Online classification and prediction of spontaneous seizure-like events in in-vitro epilepsy models

A computer-based intelligent system for real-time classification and prediction of seizure-like events (SLEs) using artificial neural networks (ANNs) on either a computer or animal model of biological neural networks (BNNs) is developed and implemented. The design of this prediction pathway is essential for the development of an intelligent seizure feedback controller since it determines whether the feedback stimulation strategy will be activated.; Three types of classifiers with (a) Mixture of Gaussians (MoGs), (b) wavelet nonlinearities and (c) hidden Markov model (HMM) are compared offline. The Gaussian Artificial Neural Network (GANN) is able to capture the manifold information by fitting the state points with Gaussian radial basis functions (RBFs). The Wavelet Artificial Neural Network (WANN) is able to capture the time-varying frequencies in the recorded potentials of seizure episodes. HMM is able to select optimal topology nonparametrically. GANN classifier is successful only in the subject-dependent situations. The WANN design does not require detailed understanding of the exact system dynamics. We show that the WANN and HMM are more robust because they capture the system invariance across subjects.; Sensitivity study, pruning analysis and automated relevance determination (ARD) are performed, demonstrating that activities in the high frequency range, above 100Hz, are critical for classification of SLE events. This study is essential for the selection of input frequency range for development of online ANN-based seizure classifier and predictor. The WANN is tested online for the in-vitro hippocampal models, capable of generating spontaneous recurrent seizure-like events (SLEs) [1].; Then, the SLE prevention pathway is explored using stochastic analysis. It is performed on a computer model to understand possible mechanisms for the effectiveness of reported SLE abolishment strategies using periodic stimulations of different frequencies. We study stochastic resonances (SR), coherence resonances (CR) on various coupling pathways between cells [2] to relate the efficacy of communication between neurons to the frequencies of the encoded information. The preliminary result suggests that gap junction and field couplings have the strongest effects on SR and CR. Rhythmicity can be broken when stimuli are given at specific frequency band.