Epileptic seizures reflect runaway excitation that severely hinders normal brain functions. With EEG recordings reflecting real-time brain activity, it is essential to both predict seizures and improve the classification of seizures in EEG signs. Towards this aim, nonlinear tools are strongly recommended to select the seizure-sensitive features prior to classification. However, the choice of the feature remains challenging. With the multitude of entropy parameters available in literature, and in order to perform a judicious selection of features that are fed to classifiers, this paper presents a comparative study of a host of candidate promising feature extraction techniques. Four entropy features namely Approximate Entropy, Sample Entropy and Renyi entropy of order 2 and Renyi entropy of order 3, were implemented as the standard techniques. Three kernel-based features namely Triangular Entropy, Spherical Entropy and Cauchy entropy were implemented. The former and latter entropies were computed from EEG recordings during induced seizures in three distinct phases: the pre-ictal (pre-seizure) phase, the ictal (seizure) phase, and the post-ictal (post-seizure) phase. Results showed that, among kernel-based methods, Spherical entropy features exhibited the largest parameter sensitivity to (Seizure-Normal) phase changes with the highest normalized relative separation (100%). The sample entropy feature in turn showed the most sensitive to EEG phase changes with the highest relative separation (94.85%), among the studied entropy alternatives.