Abstract
In recent years, Electrocardiogram (ECG) signals have emerged as a promising modality for biometric recognition, attracting significant attention. In this work, a new ECG biometric identification approach based on Spatial Pyramid Pooling (SPP) and Autoencoder (AE), referred to as the SPP-AENet, is proposed. The SPP-AENet is a hybrid architecture consisting of two parts: the SPP layer and the AE layer. The SPP layer allows non-fixed length input and, by incorporating prior knowledge, ensures that critical characteristics of ECG signals are retained while compressing redundant information. The AE layer, through its automatic feature learning mechanism, discovers latent patterns and relationships within the data. The proposed method is evaluated using three public databases acquired under different health conditions, yielding signal identification accuracies of 98.88%, 100%, and 100% for the ECG-ID, PTB, and ST-Change databases, respectively. Features of the SPP and AE layers are visualized using t-SNE to facilitate a better understanding of the model learning process.
Introduction
Biometric identification is evolving beyond traditional modalities like fingerprints and facial recognition. ECG signals, being internal, are inherently difficult to counterfeit and provide built-in "liveness" detection. However, two major hurdles persist: temporal variability (heart rate changes) and the "black box" nature of deep learning models. Most existing models require fixed-length inputs, often leading to information loss during signal cropping. SPP-AENet addresses these by handling variable-length signals and improving interpretability through a hybrid architecture.
Methodology
The SPP-AENet architecture is designed to respect the physiological structure of heartbeats while leveraging unsupervised feature learning. The process begins with denoising raw ECG signals using wavelet transforms and segmenting individual heartbeats without forcing them into fixed lengths.
The Identification Stage
The core of the system is a hybrid pipeline consisting of an SPP layer followed by an AE layer, concluding with a majority voting mechanism for final identification.
Spatial Pyramid Pooling (SPP) Layer
The SPP layer incorporates domain knowledge by treating the QRS complex as the most critical component. While the QRS fragment is kept intact, surrounding fragments (containing P and T waves) are processed via multi-scale pooling (1×1, 1×4, and 1×16). This allows the network to accept variable-length inputs and produce a standardized feature vector without losing salient morphological features.
Autoencoder (AE) Layer and Classification
The SPP output is fed into a stacked Autoencoder. Pretrained to reconstruct inputs, the AE layer learns discriminative latent patterns. A final softmax layer performs classification, and a majority voting scheme across multiple heartbeats ensures robust identification.
Results and Key Insights
The model was evaluated on three diverse Physionet databases:
- ECG-ID: 89 healthy subjects (98.88% accuracy).
- PTB: 290 subjects with mixed health conditions (100% accuracy).
- ST-Change: 28 subjects with ST-segment depression (100% accuracy).
SPP-AENet achieved over 97% accuracy with just five heartbeats, suggesting it could be used for rapid identification on wearable devices. It outperformed several complex models, including LSTMs and Capsule Networks, particularly on the larger PTB dataset. Visualization using t-SNE confirmed that while the SPP layer initiates subject grouping, the AE layer effectively untangles features to create clear decision boundaries.
Conclusion
By integrating biological prior knowledge—specifically protecting the QRS complex—SPP-AENet bridges the gap between raw biological signals and deep learning requirements. This approach handles the "accordion effect" of heart rate variability while maintaining high accuracy across diverse health conditions. Future work will focus on robustness against daily biological noise and adaptation for lower-resolution consumer-grade wearables.
Reference: Liu, X., Wang, D., et al. (2025). SPP-AENet: A New ECG Biometric Identification Approach Based on Spatial Pyramid Pooling and Autoencoder. Journal of Advances in Information Technology, Vol. 16, No. 1.
Summary updated