Publication: Electrocardiogram (ECG) biometric system using transformer (deep learning)
Loading...
Date
2023-07
Authors
Daniel, Goh Shang Choon
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Electrocardiogram signals contains sufficient inter-person variability that makes it viable to use as a unique identity token for identification and authentication purposes. Thus, feature extraction needs to be conducted to extract features of an individual ECG signals, which are conducted via deep learning. Past deep learning architecture utilized Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM). But these models have common limitations that may inhibit its performance, which includes exploding or vanishing gradients, long-term dependencies and limited to sequential input of data. Hence, this project proposes the creation of two models of ECG biometric system which are ECG Biometric Identification model and ECG Biometric Authentication model using Transformer. The first one identifies the received ECG signal by comparing with the enrolled ECG signals, whereas the second one determines whether the signal belongs to the authorized person The signals from the MIT-BIH Atrial Fibrillation Database were sent through pre-processing steps, segmentation, and preparation before being used for model training. The first model was unable to be trained due to insufficient virtual RAM in Google Colab, whereas the second model was successfully trained with two datasets with different sizes of datasets. Both instances of training showed the model is in the best fit and achieve high accuracy at 0.9946 on the first dataset and 0.9760 on the second dataset.