A Session Drop Is Not a Signal Ceiling: What Cross-Session Degradation in ECG Biometrics Measures, Why the Same Heartbeat Can Link a Patient Across Years Yet Fail to Unlock a Device, and the Longitudinal Protocol That Would Locate the Loss
(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Electrocardiogram (ECG) biometrics carries two reputations at once. In the authentication literature the heartbeat is a fragile identifier whose accuracy collapses once enrolment and probe come from different sessions; a 2026 benchmark across seven public datasets names the cause the Random Split Fallacy and concludes that the remaining failures are probably inherent to current supervised feature learning. In the privacy literature the same signal is a fingerprint robust enough that sharing a recording exposes the person who produced it. In this analysis the author reviews and synthesises published results to show that both reputations are partly right and that neither is stated in a form the evidence can test. The cross-session drop the benchmark reports is real, but it bundles four quantities: a step at the session boundary (electrodes re-attached, device or posture changed), a shift in physiological state, a slow drift with elapsed time, and a failure to generalise to identities absent from training, all read through metrics whose values also depend on gallery size. The benchmark's evidence that these losses are inherent comes from three architectures trained as closed-set classifiers on one 90-subject dataset. On the same public splits, embedding models trained on tens of thousands of identities lose much less: on the CYBHi three-month split, published verification error ranges from 19.7 percent for the benchmark's baseline to 3.9 percent for a model trained on an exercise-test cohort (and lower in one study whose protocol differs), and one fusion choice inside a single study moves cross-session error from 16.4 to 5.6 percent while leaving same-session error within 0.7 points. The privacy reputation survives a stricter test than the one most re-identification studies apply: the strongest re-identification figures split a single recording in time, yet a closed-set study on two clinical corpora with galleries of about 14,500 patients still ranks the right patient first about 56 to 64 percent of the time across a five-year gap, while the matching acceptance rate at a false-accept rate of 1 in 1,000 falls to 0.66 to 0.74. The author consolidates the measured values in Table 1 and Figure 1, separates the four components in Table 2, states a procedure for reading a cross-session claim as Algorithm 1, and specifies five measurements that would settle which part of the loss belongs to the signal and which to the learner.
Authors
- Pranay M. Mahendrakar (ORCID: https://orcid.org/0009-0003-7224-029X)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24209614
- Primary Topic
- Biometric Identification and Security
- Type
- preprint