Abstracts

Rationale: Detecting epileptic spikes in the electroencephalogram (EEG) is a crucial step during epilepsy diagnosis. Conventionally, medical experts visually analyze the EEG recordings with the aim of identifying and interpreting these patterns. This procedure is time-consuming and laborious especially for long term recordings of many hours or days. As it was shown in recent studies, the manual EEG inspection furthermore suffers from a remarkable variability between different reviewers and from a restricted reproducibility of the results. For this reason automatic spike detection systems are of great interest in order to allow an objective and accurate investigation with significant time-savings.Methods: We present a novel method that automatically detects epileptic spikes in the EEG. With our approach we firstly detect spikes with rather low specificity to include also all questionable spike candidates in the initial result. For this we use a pre-defined spike morphology based on intervals e.g. for the rising edge, for the peak, and for the falling edge. Each detected spike has to simultaneously meet different criteria regarding the EEG waveform in these intervals. In the next step we perform a source localization to estimate the 3D coordinates of the origin of the detected spikes in the brain. A clustering approach is then applied to build up groups of similar spikes using these 3D coordinates together with the morphological features derived in the initial detection. Finally, spike probability measures are derived using a statistical analysis of the results and assigned to the clusters.Results: The proposed system was tested with the EEG of three patients with a recording duration of 6.5 hours in total. The data included 282 spikes, which were annotated by experienced EEG experts. We emphasis on a reliable detection of spikes with a minimal number of false positives instead of attempting to detect all spikes contained in the data. We therefore use the precision for measuring the performance of the system. It is defined as the number of true positives in relation to all detections. After the initial detection we achieved an average precision of 85% with a sensitivity of 48%. With the subsequent clustering approach, the precision for the cluster with the highest spike probability increased up to 98% with a sensitivity of 42%.Conclusions: The proposed system automatically detects epileptic spikes in the EEG in the context of epilepsy diagnosis. The two-step approach with subsequent clustering provides a detailed overview of the contained spikes as well as their localization. Furthermore, the clustering approach leads to a very high precision of the detection results and hence fits very well to the aim of detecting clear and typical spikes only. This is achieved by the fact that artifacts or other EEG patterns like alpha waves are usually grouped together in a distinct cluster and can therefore be reliably separated from the epileptic spikes.