Abstract
Objective:
Differentiating functional/dissociative seizures (FDS) from epileptic seizures (ES) remains clinically challenging, with limited electrocardiogram (ECG) biomarker reliability. This study evaluated whether explainable machine learning applied to ECG features could identify autonomic markers for FDS–ES discrimination.
Methods:
ECG recordings from 125 patients with FDS (n = 83) or ES (n = 42) from two epilepsy centres were analysed. A number of heart rate, HRV, and morphological ECG features was extracted from interictal and preictal segments. Relative-change features were calculated by normalising preictal values to interictal baseline. Classification used Leave-One-Subject-Out Cross-Validation with mutual information filtering, SHapley Additive exPlanations (SHAP)-guided feature selection, class balancing, and hyperparameter tuning.
Results:
Entropy-based HRV measures were the most consistent discriminative features. In FDS, Sample Entropy, Fuzzy Entropy, and Dispersion Entropy decreased significantly from interictal to preictal states, whereas no significant entropy modulation was observed in ES. Dispersion Entropy showed the strongest contribution across statistical testing, SHAP interpretation, and feature-selection stability. Classification was limited in the interictal condition, and the best performance was obtained using relative-change features: XGBoost achieved 73.5% sensitivity (95% confidence interval [CI]: 62.7–82.6%) and 61.9% specificity (95% CI: 45.6–76.4%).
Conclusions:
FDS was associated with preictal reduction in HRV entropy, indicating more regular, less complex cardiac dynamics. Within-subject changes appeared to provide more discriminative than static ECG features. Although current performance does not support standalone diagnostic use, entropy-based HRV measures offer interpretable peri-ictal autonomic markers, suggesting visceral changes contribute to FDS emergence.