F08. Automatic spike detection in intracerebral depth electrode recordings

Abstract

Introduction Intracranial recordings, like intracerebral depth electrode recordings are considered to be the best choice for preoperative invasive evaluation when standard electro-clinical examinations are not conclusive. These recordings reflect a vast amount of interictal epileptic discharges, the so called spikes, which are in general abundant compared to seizure activity. Manually reviewing these recordings to find the spatial origin of seizures and spikes is still state of the art, which is a very exhausting task due to the large number of signals recorded. In order to enable a time-efficient evaluation of such recordings, we developed a spike detection algorithm which performs automatic detection, spatial clustering and smart visualization of spike clusters. Methods The automatic spike detection algorithm for depth electrode recordings is based on our existing spike detection algorithm for surface EEG. The spike detection algorithm assesses morphological and topographical potential field features. Both feature groups had to be adapted for depth electrode recordings. In these recordings spikes are shorter in duration, often have higher amplitude and the properties of the potential field compared to surface EEG are very different. Due to the morphological similarity of individual waves of the posterior alpha rhythm and spikes, the detection within such rhythmic periods is suppressed. In order to speed up the review, we developed a spatial clustering and a smart visualization of spike clusters. In the so-called overlay plot the signals of all individual spikes of each cluster are displayed one above the other in one figure, allowing for quickly distinguishing between true spikes and artifacts. In order to assess the sensitivity and precision of the developed algorithms, 1400 spikes in 3 datasets from different patients recorded at the Academic Center of Epileptology, Kempenhaeghe/MUMC were annotated by an experienced biomedical assistant. Results The automatic spike detection algorithm detected 54% of all spikes on average. The overall precision was 39%. Sensitivity values reach from 18% to 81%. However, also for the patient with 18% sensitivity the algorithm was able to detect several spikes on all those electrodes where also the human reviewer detected them. For one patient a very low precision of 6% was measured. This small precision value came from repetitive artifacts which are very similar to spikes and are even difficult to distinguish for human experts. Removing the most evident artifact clusters by using the overlay plot, the precision increases from 6% to 30%. Doing the same for all patients an average precision of 68% is achieved. Conclusion Our automatic method was able to detect spikes with a high sensitivity on average. Moreover, our algorithm always detected spikes on all the electrodes where the expert also identified them allowing for a profound diagnosis. The smart combination of detection, clustering and visualization potentially enables a more time-efficient review.