Table A3.

We show how permuting the order of the subjects can lead to a small variation in the classification performance.

Dataset	biLSTM dimension	Self-attention dimension	γ 2	Permutation	Mean AUC	Median AUC
FBIRN	100	48	0.05	Randomly done	0.86	0.861
FBIRN	128	64	0.025	Randomly done	0.865	0.875
FBIRN	100	48	0.05	Default order	0.86	0.889
FBIRN	128	64	0.025	Default order	0.858	0.875