In　this　study,　we　explore　the　approach　using　Support　Vector　Machines　(SVM)　to　predict　the　high-level　cognitive　states　based　on　fMRI　data.　On　the　base　of　taking　voxels　in　the　brain　regions　related　to　problem　solving　as　the　features,　we　compare　two　feature　extraction　methods,　one　is　based　on　the　cumulative　changes　of　blood　oxygen　level　dependent　(BOLD)　signal,　and　the　other　is　based　on　the　values　at　each　time　point　in　the　BOLD　signal　time　course　of　each　trial.　We　collected　the　fMRI　data　while　participants　were　performing　a　simplified　4*4　Sudoku　problems,　and　predicted　the　complexity　(easy　vs.　complex)　or　the　steps　(1-step　vs.　2-steps)　of　the　problem　from　fMRI　data　using　these　two　feature　extraction　methods,　respectively.　Both　methods　can　produce　quite　high　accuracy,　and　the　performance　of　the　latter　method　is　better　than　the　former.　The　results　indicate　that　SVM　can　be　used　to　predict　high-level　cognitive　states　from　fMRI　data.　Moreover,　the　feature　extraction　based　on　serial　signal　change　of　BOLD　effect　can　predict　cognitive　states　better　because　it　can　use　abundant　and　typical　information　kept　in　BOLD　effect　data.　©　2009　Springer-Verlag　Berlin　Heidelberg.