IMPROVING ELECTROCARDIOGRAM CLASSIFICATION USING TRANSFER LEARNING AND LIGHTWEIGHT DENSENET-BILSTM

Abstract

The availability of affordable and user-friendly electrocardiogram monitors has improved healthcare for patients with periodic heart arrhythmias. However, effectively diagnosing electrocardiogram records remains challenging, even for experienced medical professionals. This work introduced a transfer learning-based algorithm for electrocardiogram classification using lightweight Densely Connected Convolutional Networks (DenseNets) integrated with Bidirectional Long Short-Term Memory (BiLSTM). We first pre-trained our model on the Icentia11K dataset, the largest public dataset of continuous electrocardiogram records, then fine-tuned it on the CPSC2018 dataset. Our model demonstrated performance comparable to state-of-the-art methods, obtaining an  score of 0.839 without pre-training. With pre-training, the score further improved to 0.849. The proposed network structure outperformed existing methods in various metrics, including Area Under the Curve, , , and . The Area Under the Curve and  values were 0.986 and 0.886, respectively for CPSC2018 dataset.