Deep-Learning Premature Contraction Localization Using Gaussian Based Predicted Data

• Main Authors: Novotna, Petra, Vicar, Tomas, Hejc, Jakub, Ronzhina, Marina
• Format: Conference Proceeding
• Language: English
• Subjects: Detectors; Electrocardiography; Learning systems; Location awareness; Picture archiving and communication systems; Time series analysis; Training
• ISSN: 2325-887X
• DOI: 10.23919/CinC53138.2021.9662903

Detection of cardiac arrhythmias is still an ongoing challenge. Here we focus on premature ventricular contraction (PVC) and premature atrial contraction (PAC) and introduce a deep-learning-based method for PVC/PAC localization in ECG. Our method is based on involving the time series with non-zero values corresponding to the ground truth PVC/PAC positions into the training process. To improve the efficiency of deep model training, the transition between the non-zero and zero areas in the train output time series was smoothed by introducing a Gaussian function. When applied to the new ECGs, the output signal (time series including Gaussians) is processed by a robust peak detector with Bayesian optimization of threshold, minimal distance and peak prominence. Positions of the detected peaks correspond to the desired PVC/PAC positions. The proposed method was evaluated on China Physiological Signal Challenge 2018 (CPSC2018) using own-created ground truth positions of PVC/PAC. The proposed method reached F1 score 0.923 and 0.688 for PAC and PVC, respectively, which is better than our previous results obtained via multiple instance learning-based method.