Figure 2

Graphical representation on the method of evaluating HMMs. On the top illustration, the LFP is illustrated. Maximum Lyapunov exponent analysis is used to determine the initiation and termination of seizure activities (marked by solid vertical lines). The seizure detection horizon is defined as a 1 min time window centered on the EcSOT (marked by dotted vertical line) and is denoted by regions B and C. After the initial training process (using 40% of the overall dataset) on the full 7-D and 14-D feature space, the performance of each HMM is evaluated using the validation set (20% of the overall dataset) to determine the HMM_opt7D and HMM_opt14D topology. Afterwards, feature reduction from mRMR analysis and AICc are used to find a suitable HMM_AIC that balance the LL against the number of model parameters. The statistical tests (TP and TN) as well as optimality index (O) are then evaluated. On the bottom illustration, the intracellular activity before the EcSOT is partitioned based on its polarity characteristics into three states: hyperpolarizing (Pre_h), depolarizing (Pre_d) and a mixture of hyperpolarizing and depolarizing (Pre_m) activities [35]. This information is then compared to the γ_j(t) associated with the tonic firing patterns in the HMM to evaluate the correlation between the multiple model states with the intracellular dynamics.