A novel multi-lead ECG personal recognition based on signals functional and structural dependencies using time-frequency representation and evolutionary morphological CNN

https://doi.org/10.1016/j.bspc.2021.102766Get rights and content

Highlights

  • This paper proposes a method for the biometric recognition of a multilead ECG signal.

  • Functional and structural dependencies were calculated as medical variables.

  • Within and cross correlation were determined to estimate functional dependencies.

  • GP-based evolutionary CNN trees were employed to estimate structural dependencies.

  • CNN trees perform the deep feature learning process through morphology operators.

Abstract

Biometric recognition systems have been employed in many aspects of life such as security technologies, data protection, and remote access. Physiological signals, e.g. electrocardiogram (ECG), can potentially be used in biometric recognition. From a medical standpoint, ECG leads have structural and functional dependencies. In fact, precordial ECG leads view the heart from different axial angles, whereas limb leads view it from various coronal angles. This study aimed to design a personal biometric recognition system based on ECG signals by estimating these latent medical variables. To estimate functional dependencies, within-correlation and cross-correlation in time-frequency domain between ECG leads were calculated and represented in the form of extended adjacency matrices. CNN trees were then introduced through genetic programming for the automated estimation of structural dependencies in extended adjacency matrices. CNN trees perform the deep feature learning process by using structural morphology operators. The proposed system was designed for both closed-set identification and verification. It was then tested on two datasets, i.e. PTB and CYBHi, for performance evaluation. Compared with the state-of-the-art methods, the proposed method outperformed all of them.

Introduction

In modern society, reliable recognition of people is used for different purposes such as extreme access control in border terminals, financial systems, and personal applications including mobile phones, laptops, and PCs. An appropriate method for recognition of people is to employ biometric data. Biometrics refers to the technologies which use specific human characteristics (especially biological features that are unique for everyone) to identify people or access secure areas [1]. Different methods of biometric recognition are based on fingerprint, facial, and retinal scans as well as voice analysis. The main problem is these methods are easily bypassed through such methods as presentation attacks or voice recording [2].

Physiological signals are a class of potential biometric methods for identification. Electrocardiogram (ECG) is one of the biological signals that contain identity information. An ECG is the recording of an electrical activity in the heart. It consists of three main components: P wave, QRS complex, and T wave. The P wave occurs due to atrial depolarization, whereas the QRS complex results from ventricular depolarization. Finally, the T wave is caused by ventricular repolarization [3]. There is a unique ECG model for everyone due to differences between people’s hearts in textural and physical characteristics such as size and placement as well as internal differences between precordium and heart. Therefore, an ECG signal has widely been employed to develop biometric systems [4]. The ECG signal has a few advantages over other methods of developing biometric recognition systems. For instance, an ECG is an active signal which can be obtained only from a living individual, the biometrics of whom cannot be faked in case of demise. Moreover, an ECG signal is an intrinsic feature of an individual’s body that cannot be easily faked by others. Finally, an ECG signal can be used simultaneously in other applications because it shows an individual’s health status [5].

Although an ECG generates a contested complement for biometric modalities, studies of ECG-based person recognition are still in the gestation period. Deep feature learning through one or multiple Convolutional Neural Networks (CNNs) is a paradigm designed to extract hidden patterns in ECG time series for personal recognition, the Equal Error Rate (EER) of which was reported below 1.36 % [6]. Disregarding latent medical variables in different leads of the ECG physiological signal, most of these methods, however, considered it merely a general time series through pattern processing. This limits the use of previous methods in naturalistic settings such as different human clinical conditions.

From a medical viewpoint, an ECG signal consists of N leads. Each lead of the multi-lead ECG signals views the heart from a specific angle. Precordial leads scan the heart from different angles of an axial plane, whereas limb leads scan the heart from various angles of a coronal plane. Therefore, all of these signals report different aspects of phenomenon. Their difference lies in the fact that each of them highlights specific pieces of information; however, all of them show the status synchronously, and their information is in functional and structural dependencies [7]. For instance, although all leads leave a trace of atrial stimulation, this performance is more prominent in the double-chest lead. Addressing these medical pre-analyses in ECG signals will help discover latent medical variables and ultimately boost the final autoregression model in different human clinical conditions.

