Abstracts

Rationale: Automated seizure detection has wide applicability, from diagnostic epilepsy monitoring to responsive neurostimulation. Traditional supervised learning approaches exhibit sensitivities that range from 75-85%, with false positive rate (FPR) of 2-3 per hour, compared to the human gold standard of over 92% sensitivity with 0.12 FPR. Sensitivity and FPR improves if patient-specific approaches are used, but an example of the patient s seizures must be first identified. Unsupervised learning approaches may provide comparable or superior detection without this limitation.Methods: We utilized scalp EEG data from the channel of seizure onset from 10 patients from epilepsy monitoring unit evaluations. Using a sliding window paradigm, we extracted absolute power from five specific spectral bands from successive 2-second data epochs. Various supervised learning techniques (support vector machine (SVM), na ve bayes (NB) classifier, and k-nearest neighbors (kNN)) were used with this feature data to generate binary classifier outputs. For patient-independent approaches, 5-fold cross validation was used to generate performance characteristics. For patient specific approaches, a classifier was trained on each patient s seizure for detection of the patient s remaining seizures, with the results averaged. We used an adaptive unsupervised outlier detection approach using Gaussian Mixture Models (GMMs) to generate a sequence of binary outlier outputs (Figure 1). A buffer of the past 15 minutes of feature data was used to generate a new GMM every 7.5 minutes. Both supervised and unsupervised binary outputs were smoothed with a 10-second window, with an alarm raised if 60% of feature data were classified as seizure or outlier. We calculated sensitivity and FPR from this smoothed classifier output. Different outlier probability thresholds of the GMM were used to generate an ROC curve.Results: Preliminary results reveal that GMMs are capable of providing superior performance at certain sensitivity thresholds, with 87.5% sensitivity and FPR of 1.84/hr at the 1% outlier threshold level. The majority of the false positive detections featured arousal from sleep, followed by chewing and electrode artifact.Conclusions: Adaptive, unsupervised methods are promising for scalp EEG seizure detection and compare favorably to published supervised learning approaches. This approach has the advantages of patient specificity and adaptation to changing detection conditions, without having to identify a seizure a priori. Disadvantages include alarm from other sudden electrographic phenomenon such as state changes. Incorporation of contralateral channels of data may allow superior detection for focal changes associated with seizures.