Abstracts

Rationale: Electrocardiogram (ECG) changes are commonly observed with epileptic seizures, making it a promising candidate for monitoring seizure risk. Previous personalized seizure detection approaches have been based on heart rate (HR) or heart rate variability (HRV) analysis to detect seizures in responders, defined as patients with seizures accompanied by HR changes of 50 bpm in 100 RR-intervals. To identify ECG seizure biomarkers during distinct peri-ictal periods, we propose further segmenting peri-ictal phases using multivariate change-point detection.

Methods: We retrospectively analyzed routine 1-lead ECGs of 20 pediatric patients in the epilepsy monitoring unit at Boston Children’s Hospital (2/2015-2/2021). We identified one GTC or FBTC seizure per patient with at least 2 hours of seizure-free recording before and after seizure onset. We extracted 60-min pre- and post-ictal periods and segmented ECG signals into 60-s windows. From each segment, we computed HR, linear and nonlinear HRV metrics, and QRST morphology features (Fig. 1). We applied multivariate kernel change-point detection with a minimum segment size of 3. The optimal number of change-points was selected by minimizing Bayes Information Criterion. Permutation tests, based on the mean difference between segments, were conducted with a ±1-min buffer from seizure onset to identify seizure-associated change-points. Resulting p-values were adjusted using Bonferroni correction to control for multiple comparisons. We evaluated temporal clustering of change-points using a Kolmogorov-Smirnov test, comparing their distribution across the 2-hour window against a uniform null hypothesis.

Results: Our findings indicate that various cardiac markers reveal different trends in peri-ictal phases (Fig. 2). We identified a mean of 4.6 change-points per patient (SD: 1.35, IQR: 2.0). We found at least one significant change-point before seizure onset in 80% of patients (mean: 1.65, SD: 1.14, IQR: 1.0) and at least one significant change-point after seizure onset in 95% of patients (mean: 2.25, SD: 1.02, IQR: 1.0). We correctly detected seizure onset in 90% of patients with the detection being significant in 70% (all values p < 0.05). The Kolmogorov-Smirnov test revealed a significant deviation from a uniform distribution (KS statistic = 0.17, p = 0.002), indicating temporal clustering of change points around seizures.