Abstracts

Rationale: Seizure attacks can lead to irreversible consequences due to sudden loss of consciousness, making preparation and prevention from injury crucial. While numerous methods have been explored for predicting seizures using scalp electroencephalography (EEG), the approaches and outcomes have shown considerable variability. We aimed to predict seizure attacks using spectral power of preictal and interictal scalp EEG through machine learning classification.



Methods: The proposed algorithm comprised of preprocessing, feature extraction, and machine learning classification. Preprocessing involved applying a band-pass filter (1-100 Hz) and a notch filter (60 Hz) to remove artifacts. Spectral power within delta (1–4 Hz), theta (4–8 Hz), alpha1 (8–10 Hz), alpha2 (10–13 Hz), and beta (13–30 Hz) frequency bands was extracted from 10-second windows using a sliding-scale analysis with a half-overlap. These windows were classified as either ictal (containing seizures), interictal (occurring at least 1 hour before or after a seizure), or preictal (30 minutes prior to seizure onset) samples. Machine learning classification employing Random Forest (RF) and k-Nearest Neighbor (kNN) algorithm was used to classify preictal and interictal samples. To enhance the prediction accuracy, a double cross-validation procedure was performed five times. The models were trained on 80% of the dataset and validated on the remaining 20% to ensure their adequacy. We prospectively recruited consecutive epilepsy patients who underwent pre-surgical evaluation in an epilepsy monitoring unit at the Taipei Veterans General Hospital. Patients underwent long-term video-scalp EEG monitoring, and all seizure events were recorded during admission. Patients with fewer than three seizure events were excluded from the study.

Results: Fifteen epilepsy patients (9M/6F) were enrolled in this study. Total 59 seizure events were recorded and 59 preictal and interictal EEG samples were derived respectively. The patient-specific algorithm developed for seizure prediction demonstrated a high sensitivity of 99%-100% coupled with a low false alarm rate of 0%-1% through RF classification. Similarly, employing kNN classification yielded a high sensitivity of 94%-100% and a low false alarm rate of 0%-6%. In total, prediction using spectral power features calculated from scalp EEG had total sensitivity of 99.3% with a prediction false alarm rate of 0.5% when using RF classification and total sensitivity of 97.0% with a prediction false alarm rate of 2.5% when using kNN classification.

Conclusions: Linear features of spectral power led to high sensitivity and a low prediction false alarm rate based on RF and kNN classification. This patient-specific binary classification algorithm for scalp EEG seizure prediction stands as a valuable tool for predicting seizure events. Implementing this patient-specific algorithm can serve as a cornerstone for preventing injuries in individuals with epilepsy.

Funding: The study was funded by in part by Taiwan National Health Research Institutes [NHRI-EX111-10905NI, NHRI-EX112-11229NI (to HYY)] and Taipei Veterans General Hospital [V113B-024 (to PTL)].