This paper focuses mainly the analysis of functional and structural dependencies in the multi-lead ECG signals to design a human biometric system. A comprehensive study was conducted on two datasets of ECG signals in the disjoint acquisition session to check the generalization of the proposed method. A hybrid learning model was also introduced by analyzing the correlation of frequency bands and multi-CNN-gene in genetic programming. In particular, within-correlation and cross-correlation are first calculated to estimate functional dependencies in the multi-lead ECG signals and show it as extended adjacency matrices. Moreover, multi-CNN-gene genetic programming was introduced for the automated estimation of structural dependencies information through extended adjacency matrices. In GP, genes are the CNN trees which perform the deep feature learning process through structural morphology. In the practical application of an ECG biometric system, the medical method which extracts both functional and structural types of information from a multi-lead ECG signals would have better advantages than the methods which merely calculates general patterns of a signal. The proposed method can simultaneously be used for closed-set identification and verification. For closed-set identification, the functional and structural features extracted in the training phase are presented to a fully-connected neural network to return the user identity. In the verification phase, a biometric matcher is employed to compare the extracted features. The performance of the proposed method is evaluated on two well-known ECG datasets, i.e. PTB (PTB diagnostic ECG database) and CYBHi (Check You Biosignals Here Initiative).

This paper expands the literature on personal recognition based on ECG in two folds. The learning method is first presented. Based on medicine, this method estimates both functional and structural dependencies simultaneously through the multi-lead ECG signals model that is particularly presented as extended matrices, from which dynamic structural features are extracted for recognition of people. The matrix representation of a multi-lead ECG signals overcomes certain problems such as univariate features by finding the correlation between every pair of signal leads. The multi-CNN-gene extracts the intrinsic structural features which improves the human biometric system performance. The second contribution of this paper is the thorough analysis of proposed method on different human states during several days of experiments. This proves the generalization of the proposed method and lack of overfitting in the learning model in the face of new and unseen samples. Finally, the proposed method can be considered an extension of permanence analysis [8] and multitasking [9] in ECG-based biometrics.

This paper consists of the following sections. Section 2 briefly reviews achievements in related works. The proposed method is described thoroughly in detail in Section 3, whereas the results are presented in Section 4 that also evaluates the parameters of the proposed method in comparison with other techniques. Finally, Section 5 presents the discussion and draws a conclusion.

Section snippets

Related works

Based on the types of features used in biometric recognition systems, their ECGs can be divided into handcraft feature generation and deep feature generation [[10], [11], [12]]. In the first class, a few methods are based on the detection of fiducial points in an ECG signal, in which the QRS complex is the most important fiducial point. After the QRS complex is detected, time and frequency [13], statistical [14], and amplitude [12], and morphological features of the signal are extracted from

The proposed method

This section describes the proposed model for multi-lead ECG signals biometric recognition (Fig. 1). The model consists of two components, i.e. functional and structural dependencies estimation. The wavelet transform was employed to extract M important frequency bands from every ECG signal S(i) in S={S(1),,S(i),,S(N)} to estimate functional dependencies. In the proposed method, S is an N-lead signal for a single subject. After that, the within-correlation of the same bands and

Evaluation and results

This section presents the results of the testing of the proposed method. This test was performed on two ECG signal datasets called PTB (PTB Diagnostic ECG Database) [29], CYBHi (Check Your Biosignals Here Initiative) [30]. The PTB dataset contains ECG signals belonging to 294 subjects, of which 54 are healthy and the rest suffer from heart conditions in different days. The data contained in this database has a frequency of 1000 Hz and an amplitude resolution of 0.5μV. In this dataset, a 12-lead

Conclusion

This paper analyzed functional and structural dependencies in multi-lead ECG signals to design a human biometrics system. For this purpose, a hybrid learning model was introduced by analyzing the correlation of frequency bands and multi-CNN-gene in genetic programming. In particular, within-correlation and cross-correlation were first determined in multi-lead ECG signals to estimate functional dependencies. They were then represented as extended adjacency matrices. The multi-CNN-gene was also

CRediT authorship contribution statement

Majid Sepahvand: Conceptualization, Software, Data curation, Writing - original draft, Visualization, Investigation, Validation. Fardin Abdali-Mohammadi: Conceptualization, Methodology, Project administration, Supervision, Writing - review & editing.

