This study presents PhysioNauts Team's contribution to the PhysioNet/CinC Challenge 2021 on ECG classification for variable leads. Three types of labels were identified: those affecting cardiac rhythm, ECG morphology or both. The full model integrated handcrafted rhythm features and deep learning features into a residual neural network (ResNet) with a squeeze and excitation module and a wide 10-neuron single-layer fully connected (FC) branch to leverage the learning of both feature types. The ResNet inputs were ECG segments of 4096 samples downsampled to 257 Hz. The FC inputs were standard rhythm features extracted from the RR-series. Class imbalance was mitigated by selecting only a third of normal sinus rhythm and sinus bradycardia recordings. Moreover, threshold optimization was performed based on a grid search and the Nelder-Mead method to maximize the Challenge metric (CM). Our entry failed on the UMich test data, so it was not officially ranked and it didn't receive official scores on the full test set. The CMs obtained in the unofficial entry were 0.613, 0.585, 0.603, 0,594, and 0.582 on 12-lead, 6-lead, 4-lead, 3-lead, 2-lead, respectively.

Objective This work presents an ECG classificator for variable leads as a contribution to the PhysioNet/CinC Challenge 2021. It aims to integrate deep and classic machine learning features into a single model, exploring the proper structure and training procedure. Approach From the initial 88253 signals, only 84210 were included. Low quality and unscored recordings were excluded. Three different database subsets of 40365 recording each were created by dividing in three normal sinus rhythm and sinus bradycardia recordings. Each subset was used to train a different model with shared architecture integrated as an ensemble to provide the final classification through major voting. Models contained a deep branch composed of a modified ResNet with dilation convolutional layers and squeeze and excitation Block that took as input windowed ECG signals. This was concatenated with a wide branch that integrated 20 cardiac rhythm features into a fully connected 3-layered network. Three different training steps were studied: just the deep branch (D), wide integration and training (D+W), and a final fine tuning of the deep branch posterior to wide training (D+W+D). Main Results Results obtained in a local test set formed by a an stratified 12.5% split of the given full datasetwere presented for 2-lead and 12-lead models. The best training method was the 3-step D+W+D procedure obtaining a challenge metric of 0.709 and 0.677 for 12 and 2-lead models respectively. Significance Integration of handcrafted features and deep learning model not only may increase the generalization capacity of the network but also provide a path to add explicit information into the classification decision process. To the best of our knowledge this is the first work studying the best training procedure to properly integrate both types of information for ECG signals classification.