ConvWaveFormer: a hybrid wavelet and CNN-transformer framework for ECG biometric identification
Year: 2026
Type: article
Abstract: Electrocardiogram (ECG)-based biometrics have shown strong potential as an alternative authentication modality because ECG signals contain subject-specific cardiac morphology and provide inherent liveness information. However, recognition performance remains sensitive to signal noise, morphological variability, and detector-dependent R-peak localization errors, which may misalign segmented heartbeats and reduce feature consistency. This study proposes ConvWaveFormer , a hybrid wavelet and CNN-Transformer framework for ECG biometric identification. Following morphology-preserving preprocessing, fixed-length R-peak-centered heartbeats are processed through two branches. A Daubechies-4 wavelet branch captures multiresolution time-frequency information, while a residual 1D-CNN-Transformer branch learns local ECG morphology and within-heartbeat temporal dependencies. The deep and wavelet representations, together with their element-wise interaction, are fused to generate a compact biometric embedding. Margin-based supervision using ArcFace and CosFace further promotes intra-subject compactness and inter-subject separability. The framework was evaluated on ECG-ID, Heartprint, MIT-BIH Arrhythmia, and MIT-BIH Normal Sinus Rhythm using three R-peak detection and heartbeat segmentation methods: NeuroKit2, SciPy FindPeaks, and XQRS. Although the relative performance of the detectors varied across datasets, XQRS achieved the strongest or comparable overall performance in most cases and was therefore selected for the final five-fold evaluation. ConvWaveFormer achieved mean identification accuracies of 92.99%, 76.47%, 97.70%, and 99.63%, with corresponding EER values of 5.84%, 11.11%, 1.72%, and 0.34%, respectively. In cross-dataset evaluation, a three-source score-level ensemble achieved 87.35% Rank-1 accuracy, 6.09% EER, and 0.9808 ROC-AUC on ECG-ID without target-domain model retraining. These results demonstrate that ConvWaveFormer learns discriminative and transferable ECG representations across heterogeneous offline datasets. Nevertheless, strictly session-disjoint, cross-device, and continuous wearable evaluations remain necessary to establish robustness under unconstrained real-world conditions.
Source: Scientific Reports
Institutions: Jeju National University, Recep Tayyip Erdoğan University
Language: English
Topic: ECG Monitoring and Analysis
Subfield: Cardiology and Cardiovascular Medicine
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,690
Award: G4906405427
OpenAlex ID: W7221327617