Table 3.

Model comparison results

Model	F1@10	F1@25	F1@50	F1@75	MCC
Bi-LSTM$†$	25.9 ± 8.40	21.8 ± 9.03	15.0 ± 5.60	11.9 ± 6.26	62.4 ± 23.2
Bi-LSTM	63.7 ± 21.7	63.2 ± 22.0	50.8 ± 25.4	40.9 ± 28.4	78.8 ± 21.1
TCN	45.4 ± 16.8	42.7 ± 18.6	35.8 ± 14.8	27.0 ± 16.6	81.1 ± 12.9
ST-GCN	53.2 ± 21.2	51.5 ± 21.7	46.7 ± 22.5	37.6 ± 26.6	83.0 ± 11.5
MS-TCN	68.2 ± 29.4	66.8 ± 29.3	60.2 ± 30.5	54.9 ± 33.1	77.3 ± 22.2
MS-GCN	77.8 ± 15.3	77.8 ± 15.3	74.2 ± 21.0	57.0 ± 30.1	82.7 ± 15.5

Overview of the FOG segmentation performance in terms of the segment-wise F1@50 and sample-wise MCC for MS-GCN and the four strong baselines. The $†$denotes the sliding window FOG detection scheme. The best score is denoted in bold. All results were derived from the test set, i.e., subjects that the model had never seen