ET-UMAP integration feature for ECG biometrics using Stacking

Abstract

In recent years, personal information leakage incidents have occurred frequently. Currently, a safe and efficient identification method is essential to prevent data leakage. The electrocardiogram (ECG) has attracted more and more attention because of its high security. Researchers have made many progresses in the study of ECG identification. However, there is feature redundancy in the original heartbeats, which will lead to low identification accuracy. To address above issues, this paper proposes a new ECG biometrics configuration: the extremely randomized trees (ET) and Uniform Manifold Approximation and Projection (UMAP) integration feature (ETUMAP feature) for ECG biometrics using Stacking. First, we extract the primary features from the ECG signal with the primary learner of Stacked Generalization (Stacking), namely the extremely randomized trees, and then fuse the UMAP dimensionality reduction algorithm to generate the ETUMAP features. Second, the ETUMAP features are fed into Stacking's secondary learner, extreme gradient boosting (XGBoost), to complete the identification. The effectiveness of the proposed method is evaluated on two public databases, namely, ECG-ID and PTB database. Our method can achieve high subject identification accuracies of 98.88% and 95.77% from the ECG-ID (89 individuals) and PTB (71 individuals) database. In addition, the subject identification accuracy reaches 96.88% in classifying above 160 individuals. Experimental results show that our method effectively reduces redundancy of ECG signals, and has preferable performance and generalization of identification.