NEXT-GENERATION ECG BIOMETRICS: A DEEP LEARNING APPROACH FOR ENHANCED IDENTITY VERIFICATION

ABSTRACT

With the increasing demand for secure and efficient identity verification, ECG-based biometric authentication has emerged as a promising alternative to traditional methods. This study introduces a next-generation system that leverages advanced deep learning techniques to analyze and interpret the complex waveform characteristics of the human heart. The proposed framework comprises several key stages: signal preprocessing, feature extraction through wavelet analysis, and classification via an Artificial Neural Network (ANN). Extensive experiments on a publicly available ECG dataset demonstrate outstanding performance with an accuracy of 98% and a true positive rate of 95%, all within processing times suitable for real-time applications. The system’s robust handling of noise and inter-individual variability highlights its potential for deployment in modern security infrastructures. Future work will focus on further reducing false positives and expanding the framework to address emerging biometric challenges.

АННОТАЦИЯ

На фоне растущих требований к безопасной и эффективной проверке личности биометрическая аутентификация на основе ЭКГ становится перспективной альтернативой традиционным методам. В данном исследовании представлена система нового поколения, использующая передовые алгоритмы глубокого обучения для анализа и интерпретации сложных волновых характеристик человеческого сердца. Предлагаемая структура включает несколько ключевых этапов: предварительную обработку сигналов, извлечение признаков с помощью вейвлет-анализа и классификацию посредством искусственной нейронной сети (ИНС). Обширные эксперименты на общедоступном наборе данных ЭКГ продемонстрировали выдающуюся производительность с точностью 98% и чувствительностью 95%, при этом время обработки позволяет использовать систему в реальном времени. Надёжная работа системы в условиях шума и межличностной изменчивости подчёркивает её потенциал для применения в современных системах безопасности. В будущих исследованиях планируется дальнейшее снижение числа ложных срабатываний и расширение структуры для решения новых биометрических задач.

Keywords: Deep learning, ECG signal, artificial neural network, security, biometric authentication, signal processing, QRS complex, wavelet decomposition.

Ключевые слова: Глубокое обучение, сигнал ЭКГ, искусственная нейронная сеть, безопасность, биометрическая аутентификация, обработка сигналов, QRS-комплекс, вейвлет-анализ.

Introduction

Biometric authentication is paramount in protecting sensitive information and resources [1]. ECG signals have unique and stable characteristics, and hence they have drawn much attention as some promising physiological signals for biometric identification. Patterns of the electrical activity of the heart, representing ECG signals, are an effective way of individual recognition. This study presents a novel biometric authentication system based on ECG signals and advanced deep learning approach. As such, the details of the proposed system here aim to enhance reliability and security by exploiting the complexities of ECG waveforms. Artificial Neural Networks (ANNs) are used to extract Individual Identification and a robust, accurate identification is performed. Current biometric methods are limited and this research tries to overcome the limitations of current biometric methods and offers secure access control methods. The fact is the need for secure, reliable authentication has increased dramatically as technology has penetrated every aspect of modern life. These include facial, voice and passwords, which are all becoming more and more useless because of the vulnerability involved: no facial data, twins can't be identified, voice recordings etc. They are a universal, robust, unique, stable, easy to collect and high-performance alternative to ECG signals [2]. ECG is the recording of the heart’s electrical activity over time, revealing vital information about its rhythm and functionality. The electrocardiogram (ECG) is a graphical representation of the heart's electrical activity over time, recorded using electrodes placed on the body. An ECG signal is characterized by a repeating waveform that includes distinct components such as the P wave, QRS complex, and T wave, each corresponding to specific phases of the cardiac cycle: the P wave represents atrial depolarization, the QRS complex corresponds to ventricular depolarization, and the T wave indicates ventricular repolarization [3]. Fig. 1 shows these key components which offer very important understanding about the heart’s condition and are crucial in looking at the heart’s rhythm and functionality . These features have significantly different morphology, and temporal characteristics, between individuals, which are sufficiently unique between individuals, such that ECG signals are unique [4]. Evaluated through electrocardiographs, ECG signals also reflect variations in sensor placement, enabling enhanced waveform analysis. The standard 12-lead configuration categorizes signals into chest and limb leads, capturing periodic and recurrent patterns, with the QRS complex, encompassing the Q, R, and S waveforms, serving as the focal point of ECG analysis, typically represented in the sequence P-QRS-T [5].

Figure 1. Sequence of depolarization and repolarization events in the heart and their correlation to various heart beat wave forms in the ECG signal

The distinctiveness of ECG signals lies in their inter-individual variability, which reflects the anatomical and physiological differences between individuals. Moreover, studies have confirmed that while ECG signals exhibit some day-to-day variability, their long-term intra-individual characteristics remain stable and comparable [6]. This stability allows for reliable identification over extended periods, even with a time gap of more than a year between recordings. ECG signals, being involuntary and directly linked to the functioning of the heart, are difficult to replicate or manipulate, further enhancing their suitability for biometric authentication [7].

