We have written previously on this blog about the difficulty with atrial fibrillation (AF) screening – primarily because of a lack of cost-effective measures, screening is currently not recommended by the USPSTF. In this setting, a 2019 ground breaking study from the Mayo Clinic showed that artificial intelligence (AI) can make accurate predictions about the presence atrial fibrillation from 12-lead electrocardiograms (ECGs) taken in sinus rhythm. With the use of similar technology, we may now be entering a new area of AF detection.
Tempus AI, Inc., a Chicago based health technology company specializing in precision medicine, developed the ECG-AF algorithm, an advanced AI tool designed to identify patients at increased risk of atrial fibrillation (AF) or atrial flutter. This algorithm analyzes resting 12-lead ECGs to detect subtle patterns associated with the likelihood of developing AF within the next 12 months. In June 2024, the U.S. Food and Drug Administration (FDA) granted 510(k) clearance to ECG-AF, marking a significant milestone as the first AI-driven cardiovascular machine learning-based software cleared for AF risk prediction. (FDA Clearance)
Algorithm Development and Technical Requirements
Model and Training Data:
The ECG-AF algorithm processes resting 12-lead ECGs to extract subtle patterns indicative of AF risk. The model generates a raw risk score which then binarily categorizes patients into either an increased risk of developing AF or no increased risk based on a predefined risk threshold.
The ECG-AF algorithm was developed using machine learning techniques trained on 1.5 million ECGs from over 450,000 patients. This large dataset ensured a diverse representation of ECG patterns. For development, 80% of the ECGs were used for training and 20% for model tuning and validation, which is a typical division. The exact type of machine learning architecture has not been publicly disclosed, but it likely involves deep learning-based convolutional neural networks (CNNs) or recurrent neural networks (RNNs), which have historically been particularly effective in ECG pattern recognition.
Clinical Validation and Performance:
The algorithm underwent clinical performance validation in a retrospective observational study of 4,017 patients aged 65 and older from three geographically distinct clinical sites. The study reported:
- Sensitivity: 31%
- Specificity: 92%
- Positive predictive value (PPV): 19%
- Negative predictive value (NPV): 95%
While the sensitivity is modest, the high specificity ensures that flagged patients truly have an increased risk of AF, and while the baseline 1-year AF incidence rate in the study population was 6%, patients receiving a positive result showed an increased AF incidence of approximately 1-in-5 within the next year. This means the algorithm is conservative in flagging high-risk cases but minimizes false positives, making it useful for selectively identifying patients who may benefit from additional monitoring (i.e. can help us enhance our screening protocols).
Technical Requirements:
The system is designed without a dedicated user interface, instead integrating with other medical systems through standard communication protocols like APIs or file exchanges. It’s compatible only with ECG recordings collected using ‘wet’ Ag/AgCl electrodes with conductive gel/paste and works specifically with FDA-authorized 12-lead resting ECG machines from GE Medical Systems and Philips Medical Systems that use a 500 Hz sampling rate. The technology was validated on specific models including CSYS, MAC2K, MAC35, MAC55, MAC5K, PageWriter TC, and PageWriter Touch.
Intended Use and Patient Population
ECG-AF is designed for resting 12-lead ECG recordings collected at healthcare facilities from patients aged 65 and older who meet the following criteria:
- No prior history of atrial fibrillation/flutter
- No pacemaker or implantable cardioverter-defibrillator (ICD)
- No recent cardiac surgery (within the last 30 days)
The algorithm is not meant for continuous monitoring and should be used as a decision-support tool rather than a standalone diagnostic.
Clinical Adoption and Application
Beyond FDA clearance, the Centers for Medicare & Medicaid Services (CMS) has approved Medicare coverage for ECG-AF, effective January 1, 2025. This aligns with CMS policy to facilitate certain Software as a Service (SaaS) devices to be used in the Hospital Outpatient setting, and allows hospitals to receive reimbursement, facilitating broader clinical adoption. (CMS Decision)
Northwestern Medicine is the first major healthcare institution to integrate ECG-AF into its clinical workflows, aiming to enhance early detection and proactive intervention for at-risk patients. (Partnership Announcement)
Summary Thoughts
The ECG-AF algorithm represents a significant step forward in AI-assisted cardiovascular risk prediction. While it has limitations in sensitivity, its high specificity and Medicare coverage make it a valuable tool for selectively flagging patients for further AF evaluation. With institutions like Northwestern Medicine testing adoption, AI-driven ECG analysis could soon become a more standard part of cardiology care.
For further updates, be sure to subscribe to This week in Cardiovascular AI – an email newsletter for followers and subscribers that summarizes recent research in Cardiovascular AI.
Cardio Cypher 