Heart rate variability-based detection of epileptic seizures: Machine learning analysis and characterization of discriminant metrics

Carole Nouboue; Eva Diab; William Gacquer; Philippe Derambure; Bertille Perin; Simone Chen; Mélodie Mercier-Bryczman; Julien DE Jonckheere; William Szurhaj

doi:10.1016/j.clinph.2025.2110793

Heart rate variability-based detection of epileptic seizures: Machine learning analysis and characterization of discriminant metrics

Clin Neurophysiol. 2025 Jun 16:177:2110793. doi: 10.1016/j.clinph.2025.2110793. Online ahead of print.

Authors

Carole Nouboue¹, Eva Diab², William Gacquer³, Philippe Derambure⁴, Bertille Perin⁵, Simone Chen⁵, Mélodie Mercier-Bryczman⁵, Julien DE Jonckheere⁶, William Szurhaj²

Affiliations

¹ Clinical Neurophysiology Department, Amiens University Medical Center, France; UR 7516, CHIMERE, Jules Verne University of Picardie, Amiens, France. Electronic address: carole.julie@live.fr.
² Clinical Neurophysiology Department, Amiens University Medical Center, France; UR 7516, CHIMERE, Jules Verne University of Picardie, Amiens, France.
³ Digital Services Department, Amiens University Medical Center, France.
⁴ Clinical Neurophysiology Department, Lille University Medical Center, France.
⁵ Clinical Neurophysiology Department, Amiens University Medical Center, France.
⁶ INSERM CIC-IT 1403, Lille, France.

PMID: 40561720
DOI: 10.1016/j.clinph.2025.2110793

Abstract

Objective: We aimed to determine the most suitable cardiac metrics and machine learning algorithms (MLa) for an electrocardiography-based seizure detection device.

Methods: In a multicenter, prospective study of adult inpatients, with limited physical activity, 24-hour video-electroencephalogram recordings including ≥ 1 seizure were analyzed. Heart Rate (HR) and Heart Rate Variability (HRV) metrics were calculated continuously from the corresponding electrocardiogram. HR and HRV time series were segmented into 5-min epochs. MLa were used to classify the epochs as containing seizures or not for the whole dataset; then for convulsive and nonconvulsive seizures only, without focusing on individual results. The sensitivity, specificity and False Alarm Rate (FAR) were calculated.

Results: We included 129 patients and 313 seizures (255 nonconvulsive). The most discriminant metrics were the signal quality, maximum cardiac sympathetic index, maximum heart rate, and minimum high frequency variability index. The sensitivity, specificity and FAR were respectively 94%, 89% and 0.6 for convulsive seizures (extremely randomized trees), and 83%, 82% and 1.13 for nonconvulsive seizures (random forest).

Conclusions: In the largest unselected patient cohort study of this topic to date, seizure detection with ML analyses of cardiac metrics provides good results - even for nonconvulsive seizures.

Significance: The high FAR suggests to combine HR and HRV analysis with other metrics to increase specificity.

Keywords: Electrocardiography; HRV; Heart rate variability; Machine learning; Nonconvulsive seizures; Seizure detection.