| Popis: |
Atrial fibrillation (AF) is the most common cardiac arrhythmia and can lead to serious complications, such as stroke, systemic embolism, and heart failure. Paroxysmal atrial fibrillation (PAF) is one of the critical AF types that should be detected and managed early. PAF is diagnosable when a 10-second, 12-lead ECG shows an irregular rhythm. However, detection yield remains low due to the intermittent and often asymptomatic nature of the condition. Furthermore, detecting PAF on ECGs during normal sinus rhythm (NSR) is more difficult because the RR interval and P-wave are often similar to those of healthy individuals, making accurate diagnosis even harder. In this study, we introduce a deep learning model designed to assist cardiologists in the early diagnosis of PAF during NSR by classifying ECGs into Healthy-NSR ECG and PAF-NSR ECG categories. We propose the total-sub-length ECG (TS-ECG) network, a deep learning model that simultaneously learns two distinct features manifesting in AF. TS-ECG comprises two frameworks: one focusing on the rhythm characteristics across the total-length of the ECG and the other that learns the P-wave segments. We collected 3,591 ECGs from 785 PAF patients and 4,385 ECGs from 1,583 healthy individuals for the study. We verified the performance of the proposed model on ECG data from healthy individuals with no history of AF and from those who experienced PAF during NSR at a multicenter university hospital in South Korea. On the PAF-NSR test dataset from Inha University Hospital, the model achieved an area under the receiver operating characteristic curve (AUROC), precision, recall, and F1-score of 0.940, 0.874, 0.879, and 0.876, respectively. Code is available at https://github.com/mskim1024/TS-ECG |
