A deep multi-task learning approach for ECG data analysis

Deep learning is an advanced representation learning method and can automatically discover hidden features from raw data. Researchers have attempted to adopt it for ECG data analysis in the past few years. However, traditional deep learning algorithms usually require great efforts and experience to fine-tune the neural networks during their training processes. Moreover, these algorithms may suffer from a sharply declined accuracy when a well-trained model is directly applied to analyze the data from another group of patients. To address these issues, we propose a deep multi-task learning scheme for ECG data analysis which only requires limited efforts to fine-tune the network and can enable the trained model to be well applied to other datasets. Specifically, we first convert the ECG data analysis problem into a multi-task learning problem by dividing the ECG data analysis into multiple tasks. We then construct the multiple datasets for each task. Finally, we design a deep parameter-sharing network which inserts parameter-sharing neural layers in traditional neural networks. We conduct experiments by using the MIT-BIH database to validate the performance of our proposed scheme. Results illustrate that our proposed scheme can improve the accuracy of ECG data analysis by up to about 5.1%.