Table 2.

Mean classification accuracy across participants for a given training method, sequence length, and architecture. Number of parameters per model is displayed below each neural network architecture. Measured computational costs normalized to the cross-validated group model are displayed below each cross-participant training method. The best performing architecture for each training method and sequence length is shown in bold.

Training Method	Sequence Length	MPCRNN	BDRLSTM	BDLSTM	LSTM	2L-LSTM	CNN	ANN
{Normalized Computational Cost}		(6.2 M)	(4.2 M)	(4.2 M)	(1.9 M)	(0.7 M)	(1.8 M)	(1.1–6.5 M)
	5	0.791	0.768	0.760	0.766	0.771	0.635	0.749
Cross-Validated Group Model	10	0.820	0.785	0.777	0.778	0.793	0.657	0.777
{1.0}	20	0.850	0.839	0.839	0.811	0.820	0.732	0.834
	30	0.868	0.852	0.853	0.849	0.834	0.711	0.862
	5	0.781	0.748	0.748	0.682	0.754	0.627	0.743
Optimal-Stopping Val-Set Group Model	10	0.819	0.786	0.785	0.689	0.782	0.599	0.774
{0.12}	20	0.834	0.825	0.825	0.802	0.809	0.684	0.828
	30	0.850	0.860	0.846	0.838	0.824	0.691	0.855
	5	0.791	0.775	0.771	0.757	0.773	0.731	0.767
7-Classifier Ensemble Model	10	0.822	0.806	0.805	0.789	0.797	0.757	0.798
{0.65}	20	0.842	0.814	0.803	0.833	0.798	0.757	0.834
	30	0.865	0.833	0.812	0.808	0.808	0.732	0.838
	5	0.768	0.780	0.778	0.765	0.773	0.701	0.769
28-Classifier Ensemble Model	10	0.800	0.804	0.806	0.792	0.801	0.690	0.802
{0.52}	20	0.809	0.807	0.816	0.826	0.799	0.731	0.833
	30	0.837	0.825	0.818	0.809	0.810	0.694	0.841