Abstract—Diagnosis of epilepsy is based on the analysis of electroencephalogram (EEG) recordings. Essential epileptiform transients in the EEG are spikes, which are commonly marked manually by biomedical technical assistants which is very time-consuming and error-prone. Automatic spike detectors already exist but still have to be improved to better meet the needs of clinical experts. In this paper we discuss different automatic spike detection methods in order to improve the detection performance and to establish a user adjustable sensitivity parameter. The performances of a rule-based system, artificial neural networks (ANN) and random forests are investigated. For this retrospective study, data from an epilepsy-monitoring unit, including 12 patients comprising 130 hours recording time, were collected. The recordings were annotated by medical experts leading to a total of 5582 spikes. An artificial neural network exceeds the alternative methods in classifying the data set and achieves an average detection sensitivity of 44.1% and positive predictive value of 56.2% at a false detection rate of 19.8 per hour. Furthermore, the ANN also performs well in different sensitivity settings, enabling a user adjustable sensitivity parameter which helps the clinical experts to adjust the classifier to handle different application scenarios.

Index Terms—Epilepsy, spike detection, EEG, automatic, classification, rule-based, machine learning, minority class oversampling, artificial neural networks, random forests

Cite: Wolfgang Ganglberger, Gerhard Gritsch, Manfred M. Hartmann, Franz Fürbass, Hannes Perko, Ana Skupch, and Tilmann Kluge, "A Comparison of Rule-Based and Machine Learning Methods for Classification of Spikes in EEG," Journal of Communications, vol. 12, no. 10, pp.  589-595, 2017. Doi: 10.12720/jcm.12.10.589-595.