The　adaptivity　and　real-time　performance　of　feature　extraction　method　are　crucial　in　brain-computer　interface.　Based　on　Hilbert-Huang　transform　(HHT)　and　common　spatial　subspace　decomposition　(CSSD)　algorithm,　a　novel　feature　extraction　method,　denoted　as　HCSSD,　was　proposed.　Firstly,　the　motor　imagery　electroencephalography　(EEG)/electrocorticography　(ECoG)　was　preprocessed,　and　a　relative　distance　criterion　was　defined　to　select　the　optimal　combination　of　channels.　Secondly,　Hilbert　instantaneous　energy　spectrum　and　marginal　energy　spectrum　of　EEG/ECoG　were　calculated　to　extract　time　feature　and　frequency　feature　respectively.　Then　CSSD　was　applied　to　extract　spatial　feature.　Furthermore,　serial　feature　fusion　strategy　was　adopted　to　obtain　time-frequency-spatial　feature.　Finally,　learning　vector　quantization　neural　network　was　designed　to　classify　the　EEG/ECoG　data.　The　average　recognition　accuracy　was　92%　for　the　left　small　finger　and　tongue　motor　imagery　ECoG　tasks.　Experiment　results　show　that　HCSSD　can　enhance　the　adaptivity　and　real-time　performance　of　feature　extraction,　with　the　recognition　accuracy　improved.　This　method　provides　a　new　idea　for　the　application　of　portable　BCI　system　in　rehabilitation　field.