Classification of epileptiform events in EEG signals using neural classifier based on SOM

Abstract

Analysis of long-term electroencephalogram signals (EEG) is an important tool to clinically confirm the diagnosis of epilepsy. The characteristic electrographic events that represent epilepsy in the analysis of EEG are called epileptiform events (spikes and sharp waves). The process of EEG record analysis is performed by highly trained specialists, which identify the spikes and sharp waves throughout EEG records with minimum duration of 24 hours. Since epileptiform events have typical amplitude around 200µV and duration between 20 and 200ms the analysis of the EEG records is considered very time-consuming and tiring for the experts. Several studies for automatic detection and classification of epileptiform events have been proposed but there is still no system with widespread use and a performance that meets the needs of the specialists. The Self-Organizing Maps of Kohonen (SOM) are an unsupervised neural network algorithm that consists of two layers of neurons that has been successfully used in a wide variety of applications. The objective of this study is to test the feasibility of using Self-Organizing Maps of Kohonen for automatic classification of epileptiform events and non-epileptiform events in EEG signals. Different maps of Kohonen were developed and tested. After simulations, the results were evaluated according to classical performance indexes and the best network achieved 98.7% sensitivity, 91.9% specificity, 90.08% selectivity and 94.8% efficiency. Comparing the results of other SOM studies we obtained sensitivity 9–12% higher and selectivity 12–39% higher than the analyzed studies. Furthermore, a comparison with the results of a previous study that uses the same EEG signal database showed that the overall efficiency was quite similar (only 1% lower).