Table 2.

We compare the baseline results of three deep learning models (CNN-LSTM, ResNet, and DeepConvLSTM) with Accuracy (higher is better) and F1 score (higher is better) using three cross-validation methods (k-fold, leave-recordings-out, and leave-one-subject-out) with five sensors. The standard deviation is reported with the ± symbol. We embolden the model per column.

Model	k-Fold		Leave-Recordings-Out		Leave-One-Subject-Out
	Accuracy	F1	Accuracy	F1	Accuracy	F1
CNN-LSTM	0.832 $(\pm 0.01)$	0.830 $(\pm 0.02)$	0.632 $(\pm 0.03)$	0.630 $(\pm 0.02)$	0.565 $(\pm 0.05)$	0.556 $(\pm 0.04)$
ResNet	0.864 $(\pm 0.01)$	0.864 $(\pm 0.01)$	0.600 $(\pm 0.02)$	0.603 $(\pm 0.03)$	0.531 $(\pm 0.05)$	0.527 $(\pm 0.03)$
DeepConvLSTM	0.735 $(\pm 0.02)$	0.730 $(\pm 0.04)$	0.563 $(\pm 0.05)$	0.565 $(\pm 0.05)$	0.525 $(\pm 0.06)$	0.515 $(\pm 0.04)$