Abstracts

Rationale: Automatized seizure detection has been a research topic of renewed interest in light of development of innovative therapies including responsive neurostimulation. However, reliable early seizure onset detection remains a challenge. The complexity of seizure dynamics, the variability of seizure-onset patterns and the localization of the seizure onset foci, can often limit the efficacy of seizure detection algorithms. Modern seizure detection algorithms make use of machine learning algorithms to classify electroencephalographic features into ictal or non-ictal events. The performance of these algorithms relies highly on the training data. This is particularly difficult in the epilepsy monitoring unit when no prior ictal data is available. In this study, we investigate a seizure detection paradigm in which training of the machine learning algorithm is performed online, resulting in a dynamical adapting algorithm for seizure detection that is able to classify ictal events and is practical to implement in a clinical setting.Methods: Intracranial recordings of 5 patients with focal epilepsy undergoing presurgical evaluation were used. Five-hour epochs of EEG ending in a seizure were used from each patient. Our algorithm performs a sliding window analysis of the EEG to extract a set of 11 features per channel that have already been shown to be useful in seizure detection; these include spectral parameters, complexity and entropy measures (Jouny 2011). Features were calculated over 4 seconds window with a one second sliding window. We implemented a novelty detector using a one-class Support Vector Machine (SVM), which trains on a 20 minute window of features and shifts every minute. The output of the SVM is then processed using a Kalman filter.Results: The efficiency of the detection is measured here as the detection delay for the ictal event at the end of the testing window. As the nu parameter of the SVM is an upper bound on the fraction of outliers, it can be used to adjust the trade-off between overfitting and generalization, we analyze over a range of nu values from 1e-4 to 0.5 to assess its impact on the detection delay. For all patients, all seizures were detected with detection delays ranging from 0 to 4 seconds depending on the fraction of outliers, with lower delay for higher nu values. The false positive rate range from 0.11/h (under fitting) to 0.5/h for the largest value of nu (overfitting).Conclusions: The optimization of the choice of features and the dynamic adaptation of the detector allowed us to detect previously unseen seizure events without prior training of the ictal signature. Although all detections were performed offline in this study, the processing time of our algorithm allows for implementation of real-time seizure-onset detection. The use of a dynamically adaptative SVM is a promising paradigm for the detection of ictal events whose dynamic characteristics are unknown at the time of patient’s admission to the epilepsy monitoring unit. Funding by NIH R01 – NS75020