Hewlett Packard Enterprise Data Science Institute

Accurate and cost-effective seizure severity trackingis an important step towards limiting the negative effects ofseizures in epileptic patients. Electroencephalography (EEG)is employed as a means to track seizures due to its hightemporal resolution. In this research, seizure state detection wasperformed using a mixed-filter approach to reduce the numberof channels. We first found two optimized EEG features (onebinary, one continuous) using wrapper feature selection. Thisfeature selection process reduces the number of required EEGchannels to two, making the process more practical and cost-effective. These continuous and binary observations were used ina state-space framework which allows us to model the continuoushidden seizure severity state. Expectation maximization wasemployed offline on the training and validation data-sets toestimate unknown parameters. The estimated model parameterswere used for real-time seizure state tracking. A classifier wasthen used to binarize the continuous seizure state. Our resultson the experimental data (CHB-MIT EEG database) validate theaccuracy of our proposed method and illustrate that the averageaccuracy, sensitivity, and false positive rate are85.8%,91.5%,and14.3%respectively. This type of seizure state modeling couldbe used in further implementation of adaptive closed-loop vagusnerve stimulation applications.