Localization of Epileptiform Activity on EEG and DTW-Based Event Clustering

Abstract

This paper describes a software pipeline for the analysis of epileptiform activity in multichannel electroencephalographic recordings. The pipeline combines neural-network segmentation, event-level localization of the leading EEG lead, visual grouping of detected events, and interactive filtering of falsepositive detections. A one-dimensional U-Net-like model first produces a pointwise probability mask, which is then converted into compact event intervals. For each event, the leading lead is estimated by an ablation procedure that measures how much the model confidence decreases when a channel is suppressed inside the event interval. Event morphology is compared with dynamic time warping (DTW), which makes it possible to cluster discharges with similar shapes but slightly different duration and phase. In a test recording, segmentation with a probability threshold of 0.01 produced 246 candidate events. After expert selection of informative DTW clusters, the number of false positives decreased from 108 to 40, precision increased from 0.561 to 0.766, and event-level F1 increased from 0.612 to 0.697. The proposed post-processing layer makes automatic EEG segmentation more interpretable by attaching time boundaries, model confidence, leading lead, cluster membership, and quality changes after filtering to every detected event.