Advances in electrocardiogram signal analytics for person identification using fine-grained fusion-driven deep representation learning under arrhythmic conditions

Introduction Electrocardiogram (ECG) signals measure the minute electrical signals produced throughout the cardiac cycle, whereas biometric signals are vital indicators of human activity. Recent studies highlight ECG as a very significant tool for clinical analysis and as a novel biometric modality. Because of differences in physical features like height, heart size, muscle mass, weight, and fat mass between individuals, each person exhibits unique characteristic signals. ECG-based identification has numerous applications, including unlocking electronic doors, banking transactions through smartphones, and facilitating website login. The usage of image-based representation allows recognized computer vision structures to exploit both rhythm-related and morphological cues, enhancing discrimination across individuals. In recent times, deep learning and machine learning have been progressively applied to ECG signals for a wide range of tasks such as person identification and arrhythmia detection. Deep learning approaches like Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNNs), and Transformer based systems can learn hierarchical representations directly from raw or slightly processed ECG waveforms, capturing both temporal dynamics and morphological details of heartbeats. Methods This paper proposes FDDR-ID (FusionDriven Deep Representation Learning for ECG-Based Person Identification), a multitask deep learning framework designed to achieve robust identification performance under arrhythmic conditions. The proposed model initially preprocesses the ECG signals through three sub-processes: non-local means (NLM) filtering-based denoising, R-peak detection, and segmentation. To exploit the supremacy of deep learning approaches, every pre-processed ECG segment is converted into a two-dimensional representation using the shorttime Fourier transform to generate spectrograms. For feature extraction, a fusion of three deep learning models is used: SqueezeNet, NASNet, and EfficientNet. Moreover, a Siamese LSTM Autoencoder is utilized for the classification process, which includes three output branches: one each for person identification, age prediction, and gender classification. The model is trained with adaptive learning using the Nadam optimizer, ensuring faster convergence and improved generalization. Results and Discussion Performance validation is conducted on the proposed FDDR-ID model utilizing the ECG-ID Database and the MITBIH Arrhythmia Database. Extensive comparative analyses demonstrate that the presented method outperforms other state-of-the-art approaches. The proposed FDDR-ID method also shows important potential for real-world deployment in applications like smart authentication systems, wearable healthcare devices, remote patient monitoring, and secure biometric access platforms.