An Algorithm for Locating PVC and SPB in Wearable ECGs

Wearable electrocardiogram devices are being developed rapidly as they can be used to monitor the user's health in the long term. Their development is dependent on advances in automatic processing algorithms. Given that its intended use requires the wearable device to collect a large amount of data, and in light of the context where its hardware is limited by cost and power consumption, this paper proposes an algorithmic framework that can automatically identify position premature ventricular contraction (PVC) and supraventricular premature beat (SPB) at low computational cost by using 24-hour dynamic single-lead ECG recordings. We first use convolutional neural networks that require a small number of calculations to find problematic segments in the signal, and then use bidirectional long short-term memory networks to accurately locate the PVC and SPB. In addition to using the second-level structure that reduces the total amount of computation required by the task to 1/3.78, the algorithm has adopted focal loss to ensure the accuracy of classification of the disease samples. The algorithm was used in the Third China Physiological Signal Challenge (CPSC) 2020 and came in 10th place, which verified its effectiveness. We trained the model once again by using the data provided at CPSC 2020, and obtained the F1 scores of PVC and SPB were 0.6698 and 0.6111.