A Deep Neural Network for Interpreting Wearable Electrocardiogram Data in Atrial Fibrillation: Prospective Observational Diagnostic Accuracy Study

Clinical Snapshot

Key Findings

• P Value: P<.001 for ICC comparison

• Effect Size: Sensitivity 91.9%, Specificity 99.6%

• Primary Outcome: AF/AFL detection accuracy

• Nnt Or Sensitivity: Sensitivity 96.2% for AF, 55.8% for AFL; PPV 99.5%, NPV 93.1%

• Confidence Interval: Not reported for primary outcomes

Clinical Application

Requires wearable device compliance and technical infrastructure for AI analysis Could support RACGP guidelines for AF screening in high-risk patients; TGA approval would be required for clinical use; cost-effectiveness versus standard Holter monitoring needs evaluation Patients with suspected paroxysmal AF requiring rhythm monitoring, particularly post-cardioversion surveillance

Abstract

BACKGROUND: Atrial fibrillation (AF) and atrial flutter (AFL) are common arrhythmias associated with the risk of ischemic stroke, which can be reduced with anticoagulation therapy. Thus, early diagnosis of AF and AFL is essential. However, diagnosis may be challenging due to the paroxysmal and asymptomatic nature of these arrhythmias. OBJECTIVE: Current diagnostic workflows involve time-consuming and resource-intensive manual review of noisy signals and prolonged recordings. We evaluated a mobile system that combines a wireless wearable single-lead chest strap electrocardiogram (ECG) and a novel deep neural network (DNN)-based artificial intelligence (AI) method for detecting AF/AFL episodes, AF/AFL burden, and rhythm change and estimated the delay in the detection of rhythm change from AF/AFL to sinus rhythm. We also assessed the rhythm classification performance. METHODS: A total of 116 patients with recent-onset AF or AFL undergoing cardioversion were monitored using a mobile single-lead chest strap ECG system. Simultaneously, a 3-lead Holter ECG served as the reference. The DNN-based AI analyzed the single-lead chest strap ECG data to detect AF/AFL, non-AF/AFLrhythm, and noninterpretable segments, as well as to estimate AF/AFL burden and detect rhythm change. Performance metrics included sensitivity, specificity, positive predictive value, negative predictive value, and intraclass correlation coefficient for AF and AFL burden estimation. RESULTS: The sensitivity and specificity for detecting AF/AFL were 91.9% (204.9/223.0 h) and 99.6% (242.4/243.5 h), respectively. The sensitivity for detecting AF was 96.2% (191.5/199.0 h), whereas it was 55.8% (13.4/24.0 h) for detecting AFL. The positive predictive value and negative predictive value for AF/AFL detection were 99.5% (204.9/206.0 h) and 93.1% (242.4/260.5 h), respectively. The intraclass correlation coefficient between the AF/AFL burden estimated by the DNN-based AI method and that derived from the physician-interpreted reference ECG was 0.96 (95% CI: 0.94-0.97; P<.001). Rhythm change detection occurred within 1 minute in most cases. CONCLUSIONS: The mobile single-lead chest strap ECG system powered by a DNN-based AI algorithm demonstrated strong performance in detecting AF, estimating AF burden, and recognizing rhythm change to sinus rhythm. This AI-driven approach enables automated and accurate rhythm analysis, supporting clinical decision-making. Further validation in real-world ambulatory settings is warranted.

References

1. 1.Rantula, O. A., Lipponen, J. A., Halonen, J., Jäntti, H., Rissanen, T. T., Tarvainen, M. P., Naukkarinen, N. S., Seljola, E.-S., Santala, O. E., Sedha, J., Martikainen, T. J., & Hartikainen, J. E. K. (2026). A Deep Neural Network for Interpreting Wearable Electrocardiogram Data in Atrial Fibrillation: Prospective Observational Diagnostic Accuracy Study. *JMIR mHealth and uHealth*. https://doi.org/10.2196/82475