Premature Beat Detection from Long-term ECGs Using Modified U-Net
Shan Yang, Heng Xiang, Chunli Wang, Qingda Kong
Chengdu Spaceon Electronics Co., LTD.
Premature ventricular contraction (PVC) and supraventricular premature beat (SPB) are the most common arrhythmias, the detection of which plays an important role in ECG signal analysis. Accurate detection is a challenging task from 24-hour dynamic single-lead ECG recordings. The rulebased PVC and SPB detection methods largely depend on hand-crafted manual features and parameters, the fixed features and parameters of which require difficult offline tuning for adapting to new scenarios. In the 3rd China Physiological Signal Challenge 2020 (CPSC 2020), inspired by the popular application of U-Net in medical image segmentation, the U-Net-like architecture based on 1-D convolutional neural network (CNN) is proposed. The ResNet and ResNeXt block are introduced as backbone of encoder and decoder in the 1D U-Net model. In addition, the ECG records with frequency of 400 Hz are resampled to 200Hz, and to make the length of data fed into network is suitable, zero padding and data truncation are introduced. To increase the diversity of dataset and improve the generalization performance, some common techniques of data augmentation used in this study consist of noise addition, y-axis shift, and wavelet-based filter. The proposed method has been validated against the 3rd china physiological signal challenge data set, obtaining a PVC score of 51335, SPB score of 72488 on the hidden subtest set. Experimental results show that the proposed method acquires competitive performance.
'PythonEntry' is a folder contains the submit prediction code and model weight.
'src_preliminary' is a folder contains the preliminary training and validation code.
'src_final' is a folder contains the final training and validation code.
'TrainingSet' is a folder contains the final training data.
You can run this code by installing the requirements and running
python CPSC_YS1019.py
python CPSC_YS1020_predict.py --mtype resnet --model model_weights.h5
where mtype is a parameter of model type, model is a parameter of model weight, The ICBEB/CPSC 2020 webpage provides a training database with data files and a description of the contents and structure of these files.
python CPSC2020_score.py
which obtains score.txt, example:
S_score: -3483.0
V_score: -11086.0
See the ICBEB/CPSC 2020 webpage for more details, including instructions for the other files in this repository.