Evaluation of Unsupervised Anomaly Detection Techniques in Labelling Epileptic Seizures on Human EEG

Abstract

Automated labelling of epileptic seizures on electroencephalograms is an essential interdisciplinary task of diagnostics. Traditional machine learning approaches operate in a supervised fashion requiring complex pre-processing procedures that are usually labour intensive and time-consuming. The biggest issue with the analysis of electroencephalograms is the artefacts caused by head movements, eye blinks, and other non-physiological reasons. Similarly to epileptic seizures, artefacts produce rare high-amplitude spikes on electroencephalograms, complicating their separability. We suggest that artefacts and seizures are rare events; therefore, separating them from the rest data seriously reduces information for further processing. Based on the occasional nature of these events and their distinctive pattern, we propose using anomaly detection algorithms for their detection. These algorithms are unsupervised and require minimal pre-processing. In this work, we test the possibility of an anomaly (or outlier) detection algorithm to detect seizures. We compared the state-of-the-art outlier detection algorithms and showed how their performance varied depending on input data. Our results evidence that outlier detection methods can detect all seizures reaching 100% recall, while their precision barely exceeds 30%. However, the small number of seizures means that the algorithm outputs a set of few events that could be quickly classified by an expert. Thus, we believe that outlier detection algorithms could be used for the rapid analysis of electroencephalograms to save the time and effort of experts.