Assessment of Parkinsonian gait in older adults with dementia via human pose tracking in video data

Andrea Sabo, Sina Mehdizadeh, Kimberley-Dale Ng, Andrea Iaboni, Babak TaatiJournal of NeuroEngineering and Rehabilitation 2020 · DOI

Abstract

Gait features extracted from both 2D and 3D videos are correlated to UPDRS-gait and SAS-gait scores of parkinsonism severity in gait. Vision-based systems have the potential to be used as tools for longitudinal monitoring of parkinsonism in residential settings.

Introduction

Parkinsonism, a clinical syndrome characterized by tremors, bradykinesia (slowness), and gait instability, is highly prevalent among individuals with dementia. These symptoms can arise from neurodegenerative processes or as a side effect of antipsychotic medications. Early detection is critical, particularly for medication-induced parkinsonism, which is often reversible if identified promptly.

Currently, the gold standard for assessment involves clinical scales such as the Unified Parkinson’s Disease Rating Scale (UPDRS) or the Simpson-Angus Scale (SAS). However, these evaluations require skilled clinicians and are typically performed during infrequent clinic visits. While wearable sensors offer a more continuous alternative, they are frequently rejected by patients with dementia due to the discomfort of physical contact. This study explores the use of passive computer vision—specifically human pose tracking—to provide objective, longitudinal gait analysis without requiring patients to wear devices or alter their natural behavior.

Experimental Methodology

The researchers implemented a monitoring system within a specialized dementia unit hallway. A Microsoft Kinect v2 sensor was mounted on the ceiling to capture walking bouts. To maintain privacy and ensure accurate data association, the system only recorded when a specific participant, identified via a Radio Frequency Identification (RFID) tag, entered the hallway.

Experimental setup showing (a) the ceiling-mounted Kinect sensor, (b) the RFID tag worn by the participant, and (c) the RFID reader.

The study compared two distinct approaches to feature extraction:

  1. 3D Features: Utilizing the Kinect SDK, 16 biomechanical features (e.g., walking speed, step length, and "Margin of Stability") were extracted based on 3D joint coordinates.
  2. 2D Features: Using OpenPose, a deep learning-based pose estimation library, 8 features were extracted from standard color video. To account for the lack of depth information in 2D video, horizontal and vertical distances were normalized by the distance between the participant's hips.

Multivariate ordinal logistic regression was then used to predict clinical gait scores (ranging from 0 to 4) assigned by a clinical expert based on these automated features.

Results and Discussion

The study utilized a dataset of 401 walking bouts collected from 14 older adults with dementia. A significant finding was the parity between 2D and 3D feature performance. The models achieved a predictive accuracy for UPDRS-gait of 61.4% using 2D features and 62.1% using 3D features. For SAS-gait, the accuracies were 47.4% and 48.5%, respectively.

While these percentages may seem modest for standard classification, they represent a strong correlation in a clinical context where the ground truth is subjective. Most discrepancies between the model and the clinician were only one category apart, mirroring the natural inter-rater variability found between human experts. For 2D data, the most influential predictors included step count, cadence, and the coefficient of variation (CV) of step time. Biomechanical cues like the "Margin of Stability" and "Symmetry Index" also proved to be strong predictors, confirming that the system captures the same indicators of shuffling or asymmetry that clinicians look for.

Conclusion

The parity between 2D standard video and 3D depth data is a major win for scalability, as standard color cameras are cheaper and have a longer effective range than specialized depth sensors. Although limited by a small sample size of 14 participants, this research demonstrates that vision-based systems can effectively bridge the gap in longitudinal monitoring. Future work utilizing deep learning architectures like RNNs or LSTMs may further refine these gait "signatures," moving us closer to smart residential care environments that can automatically alert staff to mobility declines and prevent falls.

Summary updated

Abstract

Gait features extracted from both 2D and 3D videos are correlated to UPDRS-gait and SAS-gait scores of parkinsonism severity in gait. Vision-based systems have the potential to be used as tools for longitudinal monitoring of parkinsonism in residential settings.
BibTeX
@article{sabo2020assessment,
  author = {Andrea Sabo and Sina Mehdizadeh and Kimberley-Dale Ng and Andrea Iaboni and Babak Taati},
  title = {{Assessment of Parkinsonian gait in older adults with dementia via human pose tracking in video data}},
  journal = {Journal of NeuroEngineering and Rehabilitation},
  year = {2020},
  doi = {10.1186/s12984-020-00728-9},
}
Figure