A sparse Laguerre-Volterra autoregressive model has been developed as feature extraction from subdural human EEG data for seizure prediction in temporal lobe epilepsy. The use of Laguerre-Volterra kernel can compactly yield an autoregressive model of longer system memory without increasing the number of the coefficients. In 6 sets of seizure, we used a sparse Laguerre-Volterra autoregressive model with 6 coefficients and the decay parameter of 0.2 and obtained the 10-fold cross-validation prediction results of high Matthews correlation coefficients (0.7-1) and low prediction errors (<15%). These results demonstrate that the sparse Laguerre-Volterra autoregressive model is effective in the feature extraction for seizure prediction. Finally, this sparse Laguerre-Volterra method can be easily adapted to a potentially more powerful nonlinear autoregressive model as the feature extraction rather than linear autoregressive model that we are currently using.