Clinical Assessment of a Lightweight CNN Model for Real-Time Atrial Fibrillation Prediction in Continuous Wearable Monitoring

• Published in: 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) Vol. 2024; pp. 1 - 4
• Main Authors: Soares, Quenaz B, Andrade, Douglas A, Ribeiro, Estela, Verardino, Renata G S, Reis, Tallita C, Samesima, Nelson, Monteiro, Rosangela, Jatene, Fabio B, Gutierrez, Marco A
• Format: Conference Proceeding Journal Article
• Language: English
• Published: United States IEEE 01.07.2024
• Subjects: Arrhythmia; Atrial fibrillation; Atrial Fibrillation - diagnosis; Atrial Fibrillation - physiopathology; clinical assessment; deep learning; Electrocardiography; Electrocardiography, Ambulatory - instrumentation; Female; Humans; Male; Middle Aged; Monitoring; Monitoring, Physiologic - instrumentation; Neural Networks, Computer; Performance evaluation; Prediction algorithms; Predictive models; Real-time systems; Surgery; Wearable devices; wearable ECG; Wearable Electronic Devices
• ISSN: 2694-0604
• DOI: 10.1109/EMBC53108.2024.10781856

Atrial Fibrillation (AFib) represents a prevalent cardiac arrhythmia associated with substantial risk for affected individuals. The integration of wearable devices, coupled with advanced predictive models, opens pathways for non-invasive and real-time monitoring of AFib. This paper outlines the assessment of a lightweight CNN model designed to predict AFib in patients, enabling continuous monitoring through the utilization of wearable devices. In the performance evaluation, our model demonstrated robust outcomes, achieving F1 score of 0.96, AUC of 0.99, and Acc of 0.96 on the test set. To assess the clinical relevance of our model, we conducted an experiment involving 24-hour continuous monitoring using a single-lead ECG device on 28 arrhythmic subjects and 28 subjects without prior arrhythmia diagnosis. The clinical assessment achieved F1 score of 0.94, AUC of 0.99, and Acc of 0.97 when compared to the physician's true labels. These findings underscore the potential of our model for real-world applications, particularly in the continuous monitoring of patients at risk of developing arrhythmia, offering a non-invasive and effective approach for early detection of AFib.