Declaration of Competing Interest

The authors report no declarations of interest.

References (35)

  • A. Goshvarpour et al.

    Human identification using a new matching Pursuit-based feature set of ECG

    Comput. Methods Programs Biomed.

    (2019)
  • H.P. da Silva et al.

    Check your biosignals Here: a new dataset for off-the-person ECG biometrics

    Comput. Methods Programs Biomed.

    (2014)
  • M. Hammad et al.

    Parallel score fusion of ECG and fingerprint for human authentication based on convolution neural network

    Comput. Secur.

    (2019)
  • I. Odinaka et al.

    ECG biometric recognition: a comparative analysis

    IEEE Trans. Inf. Forensics Secur.

    (2012)
  • M. Ingale et al.

    ECG biometric authentication: a comparative analysis

    IEEE Access

    (2020)
  • A.S. Eltrass et al.

    A new automated CNN deep learning approach for identification of ECG congestive heart failure and arrhythmia using constant-Q non-stationary Gabor transform

    Biomed. Signal Process. Control

    (2021)
  • F. Kusumoto

    ECG Interpretation

    (2020)
  • Cited by (17)

    • Teacher–student knowledge distillation based on decomposed deep feature representation for intelligent mobile applications

      2022, Expert Systems with Applications
      Citation Excerpt :

      Large-scale statistical models and deep neural networks are employed in medicine, game applications, robotics, augmented reality, and many other areas, which often demand extensive computation resources (e.g. graphics processing units). The Convolutional Neural Network (G. E. Hinton, Osindero, & Teh, 2006; Sepahvand & Abdali-Mohammadi, 2021a), AlexNet (Krizhevsky, Sutskever, & Hinton, 2012), LeNet (LeCun, Bottou, Bengio, & Haffner, 1998), and ResNet (He, Zhang, Ren, & Sun, 2016) models can evidently perform properly in problem-solving due to having deep layers and more parameters. At the same time, intelligence is moving towards edge devices in order to decentralize the server processing workload.

    • Overcoming limitation of dissociation between MD and MI classifications of breast cancer histopathological images through a novel decomposed feature-based knowledge distillation method

      2022, Computers in Biology and Medicine
      Citation Excerpt :

      If the floating-point data type is utilized to define every parameter, the size of this model will be 84.04 M*8B = 640 MB, which is very high. Many researchers have tried to reduce the computational complexity of these models through different methods such as compression [15], model acceleration [16], Interpretable models [17,18], pruning and parameter sharing [19]. However, small-scale devices are usually used in other applications such as monitoring patients and reaching online medical diagnoses that require high-speed execution, for these devices are easily portable [20–22].

    • A novel method for reducing arrhythmia classification from 12-lead ECG signals to single-lead ECG with minimal loss of accuracy through teacher-student knowledge distillation

      2022, Information Sciences
      Citation Excerpt :

      However, the quantities of layers, filters, and hyper-parameters utilized in the CNN models might not have been selected completely optimally. Hence, the authors intend to propose a novel model called the evolutionary KD in the future by adopting a CNN architecture development approach through evolutionary algorithms (e.g., the one introduced in [3,45]). Majid Sepahvand: Conceptualization, Software, Data curation, Writing - original draft, Visualization, Validation, Investigation.

    • Emotion stimuli-based surface electromyography signal classification employing Markov transition field and deep neural networks

      2022, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      In terms of deep learning methods, convolutional neural networks (CNN) are typical deep learning methods. The advantage of this method is that it does not require many feature extraction operations, and a small number of features can also obtain a reasonable recognition rate [22]. The convolutional neural network performs well in image processing because of its local area connection, weight sharing, and down-sampling.

    • Virtual environment for assistant mobile robot

      2023, International Journal of Electrical and Computer Engineering
    View all citing articles on Scopus
    View full text