Background

H. Silvaet. Silvaet. Al. [8] "For safe healthcare systems like HIS, defining how people can use ECG biometrics to authorize access and show medical data is key. Confirming a patient's insurance and reviewing their health record are two tasks that need to be done right at the start as they form the basis for everything else. Today, biometrics is the main way to solve this issue, but regular ID check methods only confirm limited access - even though technology advances, we still need direct contact with readers confined to a single area. Current patient and doctor sign-in methods in Healthcare Information Systems aren't reliable enough and result in lower care quality since mismatched data from everywhere keeps causing errors. Study by Al. [6] explores the idea of employing the ECG heartbeat's shape and deadlines to tell people apart. These ECG readings, taken from 22 healthy participants' lead I channels during various states, help validate how well the system works by analyzing 550 test samples. The proposed method takes ECG recordings from beginning of QRS complex to T wave termination. In this paper, we developed plan for making ECG readings an essential biometric security element. Through tests on a group of 22 healthy recipients, three different access methods were created, and analysis confirmed ECG variations do offer real help in securing ECG signals as a biometric solution. Research shows that the QRS part of ECG stays constant regardless of heart rate and works well just to verify identity, making it a useful component for biological safety [9].

According to Y. Ho, Wang et. Al. [10], "Biometric Identification" identifies individuals through their unique physical and biological signal characteristics. This project looks at and assesses a way to study the ECG from one lead to tell different humans apart. When the first phase ends, the ECG data is broken down into several sections for processing, one window for each ECG heartbeat. Through identifying QRS correctly, our tool collects important data values to help with identifying who is who. Research work provides a biometric system for the structured study of a particular electrocardiogram of lead (ECG) of human authentication. The initial stage of such a system consists of a wide band-pass filter that used to noise removal as well as other artefacts produced from raw ECG signal. A. Krishnapuramet. Al. [11] "In the study "A Bayesian Method to Combined Feature Selection and Classifier Design", the authors propose using Bayesian analysis to determine the point when a classifier with fewer features becomes less accurate while also finding the most helpful factors to the classification task. The method uses high-profile importance to encourage sparsity in use of both theory capabilities and modules; such prior convictions assume the shape of regularizations for the possibility research which awards considerable clarity in the planning of knowledge. Researchers derive an expectation-maximization (EM) algorithm to effectively calculate the maximum a posteriori (MAP) point estimation of a various variables. The algorithm uses one of today's leading public algorithms that are essentially the same as support vector machine equivalents, but based on Bayesian theory. T. The author Jebaraet, man. Al. [12]  developed two techniques for SVMs: selecting kernels to deal with multiple tasks and training individual SVMs on separate but connected datasets. A technique that is beneficial if several electronic classification methods and individually marked datasets occur against even a shared input space. Distinguishable datasets will typically reinforce the traditional choice of portraits or greatly strengthen with the classification techniques. A multi-task recognition learning technique using the very extreme entropy segregation formula is identified. The consequent convex algorithms retain the global solution objectives of support vector machines. Even then, in relation to several SVM classification and regression parameters, they often collectively calculate an optimum set of attributes and an optimal kernel combination. Tests are seen in simplified datasets. Andrea Bakker and her team showed how making separate regression tasks work together improved task modeling in their 2022 article [13]. To solve this in machine learning, we use a technique called multitask learning, where the network learns output from multiple related tasks at the same time. This is often achieved by a linear mixed effects model where there is a distinction between 'set effects, which are the same for all tasks,' and 'random effects, which can differ among tasks. In this paper, we will follow a Bayesian method in that few of the parameters are distributed (the same across all tasks) and few are very closely linked across a common probability distribution which can be derived from results. Throughout this manner, they try to incorporate the better aspects of both multi-level statistical approach as well as the neural network machinery. Article [14] “Geometrical dimensions of the differences of multi-lead ECG recordings. Electrocardiogram (ECG) is used as a clinical tool for evaluating or assistance a diagnosis in a cardiac patient ever since it was recorded in a man by Waller and then enhanced by Einthoven [15]. Normal ECG readings for healthy people show big differences between each person's heart patterns. These differences come from both the body's external structure and its internal electrical activity [16]. This study measures how much geometric factors influence the ECG's electrical signals. This study also details how to adjust for these elements. The study finds that most changes in ECG readings come from variations in how blood and tissues flow through your chest. Researchers have improved ECG-based biometric authentication methods, but there's still more to study and refine. Modern systems have problems with noise interference, struggle to handle differences in human responses, and require advanced ways to pull useful data from ECG signals [17]. The system we propose uses deep learning to solve these issues, making it more reliable and precise at verifying people [2] [18].

