Smartwatch-based detection of atrial arrhythmia using a deep neural network in a tertiary care hospital. (19th May 2022)

• Record Type: Journal Article
• Authors: Fiorina, L; Lefebvre, B; Gardella, C; Henry, C; Coquard, C; Younsi, S; Ait Said, M; Salerno, F; Horvilleur, J; Lacotte, J; Mannenti, V
• Is Part Of: Europace. Volume 24:Supplement 1(2022)
• Publication Date: 2022-05-19
• Subjects: Arrhythmia -- Treatment -- Periodicals; Cardiac pacing -- Periodicals; Catheter ablation -- Periodicals; Heart -- Physiology -- Periodicals; Electrophysiology -- Periodicals; 617.4120645
• DOI: 10.1093/europace/euac053.563
• Languages: English
• ISSNs: 1099-5129

Abstract

Abstract: Funding Acknowledgements: Type of funding sources: None. Background/Introduction: Smartwatch electrocardiograms (SW ECG) have been identified as a promising noninvasive solution to assess heart rhythm abnormalities, especially atrial arrhythmias (AA) which includes atrial fibrillation, atrial flutter and supraventricular tachycardia. This study evaluates the performance of the detection of AA with a smartwatch and compares the accuracy of two algorithms, the latest version of the original companion application (Apple ECG 2.0 App) and a novel deep neural network (DNN), in a population typical of an electrophysiology department. Purpose: Determine if a novel DNN can improve the detection of AA on SW ECG in a tertiary care hospital. Methods: 101 patients from the electrophysiology department of one tertiary center were included in this ongoing study. Three simultaneous ECGs were collected for each patient: one 12-lead ECG (Mindray BeneHeart R12) and two SW ECGs (Apple Watch) taken from the left wrist (SWw ECG) and the lower left abdomen (SWa ECG). 12-lead ECGs were adjudicated by a blinded expert electrophysiologist as 52 AA and 49 not AA and considered as gold standard. The SW ECGs were processed by the ECG 2.0 App and the DNN in parallel. The proportions of inconclusive diagnoses returned and the performances were assessed and compared. Results: Overall, the ECG 2.0 App yielded inconclusive diagnoses for 19% (19/101) of all SWw ECGs while the DNN reduced that number to 0% (0/101). A similar result holds for SWa ECGs (Figure 1). Regarding the detection of AA from SWw ECGs, the ECG 2.0 App had a sensitivity of 81% (95% CI, 67%-90%), a specificity of 97% (95% CI, 87%-100%) and an accuracy of 89% (95% CI, 80%-94%) while the DNN had a sensitivity of 92% (95% CI, 82%-97%), a specificity of 90% (95% CI, 78%-96%) and an accuracy of 91% (95% CI, 84%-95%). For SWa ECGs (Figure 2), the sensitivity of the DNN was found significantly higher compared to the ECG 2.0 App: 96% (95% CI, 89%-98%) vs 76% (95% CI, 61%-87%). Conclusion(s): A novel DNN algorithm decreased the number of inconclusive diagnostics in the detection of AA from SW ECG from around 20% to 0%, which could help limit the overreading time spent by the physicians. Excluding inconclusive diagnostics, we observed no significant difference in performance between the two algorithms except for the sensitivity for SW ECG taken from the abdomen where the DNN outperforms the ECG 2.0 App. Routine application of this SW ECG analysis in tertiary care hospitals offers significant promise in arrhythmia diagnosis.