A Computational Intelligent Technique for Biometric Recognition of Electrocardiograms (ECG)

Abstract

Biometric systems have become integrated into the fabric of everyday life – deployed where and whenever secure access to a trusted instrument is required. Since the discovery of fingerprints over 100 years ago, a variety of approaches to person identification\verification have been devised. Recently, electrocardiogram (ECG) has been introduced as a new biometric trait. It distinguishes itself by being a liveliness indicator without any further processing and a difficult to falsify biometric trait. The existing ECG based biometric systems can be generally categorized according to the utilized features to fiducial and non-fiducial systems. The derivation of fiducial features significantly relies on the accuracy of detecting 11 fiducial points, which is a very challenging task by itself. On the other hand, non-fiducial approaches relax the detection process to include only the sharpest point or sometimes no fiducial detection is needed. However, they usually result in high dimension feature space. This work presents a systematic study that contributes to ECG based individual identification. A fiducial based approach that utilizes a super set of features is first introduced. Reduction of this set has been investigated using different reduction techniques like principle component analysis (PCA), linear discriminant analysis (LDA), rough sets (RS) and information gain (IG). Furthermore, in order to relax the fiducial detection process, another fiducial feature set namely PV set that is derived from only five peaks and valleys fiducial points has been also introduced. The results showed that the IG overwhelms the other considered reduction techniques, while the PV set preserves the Subject Identification (SI) accuracy with slight decrease in the Heartbeat Recognition (HR) accuracy. Moreover, a non-fiducial wavelet based approach is proposed. To avoid the high dimensionality of the resultant wavelet coefficient structure, the structure has been investigated using a proposed two-phase reduction process which has resulted in excluding roughly 65% of the structure. In addition, the proposed non-fiducial approach has been applied to different heartbeat representations such as RR, QT, QRS intervals. The results revealed the deficiencies of utilizing QT and QRS intervals only for representing heartbeats.