Predicting When an ECG Biometric Identity Becomes Unreliable: An Uncertainty-Aware Framework for Adaptive Re-Enrollment

Electrocardiographic (ECG) biometric verification can lose reliability over time because ECG morphology changes across longitudinal sessions, which makes enrolled templates increasingly likely to reject genuine users. Most existing methods focus on maximizing recognition accuracy or updating templates only after new data arrive, and they do not try to forecast how reliable a future verification will be before it happens. This work investigates whether information available before a verification attempt can be used to predict the reliability of future genuine verifications, thereby enabling uncertainty-aware decisions and adaptive re-enrollment. Experiments used the longitudinal Heartprint dataset, with earlier sessions reserved for model development and later sessions (S3R and S3L) held out for prospective evaluation. A one-dimensional ResNet produced 128-dimensional ECG embeddings that were compared via cosine similarity. Logistic Regression and XGBoost then predicted future reliability from temporal features, historical similarity measures, and signal-quality indicators. The framework also examined feature ablation, probability calibration, abstention, and adaptive re-enrollment. Under a fixed development threshold, AUROC fell from 0.971 in S2 to 0.936 in S3R and 0.916 in S3L, while the false-rejection rate rose from 9.27% to 16.99% and 20.94%, respectively. Logistic Regression showed stronger discrimination, whereas XGBoost yielded lower Brier scores and negative log-likelihood. Calibration lowered ECE from 34.18% to 11.64% on S3R and from 32.05% to 15.52% on S3L. Adaptive re-enrollment needed 88 and 84 updates on S3R and S3L, respectively—fewer than the fixed policies—although it did not achieve the lowest residual error. Overall, the results support managing ECG biometric systems with explicit attention to reliability, balancing risk, coverage, and update frequency.