Algorithm 1 Pseudo-code of individual functions

1: function Extract feature($S_{\mathrm{ecog}},\mathcal{X}$)     input: $S_{\mathrm{ecog}}=20\mathrm{s};\mathrm{all}\mathrm{channels}$     output: $\mathcal{X}\in {\mathbb{R}}^K$ 2:         $\mathcal{X}$ = $\left\{x_1,x_2,\dots x_N\right\}$ 3: end function       4: function Classify($\mathcal{X},\mathcal{Y},\mathcal{V},\mathcal{U}$)     input: $\mathcal{X}\in {\mathbb{R}}^K$     output: $\mathcal{Y},\mathcal{V},\mathcal{U}$ 5:         for each sample in $\mathcal{X}$ do 6:                 $(\mathcal{Y},\mathcal{V})=\theta \left(\mathcal{X}\right)$ (acquire soft and crisp labels) 7:         end for 8:         return $\mathcal{U}$ = $\left\{\mathcal{X}\left(P_{\mathcal{Y}|\mathcal{V}}<80\%\right)\right\}$                                              (obtain ambiguous samples) 9: end function

10: function Threshold($\widehat{\mathcal{Y}},PA$) 11:        if $70\%\mathrm{of}\widehat{\mathcal{Y}}\mathrm{is}\mathrm{classified}\mathrm{as}\mathrm{pre-ictal}\mathrm{state}$ then 12:                return $PA$ = 1 13:        else 14:                return $PA$ = 0 15:        end if 16: end function 17: function Active learner($\mathcal{X},\mathcal{U}\mathcal{Y},\mathcal{V},\mathcal{A}$)       input: $\mathcal{X},\mathcal{U},\mathcal{Y},\mathcal{V}$       output: $\mathcal{A}$ 18:        ${\theta }_{\mathrm{BGMM}}=\mathrm{BGMM}\left(\mathcal{X},\mathcal{U},\mathcal{Y},\mathcal{V}\right)$ (Obtain BGMM model) 19:        for each sample in $\mathcal{U}$ do 20:                $\left(\mathcal{Z},p\left(\mathcal{Z}\right)\right)=E\_M\left[{\theta }_{\mathrm{BGMM}}\left(\mathcal{X},\mathcal{Y},\mathcal{V}\right)\right]$ 21:        end for 22:        $\mathcal{A}=\left(\mathrm{Samples}\mathrm{in}\mathcal{Z}\mathrm{with}\mathrm{smallest}\left|p\right(\mathcal{Z}\left)\right|\right)$ 23: end function