Abstracts

Rationale: Variations of the temporal regularity of intracranial EEG (iEEG) in epileptic patients have been found meaningful with respect to seizure occurrence. While the Hölder exponent and the related spectrum quantify local and global regularity of signals, they have not been applied to characterize preictal states. The theory of complex systems also provides new frameworks for a better characterization of non linear dynamics. In particular, Tsallis and Renyí statistics provide entropies and related thermodynamic measures that can be applied to iEEG and may allow characterizing a preictal state in a similar way that we can describe a phase transition in a disordered complex system. Methods: iEEG recordings consist of 14 segments of 40min before seizure onset for 5 temporal lobe epilepsy patients and 7 background segments lasting 45 to 115min for each patient. Four to five electrodes were selected for each patient in the focus and in contralateral regions. To identify and characterize preictal states, the following measures were computed on 1, 5 and 10s sliding windows: Hölder exponent, singularity spectrum, Multiresolution Energy, Multiresolution Tsallis and Renyí entropies and Multiresolution Tsallis and Renyi temperatures. Their predictive values were assessed through ROC analysis. The best measures in term of separability served as features for seizure prediction. Clustering was used to describe the distribution of the points in this feature space. The results lead to the development of a prediction algorithm based on the rate of outliers from the interictal distributions. The algorithm is based on a probabilistic description (mixture of Gaussians) and optimization techniques. This approach allows finding abnormal dynamics in the signal that identify preictal states. Results: While there are discrepancies in the performance of the measures between patients and seizures, these measures show good ability to discriminate interictal from preictal activity. In particular, the multiresolution temperature measures and Hölder exponent exhibit the best predictive power. For the multiresolution measures, predictive values are larger in the frequency band [100,450]Hz. The best channels are mostly located in the seizure onset zone but can also be in contralateral regions. However, no general trends (preictal increase or preictal decrease) could be associated with a single measure accross patients. Clustering shows that there is no specific region in the features space made of solely preictal points. Therefore, classical pattern recognition through supervised learning is not possible. However, our algorithm based on the rate of outliers was able to detect 10/14 preictal states at various time intervals for a threshold chosen for zero false alarms. Conclusions: Individual measures did not change consistently in preictal periods across patients. The overall distribution of the variables, however, appears to change before seizures in such a way that outlying values (statistically improbable values) become more frequent. This may open new avenues in seizure prediction.