Analyzing Intracranial EEG in Pharmacoresistant Epilepsy Patients Using Hidden Markov Models and Time Series Forecasting Methods

Abstract

Various brain activities can be captured by electroencephalographic signals, which can then be used to detect epilepsy considering that the epileptic seizures are caused by a disturbance in the electrophysiological activity of the brain. Accordingly, epileptic seizure prediction usually requires a careful analysis of electroencephalography (EEG) records. In this study, we examined a large intracranial EEG dataset obtained from five pharmacoresistant epilepsy patients. Specifically, we first applied hidden Markov models to parse the amplitude under a probabilistic description considering the observed data as the outcome of either one of the three hidden states, namely, normal, subclinical (seizure) and clinical (seizure), which is in line with the setup proposed in previous studies. Our results indicate that there are indeed a maximum of 1.7% subclinical and about 1% clinical events, which comply with the observations from other studies. Next, we assumed EEG signals to form complex time series, and considered time-series prediction methods to forecast the future events. Such predictions are of interest to predetermine the possibility of a seizure onset and taking a preventive strategy. This task was performed within a deterministic framework by applying deep learning methods. An ensemble model was created using one-dimensional convolution net in conjunction with long short-term memory units and deep neural networks. We observe that the proposed time series prediction method is highly accurate as indicated by the low mean absolute error values and the high conformity of the predictions to the ground truth values.