Cardiac Arrhythmias Identification by Parallel CNNs and ECG Time-Frequency Representation

Heart abnormalities cause about 26 % of the deaths of illnesses in the world. Developing computational tools for ECG interpretation plays a critical role in the clinical diagnosis of Cardiac arrhythmias (CAs). Aims: This study aimed to develop an automated abnormal pattern recognition method for clinical decision support capable of detecting between 27 possible CAs. Proposal: An improved deep learning (DL) model was employed using raw-data and time-frequency representation (TFR) images. Methods: A vast set of ECG records were filtered and normalized. They were segmented and transformed into two sets of 2-D images. TFR images were obtained through Wavelet Synchrosqueezing (WS). The VGG-16 network was chosen, modifying the weights of the inner layers to adapt the model to the CAs detection task. A 10-fold cross-validation method was executed. Different training hyperparameters were tested to find the best model. Results: With the cross-validation on the training data, the model developed by our team UIDT-UNAM performed identifying CAs, with an overall unofficial S-score of 0.766. This model had a high performance in detecting healthy subjects with an F1 score of 0.83. We obtained these results using only the public training dataset. We plan to test these optimistic results with Physionet private dataset very soon.