Optimized detector for closed-loop devices for neurostimulation

Abstract

Brain machine interfacing (BMI) needs continuous analyses of ongoing brain activity. For a successful interaction, related brain activities and events should be reliably detected; using various approaches including machine learning techniques. To this end, a variety of characteristic signal features as well as different types of classifiers can be used. One possible application of such an interaction is for epilepsy patients. A novel approach for the group of patients with difficult to treat epilepsy is the application of electrical stimulation in the early stages of the seizure generation in a closed-loop manner which can be realized in an implant. Herein, we show results of studies on the detection of epileptic seizure patterns in human intracranial long-term recordings and their dependence on selection parameters which have to be chosen for the realization in an implant. Random forest classifier is shown to allow an energy-efficient implementation of algorithm which uses a set of time and frequency domain features for seizure detection. In this study, we searched for further possibilities to optimize the performance of our algorithm and made it more robust to signal variations for online applications. In this regard, we studied the effects of detection time window, raw data normalization, feature scaling and electrode montages on performance of random forest classifier. Results of this optimization process indicate a decrease of detection delay, which is crucial to successful seizure suppression, and increased sensitivity; while preserving the false positive detections low compared to presently available closed-loop intervention in epilepsy.