An Encoder-Decoder Based Approach for ECG Delineation

Xiangju Kong, Shuwang Zhou, Hui Liu, Tianlei Gao, Zhe Zhu

Abstract: Accurate delineation of ECG waveforms is crucial for heart disease diagnosis. Nevertheless, the extraction of deep features from ECG signals is difficult due to the variability in abnormal rhythms and noise distribution. To overcome this, the paper presents an ECG delineation method utilizing an encoder-decoder framework to identify the boundaries of the P-wave, QRS complex, and T-wave. Firstly, a multi-scale feature extraction module is introduced in the encoder, enabling multi-scale feature extraction and enhancing the model’s comprehension of the input data. Additionally, a temporal and spatial feature enhancement module is incorporated, combining temporal and spatial information to capture both dynamic and static features in sequential data. Furthermore, the features extracted by the encoder are concatenated into the decoding module for multi-scale decoding, integrating information from different scales to provide a more comprehensive understanding of the data. The proposed approach is trained and tested on the public ECG databases QTDB and LUDB, achieving an average accuracy of 96.79\(\%\) and 96.43\(\%\), an average sensitivity of 99.25\(\%\) and 99.23\(\%\), an average PPV of 98.45\(\%\) and 98.00\(\%\), and an average F1 score of 0.989 and 0.986, respectively.