Att-EMD-Unet: A Novel weakly supervised perspective for ECG segmentation

Automatic ECG segmentation has gained significant attention due to its critical role in cardiac analysis and diagnosis. However, current automatic ECG segmentation methods are hindered by the need for labor-intensive and expert-level annotations. To alleviate the annotation burden, we explore a weakly supervised perspective for ECG QT segmentation. Specifically, we employ annotator-friendly and less expert-intensive casual annotations as supervision signals for model training. In this paper, we propose a novel model called Att-EMD-Unet, which employs U-Net as base network structure and incorporates channel/ temporal attention mechanisms to predict the QT segments from original signals. Recognizing the challenges posed by casual and incomplete labels in our weakly supervised learning framework, we have innovatively incorporated an empirical mode decomposition (EMD) based R and T peaks attentive loss function during the training phase. This function is specifically designed to address and rectify potential inaccuracies or omissions in the estimation of R and T peaks within ECGs. An expert clinician conducted an evaluation of our proposed casual annotation method. The findings from this assessment indicate that it cuts down the time needed for labeling by about 46.58% for each beat in various ECG signals, compared to the usual detailed methods. And the experimental results reveal that our Att-EMD-Unet model surpasses conventional unsupervised methods and achieves comparable performance to state-of-the-art fully supervised learning methods. Our approach effectively overcomes the limitations of extensive labeling required in fully supervised learning, presenting an efficient and accurate solution for ECG QT segmentation