Table 2.

Error rates and characteristics of TCN and comparison models.

Method				# params	L-M	L-H	VT-H	MAX	Combined	$\dot{\text{V}}{\text{O}}_{2\text{peak}}$
Random forest21				298,106	20 ± 134	144 ± 187	54 ± 169	15 ± 111	62 ± 164	134 ± 151
SLSTM23				21,589	93 ± 133	52 ± 145	−58 ± 148	−16 ± 178	21 ± 160	−48 ± 243
SLSTM + {HRR,${\dot{\text{V}}}_{\text{E}}$}				21,845	52 ± 132	7 ± 163	−72 ± 152	−37 ± 163	−10 ± 159	−17 ± 205

	Filters	Receptive field (s)	Validation loss
TCN-VO2 A	24	218	0.02506	19,921	−23 ± 116	14 ± 136	−49 ± 115	−32 ± 117	−22 ± 122	18 ± 182
TCN-VO2 B	16	187	0.02525	8081	−7 ± 109	30 ± 138	−25 ± 115	−2 ± 113	−1 ± 120	62 ± 162
TCN-VO2 C	16	76	0.02585	5393	−21 ± 113	16 ± 159	−57 ± 136	3 ± 123	−17 ± 137	93 ± 214
TCN-VO2 D	16	91	0.02586	6241	−24 ± 115	31 ± 142	−54 ± 123	9 ± 117	−11 ± 128	80 ± 165
TCN-VO2 NH	24	1	0.04895	361	10 ± 127	133 ± 179	69 ± 129	75 ± 123	72 ± 149	223 ± 196
TCN-VO2 A(HR)	24	218	0.10733	19,057	197 ± 286	133 ± 373	40 ± 412	145 ± 445	127 ± 379	−220 ± 445

Errors are reported as mean ± SD [ml min⁻¹].

SLSTM stacked long short-term memory.