Precision Balance Assessment in Parkinson's Disease: Utilizing Vision-Based 3D Pose Tracking for Pull Test Analysis

Abstract

Postural instability is a common complication in advanced Parkinson's disease (PD) associated with recurrent falls and fall-related injuries. The test of retropulsion, consisting of a rapid balance perturbation by a pull in the backward direction, is regarded as the gold standard for evaluating postural instability in PD and is a key component of the neurological examination and clinical rating in PD (e.g., MDS-UPDRS). However, significant variability in test execution and interpretation contributes to a low intra- and inter-rater test reliability. Here, we explore the potential for objective, vision-based assessment of the pull test (vPull) using 3D pose tracking applied to single-sensor RGB-Depth recordings of clinical assessments. The initial results in a cohort of healthy individuals (n = 15) demonstrate overall excellent agreement of vPull-derived metrics with the gold standard marker-based motion capture. Subsequently, in a cohort of PD patients and controls (n = 15 each), we assessed the inter-rater reliability of vPull and analyzed PD-related impairments in postural response (including pull-to-step latency, number of steps, retropulsion angle). These quantitative metrics effectively distinguish healthy performance from and within varying degrees of postural impairment in PD. vPull shows promise for straightforward clinical implementation with the potential to enhance the sensitivity and specificity of postural instability assessment and fall risk prediction in PD.

Keywords: Parkinson’s disease; RGB-Depth sensor; pose tracking; postural instability; pull test; test of retropulsion.

Figure 1

Experimental setup and analysis approach. (A) The pull test execution is recorded using a single RGB-Depth sensor positioned 2 m in front of the assessed patient. A 17 keypoint pose model is then estimated from the RGB frames and projected into 3D space based on the sensor’s depth frames. (B) Pull onset and magnitude are determined from 3D shoulder acceleration, utilizing thresholding relative to baseline shoulder motion. The amplitude of retropulsion and latency of balance recovery are assessed through 3D trunk bending dynamics in the backward direction. Steps are identified by thresholding 3D bilateral ankle velocities. Abbreviations: acc: acceleration; vel: velocity.

Figure 2

Group comparison of vPull test metrics between patients and controls (* indicate a significant difference between groups). (A–D) General test characteristics, including pull magnitude, percentage of individuals displaying a stepping response, percentage of individuals showing successful balance recovery, and corresponding rating of pull test performance according to the MDS-UPDRS scheme (grade 1—slight, grade 3—moderate). (E–H) Detailed metrics characterizing the stepping response. (I,J) Detailed metrics characterizing the truncal response and balance recovery. (K) Low-dimensional embedding of the above quantitative features (E–J) labeled by pull test performance rating using UMAP. Abbreviations: HS: healthy subjects; PD: patients with Parkinson’s disease; MDS-UPDRS: The Movement Disorder Society Unified Parkinson’s Disease Rating Scale; UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction.