Tempus ECG-AF

Tempus ECG-AF is a machine learning-based software that analyzes 12-lead ECG recordings from specific asymptomatic patients aged 65 and older to identify signs indicative of an increased risk of developing atrial fibrillation or atrial flutter within the next 12 months. The technology processes standard ECG data, patient age, and sex through a trained algorithm to provide a risk notification, intended to support clinical decisions for further diagnostic follow-up rather than for direct diagnosis or treatment guidance.

The Tempus ECG-AF device is a software tool utilizing machine learning to analyze 12-lead resting electrocardiogram (ECG) recordings. It is intended for use in patients aged 65 years and older who do not have a pre-existing diagnosis of atrial fibrillation (AF) or atrial flutter, are without pacemakers or implantable cardioverter defibrillators, and have not had cardiac surgery within the preceding 30 days. The software's purpose is to detect subtle signs within these ECGs that are associated with an increased likelihood of the patient experiencing AF or atrial flutter within the next 12 months, thereby aiding healthcare providers in identifying individuals who may warrant further diagnostic evaluation for potential undiagnosed AF. The core of Tempus ECG-AF is a 'locked' machine-learning model, trained on a large dataset exceeding 1.5 million ECGs, which processes the patient's 12-lead ECG tracing along with their age and sex. This analysis generates an uncalibrated risk score, subsequently converted into a binary output—either 'increased risk' or 'no increased risk'—based on pre-defined thresholds, or an 'unclassifiable' result if data quality is insufficient. Clinical validation in a retrospective study of 4017 patients demonstrated a sensitivity of 31% and a specificity of 92% for identifying patients who would be diagnosed with AF within 12 months. The positive predictive value (PPV) was 19% and the negative predictive value (NPV) was 95% in a study population with a 6.0% one-year AF incidence, suggesting that a positive result identifies a patient group with a roughly threefold higher likelihood of developing AF compared to the baseline.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. Given the sensitivity of 31%, what proportion of patients who will develop AF within 12 months would be missed by this screening tool, and what are the clinical implications of these false negatives?; How does the model's performance vary across different ECG machine manufacturers and models beyond the specified GE and Philips systems, and with different electrode types or sampling rates not explicitly validated?; What specific ECG features or subtle patterns does the machine learning algorithm primarily rely on to identify an increased risk of future AF, and are these features novel or extensions of known subclinical ECG markers?; Considering the training dataset was predominantly White (96%), and the validation set was 82% White, what is the anticipated performance in more racially diverse populations, and are there plans for specific validation in these groups?; Since repeated testing performance has not been evaluated, what is the recommended clinical approach or testing interval if a patient initially receives a 'no increased risk' result but remains in a high-risk demographic for AF?