Abstract 4147060: Artificial Intelligence-based improvement of smartwatch detection of AF

Introduction: Cardiac arrhythmias, such as atrial fibrillation (AF), are a common health problem. Standard 12-lead electrocardiograms (ECGs) and Holter monitors may not detect them due to their brief recording duration. There is increasing consumer use and demand for Smartwatch (SW), and herein we investigate the use of an AI algorithm (Philips Cardiologs) on the Apple Watch to improve detection of AF. Hypothesis/Objective: Evaluate the performance of the AI algorithm on a SW in detecting AF in comparison to physician interpretation on the 12-lead ECG among patients undergoing cardioversion. Methods: 220 patients from 3 different US hospitals undergoing a cardioversion were included in the study. A SW single lead and a standard 12-lead ECGs were obtained before and after the procedure. 12-lead ECGs were annotated and adjudicated by three expert electrophysiologists as 91.3% AF before the procedure, 4.1% after the procedure and 49.3% of total ECGs were AF. The study design allowed us to reach a high AF:Sinus Rhythm (SR) ratio for the total ECGs dataset. The Apple ECG App 2.0 and the AI algorithm both processed the SW-ECGs to classify them as AF, SR or inconclusive and their performances were compared. Results: Among the 383 simultaneous combinations of single and 12-lead ECGs, the algorithm obtained a sensitivity of 91.21% (95% CI, 86.12%-94.89%) and a specificity of 91.48% (95% CI, 86.33%-95.15%) for AF detection. Among the same dataset, the ECG App 2.0 from Apple Watch yielded inconclusive diagnoses for 12.5% (48/383) of all SW-ECGs while the AI algorithm reduced that number to 6.5% (25/383) and reached a sensitivity and specificity of 82.9% (95% CI, 76.44%-88.12%) and 95.6% (95% CI, 91.19%-98.22%) for AF detection. Conclusions: The Philips Cardiologs AI algorithm has demonstrated strong performances in the detection of AF and SR. In comparison to the Apple Watch, the AI algorithm significantly reduced the number of inconclusive diagnoses by half (p=0.0023, Z-test), thereby significantly improving diagnostic outcomes. This AI driven ECG analysis could enhance detection of AF in SW. Further efforts are warranted to extend the achieved performance to encompass the detection of additional types of arrhythmias.