Accurate Ventricular Tachyarrhytmia Beats Detection on Low Sample Rate ECG Patch Signals Using 1D U-Net

ECG signals quality from mobile cardiac telemetry (MCT) wearable is much noisier than Holter or standard twelve leads ECG. Although, there are beats detection algorithms that has been shown to be accurate for MIT-BIH data, their performances degrade when applying to patches data and non sinus rhythms, especially when detecting ventricular beats on ventricular tachyarrhythmia. This paper presents a deep learning approach using convolutional neural network 1D U-net architecture as a core model, accompanied with miniature pre-processing and post-processing. The model consists of contracting path and expanding path. The contracting path is a sequence of multiple convolution layers and max pooling layers while the expanding path is a sequence of multiple convolution layers and up-convolution layers. There are internal connections between the contracting path and the expanding path to avoid gradient vanishing. The output of the model predicts beat probability map which can be converted to beat locations. Performance of the model on patch data gives 98.86% recall and 97.46% F1-score which is better than Pan-Tompkins by 2.48% and 0.33% respectively. For only ventricular beats, the recall is 95.21% which outperform Pan-Tompkins by 3.68%.