Objective: To investigate agreement between a recent automated FOG detection algorithm and expert video ratings, and the influence of dopaminergic medication and task context.

Background: Measuring freezing of gait (FOG) in Parkinson’s disease is a challenge due to its episodic nature. Automated FOG detection methods are promising for a number of reasons; however, they are typically validated in specific tasks while participants are OFF medication, thus raising questions about their external validity.

Method: Video recordings of 39 freezers performing 3 FOG-provoking tasks in the home situation (rapid alternating 360° turning in place, timed up and go, and personalized hotspot), ON and OFF medication, were manually annotated using detailed criteria. Automated FOG episode detection was performed using a recent open-source algorithm based on the foot FOG ratio and angular velocity obtained from 9-axis sensors. Associations between automated and expert-rated percentage time frozen and number of FOG episodes, as well as the frame-wise FOG agreement were investigated for medication and task influences.

Results: FOG occurred during 287 of the 539 trials (53.2%). Overall agreement between automated and expert ratings were negligible for the percentage time frozen (rho = -0.171, p = 0.004) and moderate for number of FOG episodes (rho = 0.583, p < 0.001). Importantly, associations between the automated and expert ratings for the number of FOG episodes was lower (Z = 3.6, p < 0.001) in ON (rho = 0.33, p = 0.001) than in OFF (rho = 0.659, p < 0.001). Conversely, frame-wise comparisons showed higher accuracy in ON than in OFF (median ON: 62.4%, OFF: 53.9%, Wilcoxon p < 0.001), driven by higher specificity (median ON: 69.5%, OFF: 64.9%, Wilcoxon p < 0.001) although with a trend towards decreased sensitivity (median ON: 28.7%, OFF: 31.5%, Wilcoxon p = 0.056). Exploring task differences revealed that the decrease in automated FOG detection sensitivity from OFF to ON was largest for turning in place (median ON-OFF difference in sensitivity – personalized hotspot: -9.1%, timed up and go: -7.9%, turning in place: -18.2%).

Conclusion: This study showed that overall agreement between the automated FOG detection and expert FOG ratings was low and further influenced by medication state and task context. Algorithms that consider these influences may deliver more robust FOG detection.

« Back to 2022 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/the-agreement-between-automated-and-expert-fog-annotations-during-an-at-home-protocol-influence-of-dopaminergic-medication-and-task/