Influence of circadian rhythms on epileptic seizure predictors based on machine learning methods

Abstract

Circadian rhythms are physiological processes that regularly oscillate with a period of approximately 24 hours. The occurrence of epileptic seizures is not entirely random in terms of their distribution over the day. The reason for this pattern is not entirely clear but evidences suggest that there is an interaction between epileptic seizure occurrence and the circadian cycle. In this work, we have analyzed the long-term behavior of a seizure prediction algorithm based on Support-Vector Machines (SVM) and on a high dimensional feature vectors obtained from multichannel scalp electroencephalography (EEG) recorded from 10 patients of the EPILEPSIAE database. Statistical significant differences were identified between day and night periods, while the power spectral densities of the models' output present a strong influence of the 24 hours period. These results suggest that pre-ictal patterns used to train the models are related to processes acting on daily rhythms. Part of the fluctuations of the models' output can be attributed to daily rhythms, which ultimately suggest that specific periods of the day may represent pro-seizures state.