Table 3.

Prediction performance of the proposed model (AT+AGCN+GMM) and comparison to baseline models. We also compare the run time of each deep learning model used in the experiment. Performance metrics are presented as mean ± 95% confidence interval. Deep learning models are indicated by italics. Abbreviations are described in text. ${}^a$ Total score with FOG item (3.11) subtracted. ${}^{†}$ p < 0.05, improvement in RF vs. SVM. ${}^{‡}$ p < 0.05, improvement in -LR vs. -L. * p < 0.05, improvement in deep learning models vs. preceding row. ${}^{\S }$ p < 0.05, improvement in multi-task vs. single task prediction.

	Medication State	FOG Score	MDS-UPDRS-III ${}^{\mathit{a}}$	Inference Time
Model	(F1)	(F1)	(RMSE)	(Seconds)
SVM-L	$0.540\pm 0.016$	$0.429\pm 0.026$	$9.346\pm 0.138$	-
RF-L	$0.594\pm 0.012{}^{†}$	$0.553\pm 0.038{}^{†}$	$9.189\pm 0.301$	-
SVM-LR	$0.616\pm 0.017{}^{‡}$	$0.608\pm 0.031{}^{‡}$	$8.714\pm 0.101{}^{‡}$	-
RF-LR	$0.657\pm 0.019{}^{†,‡}$	$0.684\pm 0.040{}^{†,‡}$	$7.918\pm 0.427{}^{†,‡}$	-
TCN [52]	$0.875\pm 0.017$ *	$0.851\pm 0.020$ *${}^{,\S }$	$4.551\pm 0.276$ *	0.0055
GCN [53]	$0.913\pm 0.015$ *	$0.929\pm 0.021$ *${}^{,\S }$	$4.023\pm 0.373$ *	0.0433
AGCN [34]	$0.949\pm 0.010$ *	$0.948\pm 0.018$ *${}^{,\S }$	$3.703\pm 0.300$ *	0.0435
AT+AGCN	$0.955\pm 0.021$	$0.955\pm 0.026{}^{\S }$	$3.555\pm 0.394$ *	0.0469
AT+AGCN+GMM	$0.975\pm 0.018$	$0.967\pm 0.022$	$2.753\pm 0.440$ *	0.0471