gms | German Medical Science

Objective: The unpredictability of epileptic seizures is responsible for injuries, stress, depression and may limit the social functioning of the affected individuals. Up to one third of patients with epilepsy remain refractory to any treatment during their lifetime. A reliable non-invasive seizure prediction method would decrease those risks and alleviate the psychological impact.

We propose the exploitation of ECG signal for seizure prediction and developed a method based on the Heart Rate Variability (HRV) changes occurring during the preictal period, reflecting an Autonomic Nervous System (ANS) modulation prior seizure onset.

Methods: Single lead ECG signal was obtained from 28 patients with epilepsy, totalling 75 seizures (median age: 42 years). Seizure onset was defined by an experienced epileptologist; seizures had to be at least two hours apart. One hour of ECG signal before seizure onset was analysed in overlapping 1-min-windows with 10 sec steps. A matrix of 355 windows x 63 HRV features (time, frequency and non-linear) was constructed for each patient. Windows containing artefacts were removed. The preictal period was defined as 10 min and the interictal period as 60 – 35 min before seizure onset.

Different models for seizure prediction were devised using either Support Vector Machine (SVM) or Random Forest Classifier (RF). HRV features were weighted and selected by an Extra-Trees Classifier. The prediction efficacy was evaluated by Leave-One-Out Cross-Validation (LOOCV).

Results: Performance for inter-patient seizure prediction using either seizure LOOCV or patient LOOCV was rather poor (ROC-AUC: 0.589 and 0.582, respectively). Stratification using sleep-wake cycles or localisation of seizure onset did not improve the performance.

For patients with ≥3 seizures (n=12) a patient-centred approach was additionally tested. Here, the models exhibited better predictive performance in the patient-specific LOOCV (mean ROC-AUC: 0.747).

Conclusion: The results imply a heterogeneous ANS behaviour between patients during the preictal period. This heterogenity can be also observed among seizures of the same individual, with seizures showing diverge feature importance as indicated by Mann-Whitney U test.

This phenomena may inherently limit the predictive potential of our proposed method. Further research is needed in order to identify additional cross-seizure and cross-patient seizure precursors, which could improve seizure prediction.