Extending RandomAD for Multichannel EEG Seizure Detection

Abstract

Time Series Anomaly Detection aims to uncover unusual patterns or windows within time series data with applications including medical diagnostics and clinical monitoring. In electroencephalogram (EEG) recordings, anomalies may correspond to seizure activity or other clinically significant changes that are difficult to detect through manual inspection alone. Existing anomaly detection methods primarily target univariate data, although real-world data is often multivariate. This project investigates how RandomAD, an efficient semi-supervised anomaly detection method, can be extended to multichannel data, with application on EEG seizure detection. RandomAD identifies anomalous windows via random convolutional kernels, kernel selection, and k-Nearest-Neighbor anomaly scoring. Its computational efficiency makes it particularly well-suited for large, high-dimensional physiological recordings.Specifically, we implement RandomAD in Python, preprocess multichannel CHB-MIT EEG  recordings, and assess four multichannel detection strategies: individual channel detection, sequential channel concatenation, max score fusion, and mean score fusion. We compare their performance across varying numbers of random convolutional kernels and kernel selection rates for seizure detection. In addition, we generate visualizations of anomaly scores and detected windows to compare each fusion method and identify trends across patient recordings. Preliminary results suggest detection performance varies strongly by EEG channel, with some channels detecting seizure aligned anomaly windows consistently while others produce off-target detections. Notably, score fusion improves detection in several parameters while concatenation is more computationally expensive. We anticipate that these findings will have significant impact in improving multivariate time series anomaly detection methods for EEG-based seizure identification and preventative clinical applications.