Identification of Cardiac Arrhythmias from 12-lead ECG Using Beat-wise Analysis and a Combination of CNN and LSTM
Alkhodari M., Hadjileontiadis LJ., Khandoker AH.
Throughout the years, there have been many attempts to develop an accurate cardiac arrhythmias identification algorithm. However, despite achieving acceptable results, they have been only applied on either small or homogeneous data-sets. A study was developed herein to identify cardiac arrhythmias from varied-length 12-lead ECG signals obtained from the PhysioNet/Computing in Cardiology Challenge 2020 and acquired from a wide set of sources. Our team, Care4MyHeart, developed an approach that starts by analyzing the labels of the database. Then, applying various signal processing techniques to denoise the 12-lead signals. After that a beat-by-beat segmentation procedure was followed to identify the most significant beats in exhibiting the arrhythmia within the signals. A CNN+BiLSTM model was then trained and evaluated on the training set using 10-fold cross-validation scheme as well as on hidden validation and testing sets. Our approach achieved a challenge validation score of 0.379 and full test score of 0.146 on the hidden validation and testing sets, respectively. Our team was ranked the 26th out of 41 entries in this year's Challenge.