Tempus wins breakthrough status for AI-based tool to detect atrial fibrillation



By Meg Bryant March 24, 2021

The U.S. FDA has granted breakthrough device designation to Tempus Inc. for its ECG Analysis Platform. Developed in collaboration with Geisinger, the artificial intelligence (AI)-powered platform helps clinicians identify patients at increased risk of developing atrial fibrillation (AF) or atrial flutter.

AF – an irregular, often rapid, heart rate – affects as many as 2.2 million Americans and is a leading cause of stroke, increasing an individual’s risk by four to six times, on average. However, detecting future risk of AF in asymptomatic patients has been challenging.

Fills an unmet need

“There are currently no available tools to help identify patients without current AF or AF history who are likely to experience future AF within a clinically actionable time frame and could derive the greatest benefit from additional cardiac monitoring within the goals of timely AF diagnosis, proactive stroke risk reduction and prevention of negative outcomes associated with a potential future stroke,” Joel Dudley, Tempus’ chief scientific officer, told BioWorld. “The [AI] algorithm is intended to be used by health care providers in combination with a patient’s medical history and clinical evaluation to inform clinical decision-making.”

Dudley stressed that the information should be interpreted only in conjunction with the patient’s initial ECG recordings, clinical history, symptoms and other diagnostic tests, and is not intended to be used alone to determine patient care.

The platform is intended for use with patients 40 and older who have not been diagnosed with AF or atrial flutter but are at increased risk for stroke, based on a CHA2DS2-VASc score of 4 or higher a common means of assessing stroke risk. The device analyzes the results of a 12-lead electrocardiogram (ECG) administered during routine care to provide insights into a patient’s future risk of AF or atrial flutter.

The algorithm is able to provide results almost immediately after the ECG data is fed into the system; however, real-world turnaround would depend on the how ECG is implemented by the health care provider, Dudley said. Moreover, the value of immediacy would depend on the clinical scenario. For example, the AF predictor could run in a batch mode on ECGs from a large population to flag individuals who require additional attention or follow-up. It could also be used within an episode of care to provide a future risk probability estimate for a single patient.

“Every year, hundreds of millions of ECGs are performed in the U.S. to detect cardiac abnormalities as part of routine clinical care,” Dudley said. “We are making ECGs smarter so that they can also identify the risk of future clinical events of interest, such as AF, thus enabling clinicians to act earlier in the course of disease and improve patient outcomes.”

Clinical evidence

In a recent study published in Circulation, researchers at Geisinger and Chicago-based Tempus demonstrated the ability of AI to predict new-onset AF. The team used 1.6 million resting 12-lean digital ECGs from 430,000 patients collected between 1984 and 2019 to train a deep neural network to predict patients with no history of AF who would develop it within the next 12 months. To simulate real-world use, they trained a separate model using ECGs from before 2010 and assessed its performance on ECGs from 2010 to 2014 that were linked to a stroke registry.

In term of predicting new-onset AF within one year of an ECG, the area under the receiver operating characteristic curve and area under the precision recall curve were 0.85 and 0.22, respectively. The hazard ratio for predicted high-risk vs low-risk groups over 30 years was 7.2 (95% CI, 6.9-7.6). When tested in the simulated real-world scenario, the model predicted new-onset AF at one year with 69% sensitivity and 81% specificity. The simulated deployment scenario identified patients at high risk for new-onset AF in 62% of patients who experienced an AF-related stroke within three years of the index ECG.

Brandon Fornwalt, chair of Geisinger’s department of translational data science and informatics, welcomed the FDA’s breakthrough device designation. “We showed in the Circulation study that nearly two-thirds of patients who have a stroke that is associated with atrial fibrillation might be able to be identified and treated earlier through using AI algorithms to identify high-risk patients,” he told BioWorld.

“This is ultimately about helping patients and fulfilling the promise of precision health by supporting clinical decision-making with additional patient-specific information, and we are excited that the FDA recognizes the importance of this work,” he said.

The breakthrough device designation opens the way for more interactions with the FDA throughout the regulatory pathway. Companies are also fast-tracked for reimbursement of their products once clearance or approval or granted.

Tempus has a number of ECG-based algorithms in the pipeline for predicting other cardiovascular events and will share details in peer-reviewed publications when the time is right, Dudley said.

“We are finding that AI-based analysis of ECGs appears to have promise of utility in an appreciable number of cardiovascular disease management scenarios,” he added. “We think we are just scratching the surface of what’s possible.”