Why the study?
ECG-based seizure prediction has predominantly relied on heart rate variability, whereas intra-beat ECG morphological features have remained underused, leaving uncertain whether they offer complementary predictive information.
Population
Patients in the SeizeIT2 and Siena external validation datasets
Comparison
Integrated HRV and ECG morphology inputs vs HRV-only vs ECG morphology-only inputs
Design
Deep Support Vector Data Description one-class anomaly detection modeling with internal and external validation
Answers are grounded in this paper and cite their sources.
May support multimodal ECG seizure prediction algorithms; leaves open prospective validation before clinical adoption.
Electrocardiogram (ECG)-based seizure prediction has predominantly relied on heart rate variability (HRV), while intra-beat information represented by ECG morphology (ECG Morp) has remained underused. This study evaluated whether ECG Morp features provide complementary information to HRV within a one-class anomaly detection framework. We developed three Deep Support Vector Data Description models using HRV-only, ECG Morp-only, and integrated HRV and ECG Morp inputs. Models were trained with interictal sequences from SeizeIT2 and evaluated on a patient-level SeizeIT2 internal validation set and the Siena external validation dataset. Seizure-wise sensitivity and time in warning (TiW) were assessed at target false alarm rates (FARs) of 0.1, 0.2, 0.5, and 1.0 false alarms per hour. Circular-shift surrogate analysis and Integrated Gradients (IG) were used to assess above-chance temporal association and feature contributions. At 0.5 false alarms per hour, the integrated model achieved sensitivity/TiW values of 0.424/6.7% in SeizeIT2 versus 0.249/9.7% for the HRV-only model, and 0.500/7.3% in Siena versus 0.375/10.5%. At 1.0 false alarms per hour in Siena, the HRV-only model achieved higher sensitivity than the integrated model (0.900 vs. 0.850), although the integrated model retained lower TiW (20.8% vs. 28.0%). Surrogate analysis supported temporal associations beyond those expected under the circular-shift model in both datasets. HRV and ECG Morp showed broadly comparable contributions (0.442–0.500 and 0.500–0.558, respectively), indicating that the integrated model did not rely exclusively on either feature group. Overall, ECG Morp provided complementary information to HRV and generally improved the balance between seizure-wise sensitivity and warning burden.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: