Abstracts

Rationale: High frequency oscillations (HFOs) are believed to be associated with epileptic properties of the neuronal tissue in the brain. The detection of HFOs is in most cases done manually by trained EEG analysts. This is labour intensive and results are vulnerable to subjective decisions. This study aims primarily to design an automated algorithm that can match the performance of a human observer, and the same time provide reproducible output. As secondary objective, by quantifying certain non-linear features of high frequency EEG components we address the issue of more reliable selection of regions that may be closest to the seizure onset zone (SOZ). This way, automatic prediction of the SOZ during an interictal period could both drastically shorten the registration time of intracranial EEGs for presurgical evaluation and at the same time increase the accuracy of the localisation.Methods: Intracranial electroencephalograms (EEGs), obtained with depth electrodes, of 6 patients with intractable temporal lobe epilepsy were analysed retrospectively. All patients were candidates for epilepsy surgery, and the intracranial EEGs were part of the presurgical workup. Five minutes of slow wave sleep from an interictal period were selected from each EEG. EEGs were pre-processed, obtaining the first component from an empirical mode decomposition of the signals from each channel, thereby extracting the fast frequencies of the signals. 80-sample, 50% overlapping EEG windows were used to fit autoregressive models (ARM) of order 3, thereby obtaining the residual variation of the signal for each window. This residual is interpreted as evidence for non-linear dynamic processes. The Percentage of Windows with Excess Residual (PWER) for each channel was obtained, using a threshold of 5 for excess residual. Subsequently, the maximum PWER (PWERmax) per depth electrode bundle was obtained. Bundles with PWERmax>0.12 were considered to be in the SOZ. Results per bundle were compared to the current gold standard for finding the SOZ, i.e. visual observation of ictal recordings (intracranial EEG with video).Results: Results per patient are shown in Table 1. Electrode bundles found to be in the SOZ by gold standard, had, with one exception, high PWERmax values (>0.12). The novel method has a sensitivity of 95% and a specificity of 85% for bundles located in the SOZ. False positives could have been caused by the epileptogenic nature of the area in which the electrode bundles were situated (contradictory to gold standard SOZ findings), by muscle activity, or by artefacts, and remain to be investigated.Conclusions: These data indicate that the presence of high residual signal variation (after ARM fit), interpreted as non-linearity, can be used to automatically predict the SOZ from high frequency components in short interictal periods of intracranial EEGs.