Method

Our proposed system combines multiple parts as shown in Fig. 2: signal processing, feature picking, wavelet analysis, and an Artificial Neural Network (ANN) biometric comparison tool.

Figure 2.  Proposed block diagram

This deep learning method allows our system to uncover complex wave patterns in ECG signals, making it better at both recognizing and interpreting heart activity. The way the ECG signals are processed both inside the system and averaged out helps it detect heart issues better by ignoring unwanted background noise.

Data Acquisition

The ECG data used in this research was obtained from the ECG-ID dataset, a publicly available and widely recognized dataset for biometric authentication research. The dataset includes ECG signals from 90 subjects, providing a diverse and comprehensive range of signals for analysis. For this study, 40 subjects were selected, ensuring diversity in terms of age and gender to enhance the robustness and generalizability of the system.

Signal Pre-Processing and Feature Extraction

The first step was to apply a median filter to raw ECG signals, reducing noise and enhancing quality. Key features like the P-wave, QRS complex, and T-wave were then extracted, representing unique cardiac cycle phases for individual identification. Wavelet decomposition analyzed the signals in both time and frequency domains, isolating bands to highlight authentication-relevant details.

Wave Analysis and Modelling

The QRS complex analysis included detection of waves followed by modeling and feature extraction steps while comparing against a reference template using distance thresholds and deviation and averaging thresholds as robustness measures.

Classification and Output Generation

The refined features were processed using an Artificial Neural Network (ANN) classifier, trained on a dataset of known ECG signals. The ANN utilized intricate patterns within the signals to authenticate individuals. The final system output provided a binary decision, indicating whether the input ECG signal was authentic or not. This decision was derived based on the results of the averaging threshold and ANN classification, showcasing the system's ability to deliver accurate and reliable biometric authentication.

Methodology

Fig. 3 shows the block diagram of the proposed approach. It consists of pre trained ANN classifier for the performance evaluation of ECG signal classification for person authentication. In the proposed methodology the steps as follows.

Figure 3. Proposed methodology

The collected ECG signals were retrieved from a pre-existing database and subjected to a series of pre-processing steps to enhance their quality. Initially, the signals were filtered to remove high-frequency noise and artifacts that could interfere with subsequent analysis. Pre-processing was carried out using a Band Pass Filter (BPF), which allows signals with frequencies below a specified cut-off to pass while attenuating those with higher frequencies. This technique ensures that only the relevant frequency range of the ECG signal is preserved for further processing. Following pre-processing, the filtered signals were prepared for input into the Artificial Neural Network (ANN) classifier. During this step, the layers of the ANN classifier were carefully configured, with all layers concatenated except for the first and the last three, optimizing the model’s ability to analyze and classify the ECG signals effectively.

Results

Fig. 4 represents the raw ECG signal acquired from the individual. The ECG signal typically consists of P, Q, R, S, and T waves, each corresponding to different phases of the cardiac cycle. The irregularities and unique patterns in this signal form the basis for biometric authentication.

Figure 4.  Input ECG Signal

During acquisition, the ECG signal is frequently contaminated by various types of noise. After removing unwanted noise from the ECG signal using Band pass filter (BPF), the ECG signal is displayed in Fig.5. It is this step that significantly improves the quality and accuracy of following signal processing and analysis.

Figure 5. Noise removed signal via BPF

PQRS points represent the key landmarks in the ECG signal, namely the P-wave, QRS complex, and T-wave. Fig. 6 illustrates the identified PQRS points, providing a visual representation of the critical features used for subsequent analysis and feature extraction.

Figure 6. PQRS points representation

The accurate detection of R-peaks within the QRS complex is very important. Specifically detected R peaks in the ECG signal, representing a foundation for subsequent processing steps are shown in Fig. 7.

Figure 7. Detected R in ECG Signal

To mitigate the impact of noise and irregularities, a smoothing process is applied to the ECG signal. Fig. 8 shows the signal after smoothing, with a clearer underlying cardiac activity.

Figure 8.  Smoothed Signal

The system achieved 98% accuracy in identifying individuals, with 95% success in recognizing authorized users. However, it struggled to distinguish unauthorized users, rejecting them only 20% of the time and showing 75% false negative accuracy. This highlights a key challenge in differentiating positive and negative cases. The model performed within 10 seconds, meeting current operational speed requirements. Figure 9 summarizes the accuracy, sensitivity, and selectivity results.

Figure 9.  Performance Metrics

The confusion matrix from Fig. 10 helps us assess how well the system works. Looking at the matrix, we find that the system correctly classified 95% of authentic face images (True Positives) and 20% of non-authentic ones (True Negatives).

