Semantic Segmentation of QRS complex in 12-Lead ECG Signals

In this work we present a method for semantic segmentation of the QRS complex in 12-lead ECG signals. The dataset used in the work was St Petersburg INCART 12-lead. The proposed method is comprised or three steps: 1. data preprocessing, where the binary vectors corresponding to the ground truth are obtained. In the ground truth vector, the samples corresponding to the QRS complex are identified as logical 1, while the other samples are identified as logical 0; 2. segmentation of the QRS complex, using autoencoders. Three different autoencoder architectures, with different number of kernels in convolutions layers are evaluated; 3. post-processing filters to remove noise, desgined using heuristic rules. The result obtained for detecting the QRS complex, an F1-score over 99.90%, is the state-of-the-art and surpasses the performance obtained with other methods previously published in the literature.