Development and Multinational Validation of an Ensemble Deep Learning Algorithm for Detecting and Predicting...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Devices

Background:

• 12-lead ECG AI shows promise for detecting structural heart diseases (SHDs).
• Current AI-ECG tools have limitations in community-based screening.
• Single-lead ECGs from wearable devices offer potential for broader accessibility.

Purpose of the Study:

• To develop and validate a noise-resilient, single-lead AI-ECG algorithm for SHD detection.
• To assess the algorithm's ability to predict future SHD risk.
• To enable SHD screening and risk stratification using portable devices.

Main Methods:

• Developed ADAPT-HEART, a deep-learning algorithm using 266,740 lead I ECGs.
• Defined SHD as LVEF<40%, moderate/severe left-sided valvular disease, or severe LVH.
• Validated ADAPT-HEART in external hospital sites and the ELSA-Brasil cohort, and assessed predictive performance in UK Biobank.

Main Results:

• ADAPT-HEART demonstrated strong performance in detecting SHD (AUROC 0.879 in development, consistent in external validation).
• The algorithm showed good calibration for SHD detection.
• High ADAPT-HEART probability significantly increased future SHD risk (2.8- to 5.7-fold) in individuals without baseline SHD.

Conclusions:

• A novel AI model detects and predicts SHDs from noisy, single-lead ECGs.
• The algorithm is suitable for use with portable/wearable devices.
• This provides a scalable strategy for community-based SHD screening and risk stratification.