Ensemble Deep Learning Models for ECG-based Biometrics

In this paper, we study ensemble intelligent system for individual identification using ECG (electrocardiogram) signal by combining several time-frequency representations and CNN models. The ECG signal is a bio-signal that measures the minute potential difference generated by heart activity through the skin of the body surface. ECG signals have an advantage in personal identification because the source of the signal is hidden inside the body and the measurement equipment is inexpensive. However the ECG signals are sensitive to noise. Time-frequency representation gives observation of frequency changes in signal over time and allows robust analysis of noise. The log-spectrogram, mel-spectrogram, spectrogram, MFCC(Mel Frequency Cepstral Coefficient), and scalogram are considered as time-frequency representations. CNN(Convolutional Neural Network) is a neural network designed to enable feature extraction and classification of images to be learned in a single structure. Xception, ResNet, and DenseNet are famous models in CNN and are used in this study. PTB (Physikalisch-Technische Bundesanstalt)-ECG database is used for experiment. The ensemble CNNs have shown higher than single CNN.