1D Physiological Signals: ECG and PPG

The 1D Physiological Signals: ECG and PPG project investigates computational-intelligence and deep-learning methods for analysing cardiac signals acquired through electrocardiography and photoplethysmography. The research focuses particularly on biometric recognition and continuous authentication, while also considering wearable monitoring, body-area networks, mobile devices, and integrated health-and-security applications.

ECG records the electrical activity generated by the heart, whereas PPG measures blood-volume variations through optical sensors. Both signals contain characteristics that differ among individuals and can therefore support identity recognition. ECG also provides intrinsic liveness information, is difficult to reproduce artificially, and can potentially combine biometric authentication with health monitoring.

PPG offers complementary advantages because it can be acquired continuously through inexpensive and comfortable sensors already integrated into smartwatches, pulse oximeters, fitness devices, and mobile platforms. This creates opportunities for low-cooperation and continuous authentication without requiring users to perform an explicit biometric acquisition.

An early research direction studied correlation-based recognition using ECG and PPG waveforms. The methods analyse stable cardiac patterns over time and compare signals acquired during enrollment and verification. Experiments showed that PPG contains sufficient subject-specific information for authentication scenarios that do not require extremely high accuracy, and that continuous enrollment can improve performance by progressively updating the user model.

The project also introduced HeartCode, a compact binary representation of ECG signals. The method extracts QRS complexes, transforms them using one-dimensional Log-Gabor filters, and converts the resulting phase information into binary strings that can be compared using the Hamming distance.

HeartCode templates require approximately 150 bits in the reported configuration, making them suitable for wearable devices, body-area networks, and cryptographic template-protection schemes. On 8,400 samples from Holter recordings acquired under uncontrolled conditions, the method achieved an equal error rate of 8.58%. Matching time was reduced from approximately 112 microseconds for a correlation-based approach to about 1 microsecond using the binary representation.

A later research direction introduced Deep-ECG, a convolutional neural-network architecture for learning discriminative representations directly from cardiac signals. The original approach processes selected QRS complexes, exploiting the fact that this portion of the ECG is comparatively stable over time and contains subject-specific morphological information.

The most recent development simplifies the pipeline by removing heartbeat detection and QRS selection. Minimally processed raw ECG segments are fed directly to the network, reducing computational complexity and allowing the model to exploit not only heartbeat morphology but also temporal and heart-rate-variability information.

This revised Deep-ECG approach was evaluated on 24-hour Holter recordings from 139 individuals acquired under uncontrolled conditions. With two-second segments, it achieved accuracies above 80% using individual leads and above 90% when multiple leads were combined. Around 100 training samples per person were sufficient to exceed 80% accuracy in the investigated configuration.

The experiments also showed that combining multiple ECG leads generally improves identification because each lead provides a different view of cardiac electrical activity. The minimally preprocessed model obtained performance comparable to the original, more complex Deep-ECG pipeline while being more suitable for real-time, wearable, and edge-device implementation.

Another major contribution is MultiCardioNet, a Siamese convolutional neural network designed to compare spectrogram-based representations of ECG and PPG signals. Unlike conventional biometric systems that require enrollment and verification using the same sensing modality, MultiCardioNet supports ECG-to-ECG, PPG-to-PPG, and cross-modal ECG-to-PPG comparisons.

This interoperability is relevant when the user is enrolled through a clinical ECG device but later authenticated through a wearable PPG sensor, or vice versa. Apart from modality-specific signal enhancement, the same feature extraction and matching architecture is used for both signal types.

MultiCardioNet uses a specifically designed lightweight Siamese architecture rather than a large pretrained image model. An adaptive training strategy iteratively selects informative genuine and impostor pairs, addressing the class imbalance and limited sample sizes typical of physiological-signal datasets.

The method was evaluated using 1,008 ECG samples and 1,008 corresponding PPG samples from 42 individuals. It achieved equal error rates of 2.15% for ECG-to-ECG verification, 2.42% for PPG-to-PPG verification, and 4.76% for heterogeneous ECG-to-PPG or PPG-to-ECG comparisons. The relatively limited decrease in the cross-modal scenario demonstrated the feasibility of interoperable cardiac biometrics.

The project also provides a broader analysis of PPG biometrics, covering acquisition devices, signal-processing techniques, handcrafted and deep features, public datasets, application scenarios, and unresolved research challenges. This work highlights the growing role of PPG in continuous authentication enabled by wearable and mobile technologies.

Important open issues include signal quality, motion artefacts, sensor heterogeneity, temporal stability, physiological changes, dataset size, cross-session evaluation, privacy protection, and the integration of biometric recognition with health monitoring. The project addresses several of these limitations through compact representations, deep feature learning, adaptive enrollment, and interoperability between sensing modalities.

Overall, the research establishes a progression from correlation-based ECG and PPG recognition to compact binary templates, deep identification models, and multimodal cardiac biometrics. Its main contribution is a family of efficient and increasingly flexible methods for authenticating individuals from physiological signals acquired by medical, wearable, and edge devices, with possible applications in continuous access control, telemedicine, personal health systems, and secure body-sensor networks.