of baseline hypertension status or the presence of overt ECG abnormalities.

High HTN-AI risk corresponds with a higher mean 24-hour systolic blood pressure and greater odds of abnormal 24-hour ambulatory blood pressure monitoring results. HTN-AI shows superior or comparable discrimination to established clinical risk scores, such as the Pooled Cohort Equation, for predicting CVD outcomes. Standard ECG measurements, and saliency maps highlight regions of the ECG contributing to HTN-AI predictions, influencing the model’s predictions.

Why is this important?

Hypertension affects over 1 billion individuals worldwide and is a major modifiable risk factor for CVD. Office blood pressure measurements can be variable and may not accurately reflect ambulatory blood pressure, leading to potential underdiagnosis of ambulatory hypertension.

Guidelines recommend serial blood pressure measurements and ambulatory BP monitoring, but clinicians often do not rigorously implement these in clinical settings. Clinicians widely use ECGs as inexpensive diagnostic tools that provide valuable insights into cardiac health. Deep learning models have previously utilized ECGs to classify various cardiac conditions, including atrial fibrillation, left ventricular systolic dysfunction, and valvular disease.

Next Steps

The use of the hypertension detection model HTN-AI could reduce healthcare costs by identifying patients at high risk of CVD earlier, allowing for more targeted and efficient management of hypertension. This could lead to a reduction in the incidence of costly CVD outcomes. By accurately identifying patients with hypertension and stratifying their risk of CVD, HTN-AI can facilitate early detection and management of hypertension, potentially reducing the incidence of heart failure, myocardial infarction, stroke, and other cardiovascular events.

HTN-AI could be deployed in primary care settings to opportunistically screen for masked hypertension using 12-lead ECGs, prompting further diagnostic testing with ambulatory BP monitoring for those at high risk. This approach could enhance the detection and treatment of ambulatory hypertension, which is often underdiagnosed. Before clinical implementation, prospective validation of the hypertension detection model HTN-AI against gold-standard measures of hypertension, such as 24-hour ambulatory BP monitoring, is necessary to ensure its accuracy and reliability in diverse patient populations.