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

Table 5 Overview of the performance of Model_Clinical and Model_Stop

Model	Test data	ICC (%TF)	ICC (#FOG)	Segment-F1@50	Sample-F1
Model clinical	TUG	0.95, CI = [0.86, 0.99]	0.97, CI = [0.89, 0.99]	0.67	0.72
	360Turn	0.94, CI = [0.78, 0.98]	0.84, CI = [0.57, 0.95]	0.45	0.58
	Off-state	0.94, CI = [0.74, 0.99]	0.94, CI = [0.76, 0.99]	0.55	0.65
	On-state	0.92, CI = [0.76, 0.98]	0.90, CI = [0.70, 0.97]	0.64	0.69
	All trials	0.92, CI = [0.68, 0.98]	0.95, CI = [0.72, 0.99]	0.60	0.67
Model stop	TUG	0.78, CI = [0.42, 0.93]	0.77, CI = [0.41, 0.93]	0.65	0.67
	360Turn	0.98, CI = [0.90, 1.00]	0.65, CI = [0.16, 0.89]	0.33	0.47
	Off-state	0.89, CI = [0.52, 0.97]	0.83, CI = [0.53, 0.95]	0.55	0.62
	On-state	0.98, CI = [0.96, 1.00]	0.69, CI = [0.22, 0.91]	0.62	0.64
	All trials	0.95, CI = [0.73, 0.99]	0.79, CI = [0.46, 0.94]	0.59	0.63

We investigated the overall performance of the generic model trained for clinical settings (i.e., excluding stopping) with standardized measurements: Model_Clinical and the generic model trained to work towards FOG detection in daily life (i.e., including stopping): Model_Stop. Results show that Model_Clinical has a strong agreement with the experts in terms of %TF (ICC = 0.92) and #FOG (ICC = 0.95). Also, Model_Stop has a strong agreement with the experts in terms of %TF (ICC = 0.95) and a moderately strong agreement in terms of #FOG (ICC = 0.79). We also showed the relative performance of the two models for each of the four conditions with less data variety: (1) TUG trials, (2) 360Turn trials, (3) Off-medication trials, and (4) On-medication trials. For the four conditions, Model_Clinical was evaluated on trials that excluded stopping, while Model_Stop was evaluated on trials that included stopping

ISSN: 1743-0003