Feasibility of Using Sensor Insoles to Automatically Segment Timed-up-and-go Tests in Parkinson’s Disease Population

Research Objectives: To investigate the feasibility of using sensor insole data and rule-based algorithms to automatically segment the Timed-Up-and-Go (TUG) test in patients diagnosed with Parkinson’s Disease (PD).

Design: Nonrandomized controlled trial

Setting: Research facilities at the Johns Hopkins Outpatient Center and Center for Limb Loss and Mobility at the VA.

Participants: Using population-based sampling, data were collected from 41 participants (17F/24M) with a physician-diagnosis of PD (66±9 years, modified Hoehn & Yahr score of 2.3±0.53, MDS-UPDRS Part III score of 23 ± 12). Ten participants were excluded from analysis due to lack of sensor insole data or unanticipated movements during the task.

Interventions: Not applicable

Main Outcome Measures: Participants performed the TUG test while data from sensor insoles were collected. Data included three degree of freedom acceleration and angular velocity and force from insoles, and frame-by-frame annotations of the participants’ actions. All data are accessible on the Synapse data repository by SAGE Bionetworks under the “WearGait-PD: Wearables for gait in Parkinson’s Disease and age-matched controls” Version 1.0 dataset. Raw data were filtered, normalized, then reviewed to generate rules for the segmentation algorithm. The algorithm used total force, acceleration magnitude, and rotational velocity about the Y and Z axes to segment the TUG test. Krippendorff’s alpha was calculated to quantify agreement between the manual and automatic segmentation. The Krippendorff’s alpha was interpreted as acceptable (α>0.80), moderate (0.67< α < 0.79), or poor (α < 0.67). Qualitative analysis of scatter plots between Krippendorff’s alpha and MDS-UPDRS Part III score was also done.

Results: Based on our analysis, 71% (n=22) of the TUG trials had acceptable agreement between the automated and manual segmentation methods. 22.5% (n=7) had moderate agreement, and 6.5% (n=2) had poor agreement. Scatter plots between Krippendorff’s alpha and MDS-UPDRS Part III score showed patients with higher MDS-UPDRS Part III scores tended to have poor agreement between manual and automatic segmentation.

Conclusions: Using our rule-based algorithm for automated TUG segmentation, 71% of TUG trials had acceptable agreement with the manually annotated data. These findings suggest there is merit to utilizing insole data to develop such algorithms for use in more refined TUG analysis. 29% of the TUG trials had moderate to poor agreement, indicating further tuning of the algorithm may be needed to account for movements characteristic of severe mobility issues in PD.

Author(s) Disclosures: None