ABSTRACT
In purpose of screening arrhythmia, wearable adhesive patch-type electrocardiographs that can measure electrocardiogram continuously for 14 days have been replacing the 24-hour Holter monitor. The reason for that is the patch-type electrocardiograph being smaller and lighter than the Holter monitor, making it more convenient for patients to coexist with in their daily lives. However, this type of electrocardiograph generates a lot of noise signals due to movements during various physical activities and extended wear time.While analyzing electrocardiograms automatically using software, noise signals make the analysis difficult and they may be misclassified as arrhythmia signals. These misclassified signals require a lot of effort and time from clinical technicians to reclassify them as noise. To resolve this problem, this study hypothesized that a deep learning algorithm could be used to screen noise signals. We used 7,467 noise signals and 15,638 ECG signals collected from arrhythmia patients and healthy people. The signals were divided into 10 seconds segments and labeled by cardiologists. We split the data into training and test datasets, ensuring no patient overlap.A hybrid noise classification model, Squeeze and Excitation - Residual Network - Vision Transformer (SE-ResNet-ViT) was developed using the training and validation datasets with an 8:2 ratio. We evaluated the performance of the model using a test dataset. The best F1 score was 0.964. The proposed model can effectively screen for noise signals and potentially reducing the time and effort required by clinical technicians.