Accurate and cost-effective seizure severity tracking is an important step towards limiting the negative effects of seizures in epileptic patients. Electroencephalography (EEG) is employed as a means to track seizures due to its high temporal resolution. In this research, seizure state detection was performed using a mixed-filter approach to reduce the number of channels. We first found two optimized EEG features (one binary, one continuous) using wrapper feature selection. This feature selection process reduces the number of required EEG channels to two, making the process more practical and cost-effective. These continuous and binary observations were used in a state-space framework which allows us to model the continuous hidden seizure severity state. Expectation maximization was employed offline on the training and validation data-sets to estimate unknown parameters. The estimated model parameters were used for real-time seizure state tracking. A classifier was then used to binarize the continuous seizure state. Our results on the experimental data (CHB-MIT EEG database) validate the accuracy of our proposed method and illustrate that the average accuracy, sensitivity, and false positive rate are 85.8%, 91.5%, and 14.3% respectively. This type of seizure state modeling could be used in further implementation of adaptive closed-loop vagus nerve stimulation applications.