Long-Term, Patient-Specific Seizure Prediction using Absolute Mean Instantaneous Frequency Difference (AMIFD) and Seizure Clustering.

Abstract

Long-term seizure prediction in epileptic individuals is challenging, mainly due to signal non-stationarity and seizure type variability. This research presents a patient-specific prediction pipeline for intracranial electroencephalography (iEEG) based on the novel Absolute Mean Instantaneous Frequency Difference (AMIFD) biomarker. The approach addresses seizure variability using the minimum uncertainty and sample elimination (MUSE) feature-ranking technique to automatically cluster seizure types based on their top-ranked AMIFD features, distinguishing the interictal baseline from the preictal phase (defined as a 60-min window ending 5 min before onset). For each identified seizure cluster, a specialized random forest classifier is trained independently, creating a multi-model ensemble tailored to specific seizure morphologies. The final prediction alarm is triggered using a k-of-N aggregation logic to enhance reliability and minimize false alarms. The AMIFD biomarker shows a significantly larger effect size than several standard electrophysiological features (adjusted p < 0.04). When tested on 10 epileptic subjects, the patient-specific framework achieved a mean sensitivity of 92.08% and a false-positive rate of 1.32/day, outperforming the minimum Redundancy Maximum Relevance (mRMR) and MUSE baselines. This patient-specific, multi-model strategy establishes a viable pathway toward more accurate and clinically reliable personalized seizure prediction systems.