Deep Convolutional Neural Network for Imbalanced Multi-label ECG Classification

Xiaoyu Li1, Buyue Qian1, Jishang Wei2, Yuhua Wei1, Haochen Han1, Xiyang Li3
1Xi'an Jiaotong University, 2HP Labs, 3Xi'an Jiaotong University Health Science Center


Though hand-crafted features have been proven effective for certain arrhythmia, there is much room for deep learning methods to explore. In this work, we adopt InceptionTime network as baseline and explore the design space in imbalanced multi-task classification setting. Currently, we design InceptionTime network with network architecture search, apply focal loss, and utilize snapshot ensemble method without any other ensemble learning. Such model achieves geometric mean of F2 and G2 as 0.646. We will study further about enhancing InceptionTime with attention and also ensembling with feature engineering-based classifiers.