- The paper introduces a novel ArcFace-Inception method for ECG biometrics using 512-dimensional, L2-normalized embeddings, validated on large public datasets.
- It employs comprehensive stress tests—including scale, temporal drift, and cross-domain evaluations—to reveal systematic performance degradation.
- The study highlights effective post-hoc reranking with AS-norm and robust confidence calibration, emphasizing challenges for real-world deployment.
Large-Scale, Multi-Domain ECG Biometrics with ArcFace-Inception: External Validation and Stress Testing
Introduction
This work rigorously evaluates the operational limits of ECG biometrics by establishing a tightly controlled, reproducible large-scale benchmarking framework. A 1D Inception-v1 backbone trained with ArcFace loss is externally validated on MIMIC-IV-ECG and HEEDB, currently the largest public corpora for ECG biometrics. The study’s core contribution is the comprehensive closed-set stress testing of deep metric embeddings, embracing both domain shift and long temporal intervals. The methodology and results set a new bar for analytical rigor in this subfield.
Methodological Framework
The model architecture integrates a 1D Inception-v1 network with an ArcFace head during training, yielding 512-dimensional, L2-normalized ECG embeddings. Following evidence from prior work (2602.02776), ArcFace is selected due to its stable and discriminative behavior under large gallery sizes—a scenario consistently neglected in prior literature, where pairwise and triplet losses often falter in optimization and retrieval calibration.
Training utilizes the ASUGI-DB corpus, comprising 164,440 ECGs from 53,079 patients, filtered for recording quality, temporal separation, and device uniformity. Evaluation is conducted both internally and on two large public, multi-domain datasets: MIMIC-IV-ECG (231,329 ECGs, 63,895 patients) and HEEDB (385,079 ECGs, 118,756 patients). Derived evaluation datasets and protocols are explicitly stratified to enforce either cohort disjointness or controlled stratification, preventing patient overlap and ensuring strict external validation.
Signal preprocessing and normalization are standardized to z-score per channel and 500 Hz sampling across all datasets, optimizing computational efficiency and harmonization with external corpora.
Benchmark Protocols
Four major protocols structure the evaluation:
- General Comparability: Leave-one-out closed-set identification, testing every ECG as a query against all remaining ECGs from other patients.
- Scale Analysis: Systematic grid search varying gallery sizes (500–7,000 identities) and examinations per patient (2–7), quantifying identification degradation as scale increases.
- Temporal Stress Test: Fixed-size galleries with temporal gaps (1–5 years) between enrollment and probe ECGs, isolating the effect of longitudinal drift.
- Post-hoc Reranking: Performance augmentation using score normalization strategies (AS-norm, S-norm, Z-norm, T-norm, C-norm), graph-based diffusion, and (α-)query expansion—analyzed for both identification improvement and confidence calibration.
Confidence estimates of rank-1 matches are assessed by a logistic regression calibrator fed with leading cosine scores, yielding robust domain-dependent calibration and selective prediction analyses as recommended in ISO/IEC 19795-1:2021.
Main Results
Generalization Across Domains
Under the strictest closed-set comparability, the system achieves:
- ASUGI-DB: Rank@1 = 0.9506, TAR@FAR=1e-3 = 0.9675
- MIMIC-GC: Rank@1 = 0.8291, TAR@FAR=1e-3 = 0.8188
- HEEDB-GC: Rank@1 = 0.6884, TAR@FAR=1e-3 = 0.7210
Cross-domain performance is strongly conditioned by both acquisition heterogeneity (site, device, curation) and the size of the gallery. The substantial drop when transitioning from the internal to external validation demonstrates the fragility of results reported on homogeneous, small cohorts typical in prior studies.
Large-Scale and Temporal Robustness
Scale analysis reveals Rank@1 decays monotonically with gallery size (from 0.8761 at 500 subjects to 0.8107 at 7,000 in HEEDB), but increases systematically as more templates per patient are used (up to 0.8928 at seven exams/patient). The temporal stress test isolates the effect of time intervals, with Rank@1 on MIMIC decreasing from 0.7853 (1 year) to 0.6433 (5 years), and on HEEDB, from 0.6864 to 0.5560 over the same period. These outcomes quantitatively illustrate the identity information retained under severe operational regimes but also highlight substantial temporal vulnerability.
Reranking and Confidence Calibration
Among post-hoc re-ranking methods, AS-norm yields the strongest improvement, with Rank@1 reaching 0.8005 in HEEDB-RR for optimal parameters (N=4), compared to 0.7765 baseline. Other methods provide marginal or even detrimental effects, with graph-based diffusion notably degrading both retrieval and calibration.
Confidence analysis shows persistent discriminative separation of means (Δ up to 0.4882 on ASUGI-DB), but high-confidence prediction coverage degrades out-of-domain (from 88.05% in ASUGI-DB to 35.73% in HEEDB-GC at the 0.90 threshold). This underscores that while well-calibrated models can selectively predict in easier domains, sparseness and error of high-confidence matches rise in more complex or shifted cohorts.
Implications and Discussion
Empirical Rigor Over Literature Baselines: Performance metrics are lower than those often reported for ECG biometric systems, but prior studies frequently use insubstantial cohort sizes and ignore domain, device, and temporal factors. Here, gallery sizes reach up to 118,756, representing an order-of-magnitude increase compared to typical literature [melzi2023], with known and controlled sources of temporal and cross-hospital variation.
Limitations: Despite the robustness to scale and domain, these results stem from closed-set identification, not open-set or deployment scenarios. Asymmetric curation (i.e., the inability to perform the same metadata-driven filtering in external sets) confounds pure domain-shift attribution. Confidence calibrators remain domain-anchored, and it is unclear how they generalize across further distributional shift.
Future Directions: Prospective work should address open-set identification, enforce robustness to unseen-identity rejection, and explore stratification by clinical condition and pathology. The public availability of MIMIC-IV and HEEDB enables reproducibility and further extension, including longitudinal studies of disease-modulated biometric drift and experiments employing alternative deep or self-supervised representations.
Conclusion
This study demonstrates that ECG-based biometric identification with ArcFace-Inception remains feasible against hundreds of thousands of identities and retains quantifiable discriminative power even with multi-year temporal drift and substantial cross-domain shift. Performance degrades systematically with increasing operational challenge, but remains robustly above-chance. Reranking, especially via AS-norm, delivers moderate retrieval improvements, while confidence metrics reliably identify high-certainty matches primarily within well-curated, in-domain sets.
These results establish large-scale, external, temporally-aware benchmarking as essential for realistic ECG biometrics research, emphasizing the need for continued investigation into open-set validation, pathology-dependent drift, and robust calibration when deploying identity models in clinical or security-critical environments.
Reference: "ECG Biometrics with ArcFace-Inception: External Validation on MIMIC and HEEDB" (2604.04485)