Semi-Supervised anomaly detection for the prediction and detection of pediatric focal epileptic seizures on fused EEG and ECG data

Abstract

Listen

Translate article

We introduce a semi-supervised multi-modal anomaly detection framework for efficient seizure prediction and detection in pediatric patients suffering from focal epilepsy. Our approach combines temporal, spectral and nonlinear features derived from two modalities: Electroencephalogram (EEG) recordings recorded in electrodes placed adjacent to the seizure loci and Electrocardiogram (ECG) recordings from which we derive Heart Rate Variability (HRV) parameters. We employ anomaly detection models, such as the Minimum Covariance Determinant (MCD) estimator, Isolation Forest (IsF), One-Class Support Vector Machine (OCSVM), and Local Outlier Factor (LOF), which were trained in a semi-supervised way to identify EEG / HRV anomalies during or prior to seizure onset. Our method demonstrates robust performance in detecting segments of ictal activity, achieving ROC-AUC scores of up to 95% using patient-specific thresholds and 91% using patient-in specific thresholds. Additionally, our framework was able, in most cases, to anticipate seizures 16–23 s prior to seizure onset, with a minimal duration of false warnings. These findings support the potential for developing a lightweight, wearable EEG / ECG monitoring device optimized for pediatric patients.