Freezing of gait assessment with inertial measurement units and deep learning: effect of tasks, medication states, and stops

Table 7 Model comparison results in terms of training on different FOG-provoking task or medication state

Model	Test data	Segment-F1@50		Sample-F1
Model	Test data	Mean	t-test (#pair)	Mean	t-test (#pair)
Model_TUG	TUG	0.70	t(12) = 6.19, p < 0.005	0.75	t(12) = 6.07, p < 0.005
Model_360Turn	TUG	0.10	t(12) = 6.19, p < 0.005	0.15	t(12) = 6.07, p < 0.005
Model_360Turn	360Turn	0.53	t(12) = 2.14, p = 0.055	0.62	t(12) = 6.07, p < 0.005
Model_TUG	360Turn	0.44	t(12) = 2.14, p = 0.055	0.56	t(12) = 6.07, p < 0.005
Model_Off	Off	0.52	t(12) = − 0.06, p = 0.952	0.66	t(12) = 0.05, p = 0.957
Model_On	Off	0.53	t(12) = − 0.06, p = 0.952	0.65	t(12) = 0.05, p = 0.957
Model_On	On	0.68	t(11) = 0.57, p = 0.579	0.76	t(11) = 1.07, p = 0.307
Model_Off	On	0.65	t(11) = 0.57, p = 0.579	0.71	t(11) = 1.07, p = 0.307

We investigated the generalization of task-specific models to unseen tasks and the generalization of medication-specific models to unseen medication states. The third and fourth column depicts Segment-F1@50 averaged over all subjects and the paired t-test result. The fifth and sixth columns depict the Sample-F1 averaged over all subjects and the paired t-test result. The number of subjects (pairs) was 12 for TUG, 360Turn, and Off-state, while only 11 subjects were considered for On-state due to technical problems for subject 5 during On-medication state measurements

ISSN: 1743-0003