ALGORITHM 1.

Training and optimization procedures of EEG-fNIRS fusion based on our proposed framework.

Inputs: time domain features of EEG E1 = {(E1xm, ym)}; frequency domain features of EEG E2 = {(E2xm, ym)}; time domain features of fNIRS F1 = {(F1xm, ym)}; frequency domain features of fNIRS F2 = {(F2xm, ym)}; number of primary learning models T; primary learning models λ1, λ2, λ3, λ4; secondary learning model λ

Outputs: trained model M based on our proposed method; predicted labels L;

1. set E = E1 ∪ E2 = {(Exm, ym)},

2.   F = F1 ∪ F2 = {(Fxm, ym)}% Combine the time and frequency domain                  features

3. D1 = E’, D2 = F’% feature selection using ASO algorithm for E and F:

4. for t = 1, …, T:

5.   ht1 = λt (D1)% Train a primary individual learner ht1 and ht2 by           applying the primary

6.   ht2 = λt (D2)% learning models

7. end for

8. D’ = ∅; % Generate a new data set

9. for i = 1,…,k:% k-flod cross-validation

10. for t = 1,…,T:

11.   ht1’ = ht1 (E1xi), ht2’ = ht2 (F1xi),

12.   ht = ht1’ ∪ ht2’

13.   Zit = ht (xi)

14. end for

15. D’ = D’ ∪ ((Zi1, Zi2, …, ZiT), yi)

16. end for

17. h’ = λ (D’);% Train the secondary learner h’ by applying the secondary          learning

18.   % algorithm λ to the new data set D’

19. M (x) = h’ (h1 (x), h1 (x),…, hT (x))

20. return predicted labels L and trained model M based on our proposed      method