Figure 10. Confusion Matrix

The Receiver Operating Characteristic (ROC) curve is plotted as shown in figure 11 to evaluate the trade-off between sensitivity (True Positive Rate) and specificity (False Positive Rate). The curve showed an Area Under the Curve (AUC) of 0.98, indicating excellent performance in distinguishing between authentic and non-authentic individuals. However, the curve also revealed that the system's performance drops at higher specificity levels, which aligns with the low selectivity observed in the performance metrics.

Figure 11. ROC Curve for ECG-Based Authentication System

Conclusion

A new biometric security system uses ECG signals, advanced signal processing, and deep learning to create a robust, accurate identification method. It improves on current biometrics by combining signal modification, feature extraction, and neural networks for better real-world performance. The system uses previous research findings but it solves existing challenges by seeking larger validation datasets. The system will undergo different security tests that utilize deep learning and signal processing techniques to manage erratic input data. The framework will evolve to address future challenges, integrate with emerging technologies, and maintain cutting-edge performance in biometric authentication. Ongoing research is key to advancing security standards and meeting reliable identification needs.

References:

1. Ingale M. et al. ECG Biometric Authentication: A Comparative Analysis // IEEE Access. ieee institute of electrical electronics engineers, 2020. Vol. 8. P. 117853–117866.
2. Ahmed M.J. et al. CardioGuard: AI-driven ECG authentication hybrid neural network for predictive health monitoring in telehealth systems // SLAS Technology. sage, 2024. Vol. 29, № 5. P. 100193.
3. Liu J. et al. A novel P-QRS-T wave localization method in ECG signals based on hybrid neural networks // Computers in biology and medicine. n tab tab pergamon n tab tab n tab tab elmsford ny usa n tab, 2022. Vol. 150. P. 106110.
4. Schwerdtfeger A.R. et al. Heart rate variability (HRV): From brain death to resonance breathing at 6 breaths per minute // Clinical Neurophysiology. elsevier ireland, 2019. Vol. 131, № 3. P. 676–693.
5. Sohn J. et al. Reconstruction of 12-Lead Electrocardiogram from a Three-Lead Patch-Type Device Using a LSTM Network. // Sensors. molecular diversity preservation, 2020. Vol. 20, № 11. P. 3278.
6. Qu H., Pang L., Gao X. Classification of mental workload based on multiple features of ECG signals // Informatics in Medicine Unlocked. elsevier, 2021. Vol. 24. P. 100575.
7. Xiong P. et al. Short-term paroxysmal atrial fibrillation detection with intra- and inter-patient paradigm based on R-R intervals // Biomedical Signal Processing and Control. american geophysical union, 2023. Vol. 89. P. 105750.
8. Da Silva H.P. et al. Finger ECG signal for user authentication: Usability and performance. institute of electrical electronics engineers, 2013. Vol. 2. P. 1–8.
9. El Boujnouni I. et al. A wavelet-based capsule neural network for ECG biometric identification // Biomedical Signal Processing and Control. american geophysical union, 2022. Vol. 76. P. 103692.
10. Meltzer D., Luengo D. Efficient Clustering-Based electrocardiographic biometric identification // Expert Systems with Applications. elsevier, 2023. Vol. 219. P. 119609.
11. Wagner P. et al. Explaining deep learning for ECG analysis: Building blocks for auditing and knowledge discovery // Computers in Biology and Medicine. n tab tab pergamon n tab tab n tab tab elmsford ny usa n tab, 2024. Vol. 176. P. 108525.
12. M S., Gupta N. A Survey on Biometric Authentication Using Electrocardiogram // SSRN Electronic Journal. elsevier, 2020.
13. Yang W., Wang S. A Privacy-Preserving ECG-Based Authentication System for Securing Wireless Body Sensor Networks // IEEE Internet of Things Journal. institute of electrical electronics engineers, 2022. Vol. 9, № 8. P. 6148–6158.
14. Khan M.U. et al. Biometric Authentication System Based on Electrocardiogram (ECG). institute of electrical electronics engineers, 2019. P. 1–6.
15. Issa M.F. et al. Heartbeat classification based on single lead-II ECG using deep learning // Heliyon. elsevier bv, 2023. Vol. 9, № 7. P. e17974.
16. Van Dam P.M. et al. The relation of 12 lead ECG to the cardiac anatomy: The normal CineECG // Journal of Electrocardiology. elsevier, 2021. Vol. 69. P. 67–74.
17. Uwaechia A.N., Ramli D.A. A Comprehensive Survey on ECG Signals as New Biometric Modality for Human Authentication: Recent Advances and Future Challenges // IEEE Access. ieee institute of electrical electronics engineers, 2021. Vol. 9. P. 97760–97802.
18. Wang D. et al. A Novel Electrocardiogram Biometric Identification Method Based on Temporal-Frequency Autoencoding // Electronics. mdpi ag, 2019. Vol. 8, № 6. P. 667.