Learning Cardiac Features: ECG Biometrics Across Time and~Exercise
Luca Thiebaud, Paul Chauchat, Mustapha Ouladsine, Stéphane Delliaux
cs.AI · q-bio.TO
Quick verdict
Claims first extensive ECG-biometrics test under exercise and time drift; sound protocol and strong EERs, but private data and no code limit verification.
Summary
The paper evaluates ECG-based biometric authentication under two stressors rarely tested together: exercise and time. The authors train a Siamese 1D ResNet-18 on beat-level pairs from an in-house cardiopulmonary exercise test (CPET) dataset of 1,651 patients (128 with at least two sessions; median inter-session interval 346 days), using patient-level splits to prevent leakage. The proposed late fusion — averaging match scores of eight independently trained single-lead Siamese networks — reaches 1.0% intra-session and 5.6% cross-session EER on CPET, while early fusion degrades from 1.7% to 16.4%. Phase-specific analysis on the 64 multi-session test patients shows rest-to-peak intra-session EER of 1.7% (95% CI 0.8-2.8), with cross-session peak comparisons rising to 7.8-8.1%. Without fine-tuning, replicating the ECGXtractor benchmark protocol, the method reports 2.1% EER on PTB (matching the baseline), 3.9% on CYBHi (versus 8.0), and 10.0% on Heartprint (versus a single prior 53.9% result). Main limitations: the dataset and trained models are not released, so headline claims cannot be independently verified; bootstrap confidence intervals cover only the 64-patient phase analysis, where cross-session CI widths exceed 5 points; and per-phase beat-retention rates after quality filtering are unreported, leaving a possible selection bias toward rest conditions unquantified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Late multi-lead score fusion of independent single-lead Siamese ResNets cuts cross-session EER to 5.6% versus 16.4% for early fusion on the CPET dataset (Fig 4).
- ECG biometric authentication stays accurate under exercise: rest-to-peak intra-session EER is 1.7% (95% CI 0.8-2.8) on the 64 multi-session test patients (Fig 5).
- Without fine-tuning, the method achieves 3.9±0.9% EER on CYBHi, 2.1±0.3% on PTB (matching ECGXtractor), and 10.0±1.0% on Heartprint (Table 3).
- The results support an intrinsic cardiac signature resilient to physiological and temporal drift (Conclusion).
Strengths
Leakage-resistant evaluation protocol
Patient-level splits (64/64 multi-session, 85/15 single-session, Sec 5.1) directly fix the train/test patient overlap the authors document in identification-protocol literature (Sec 2.2), making the cross-session numbers methodologically meaningful.
Clean and consequential fusion ablation
Fig 4: late score fusion achieves 1.0%/5.6% intra/cross-session EER vs 1.7%/16.4% for early fusion — early fusion even falls below the best single lead (Lead II, 13.5% cross) — evidencing that joint lead processing overfits session-dependent inter-lead correlations.
First large-cohort authentication analysis on the exercise axis
Benchmark-compliant cross-dataset evaluation without fine-tuning
Candid reporting of degradation
Weaknesses
No code, model, or data release
Ambiguous training data for public-benchmark results
Unquantified selection bias from the quality filter
Thin statistical support for headline comparisons
Under-justified configuration choices
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.21962) studies ECG-based biometric authentication (1:1 verification) under two axes of variability the literature largely treats separately: exercise-induced physiological change and cross-session temporal drift. Three claims anchor the abstract and Section 1: (1) a Siamese ResNet with late multi-lead fusion is more robust across sessions than early fusion; (2) this is the first extensive evaluation of ECG biometric stability under exercise stress; (3) the method reaches state-of-the-art authentication results on public benchmarks without fine-tuning. The headline numbers are an intra-session rest-to-peak EER of 1.7% and a 5.6% cross-session EER on a private CPET cohort. Claims 1 and 3 are directly supported by the reported experiments; claim 2 rests on a definition of 'extensive' (cohort scale plus authentication protocol) that Section 2.3 legitimately earns by contrasting with prior 20-70-subject identification-only studies, though the abstract states it without that hedge.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What error rates does this ECG biometric authentication method achieve?
On its own CPET dataset it reports 1.0% intra-session and 5.6% cross-session EER with late fusion (early fusion: 1.7% and 16.4%). On public benchmarks without fine-tuning it reports 2.1% EER on PTB (tying the ECGXtractor baseline), 3.9% on CYBHi (baseline 8.0%), and 10.0% on Heartprint.
Does exercise break ECG-based biometric recognition?
Not according to this paper. Comparing rest beats against peak-exercise beats within the same session yields only 1.7% EER (95% CI 0.8-2.8) on 64 patients, far better than the large drops reported in earlier studies with 20-70 subjects. Across different sessions, however, peak comparisons degrade to roughly 8% EER.
Is the code or the CPET dataset publicly available?
No. The paper contains no code or data availability statement. The CPET data are clinical exercise-test records collected at a Marseille hospital between 2010 and 2024, so headline results cannot currently be reproduced by third parties.
What is 'late fusion' in this paper?
Instead of feeding all ECG leads into one network, the authors train eight separate Siamese ResNet networks, one per lead (I, II, V1-V6), and average their verification scores. This design halves cross-session error relative to processing all leads jointly (5.6% vs 16.4% EER).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are images; ROC curve shapes and Fig 2 waveform morphology cannot be verified from extracted text, though the key numeric values are present
- Text extraction interleaves figure captions with axis labels in places (e.g., Fig 5 region), so some values were matched to labels carefully rather than trivially
- Whether supplementary material or an appendix exists is unknown from this extraction
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.