Table 5.

The classification results for different models with offline and on-device training are compared. We show three different models (CNN-LSTM, ResNet, and DeepConvLSTM) with both offline and on-device training for the nursing activity dataset. The best scores are emboldened row-wise per metric. We report the same metrics for each combination of input data and evaluation methods. We embolden the best model per column.

Model	Offline Training		On-Device Training
	Accuracy	F1	Accuracy	F1
CNN-LSTM	0.352 $(\pm 0.07)$	0.310 $(\pm 0.06)$	0.527 $(\pm 0.02)$	0.492 $(\pm 0.03)$
ResNet	0.487 $(\pm 0.04)$	0.473 $(\pm 0.06)$	0.489 $(\pm 0.03)$	0.495 $(\pm 0.04)$
DeepConvLSTM	0.320 $(\pm 0.08)$	0.193 $(\pm 0.09)$	0.416 $(\pm 0.04)$	0.237 $(\pm 0.05)$