An Attention Based Hierarchical LSTM Architecture for ECG Biometric System

The electrocardiogram (ECG) based biometric sys-
tem has recently gained popularity. Easy signal acquisition and
robustness against falsification are the major advantages of the
ECG based biometric system. This biometric system can help
automate the subject identification and authentication aspect of
personalised healthcare services. In this paper, we have designed
a novel attention based hierarchical long short-term memory
(LSTM) model to learn the biometric representation correspond-
ing to a person. The hierarchical LSTM model proposed in this
paper can learn the temporal variation of the ECG signal in
different abstractions. This addresses the long term dependency
issue of the LSTM network in our application. The attention
mechanism of the model learns to capture the ECG complexes
that have more biometric information corresponding to each
person. These ECG complexes are given more weight to learn
better biometric representation. The proposed system is less
complex and more efficient as it does not require the detection
of any fiducial points. We have evaluated the proposed model for
both the person verification and identification problems using
two on-the-person ECG databases and two off-the-person ECG
databases. The proposed framework is found to perform better
than the existing fiducial and non-fiducial point based methods.