Abstract: SA-PO585
Automatic Detection of Intradialytic Paroxysmal Atrial Fibrillation and Flutter in Single-Lead ECG
Session Information
- Hemodialysis: Biomarkers, Translational Research
November 04, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 801 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Yun, Donghwan, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
- Kim, Dong Ki, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
- Oh, Kook-Hwan, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
- Han, Seung Seok, Seoul National University Hospital, Jongno-gu, Seoul, Korea (the Republic of)
Background
Intradialytic paroxysmal atrial fibrillation and flutter (AF/AFL) can result in hemodynamic instability and suboptimal dialysis dose. To achieve appropriate prevention and management of intradialytic paroxysmal AF/AFL, the first step is precise detection during hemodialysis session. Herein, we developed a Transformer-based model to automatically segment AF/AFL in single-lead echocardiography (ECG) by self-supervised learning and masked signal modelling.
Methods
To develop a model, we used 11 open source databases containing 5,684-hour ECG signals from 2,017 patients. To validate a model, intradialytic ECG signals were retrieved from 4-hour 30 hemodialysis sessions at Seoul National University Hospital. AF/AFL was defined as ≥10-second duration, and the Swin Transformer model with wide window-size was adopted as a main architecture. The model performance was evaluated with F1 score.
Results
The model achieved F1 scores of 0.9612 and 0.9731 in segmenting AF/AFL and other rhythms, respectively. Throughout cross-database, F1 scores ranged between 0.8889 and 0.9896 depending on the prevalence of AF/AFL. Ablation analysis identified that the performance was attributable to pretraining with masking random signals and unlabeled database. The present model was superior to previous models in detecting AF/AFL particularly when training and testing databases were matched. When applying to intradialytic ECGs, the model showed favorable performance in segmenting AF/AFL areas.
Conclusion
The model with self-supervised learning and masked signal modelling maintains robust performance in various databases, which will help automatic detection of intradialytic paroxysmal AF/AFL.