Toward improving ECG biometric identification using cascaded convolutional neural networks

Abstract

Biometric identification via Electrocardiogram (ECG) signals, which can be captured by devices with ECG sensors, have been explored for human identification for decades. Whereas, the problems of generalization and efficiency for ECG biometric recognition are still challenging. In this paper, we propose a new generic convolutional neural network (CNN) approach (i.e., Cascaded CNN) to realize human identification via ECG biometric identification. In our method, two CNNs are trained progressively. The first CNN called F-CNN is used for feature extraction of ECG heartbeats, and the second one called M-CNN is used for biometric comparison (identification). The trained F-CNN and M-CNN are cascaded to compose the Cascaded CNN as the final identification network. One of the main characteristics of the proposed method is the strong generalization ability. Once the Cascaded CNN is constructed, it can be used for various groups with variable number of members for human identification, which meets the practical demands greatly. Experiments are conducted on five public datasets in PhysioNet to evaluate the performance of the proposed method. By the Cascaded CNN, an average identification rate of 94.3% is achieved without re-training and any fine-tuning for the four test datasets. Moreover, only two milliseconds are needed for once comparison operation. Because of the generalization ability and real-time efficiency, it is feasible to promote the application of the proposed method for biometric identification via ECG in practice.