Variational Bayesian learning for Gaussian mixture HMM in seizure prediction based on long term EEG of epileptic rats

Abstract

Epilepsy is a common neurological disorder characterized by abnormal excessive or synchronous neural activity in brain. In this study, we develop an unsupervised learning for seizure prediction. Extracting wavelet features of brain electroencephalogram (EEG), we propose a Hidden Markov Model (HMM) with a mixture of Gaussian observation model as an unsupervised learning setting for seizure prediction, where the seizure predictions are derived from the posterior distributions over the hidden states in the HMM. By using the Variational Bayesian (VB) method instead of the Maximum Likelihood estimation, which is the method commonly used for training HMMs, we overcome data overfitting and make it possible to compare models with different model orders by means of the variational free energy. VB learning also improves results in terms of convergence speed and achieved performance. The proposed method was evaluated using 20h of labeled EEG recordings from 7 epileptic rats with total number of 350 seizures. Our method obtained a high sensitivity of 90.7% and a specificity of 88.9% with early detection of 1.3s, which makes it more reliable than ML estimation with a sensitivity of 82.1% and a specificity of 86.2% and late detection of 4s. Keywords-Electroencephalogram (EEG); Epilepcy; seizure prediction; hidden Markov model (HMM); variational Bayesian . I. , {π θ K J J J J

Dynamic causal modelling of seizure activity in a rat model

“…These include stochastic models to help predict occurrence of seizures (Esmaeili et al 2014;Wong et al 2007); phenomenological models of EEG dynamics that can be used to identify contributing factors leading to seizure onset (Benjamin et al 2012;O'Sullivan-Greene et al 2009;Geier et al 2015); and biophysically informed dynamic models attempting to identify abnormal synaptic parameters leading to epileptic activity in simplified models of population dynamics (neural mass models) (Goodfellow et al 2012;Breakspear et al 2006;Aram et al 2015). This work has gained particular traction since the development of model inversion techniques that allow model parameters to be estimated directly from empirical data, such as the variations of the Kalman filter (Ullah & Schiff 2010;Haykin & Arasaratnam 2009), dynamic causal modelling (Papadopoulou et al 2015;Cooray et al 2015;Moran et al 2011), and related machine learning techniques such as Markov chain Monte Carlo (MCMC) (Luna et al 2014;Wulsin et al 2014).…”