Preserving Clinically Relevant Gait Features in Wearable-Sensor Pose Estimation for Parkinson’s Disease
1The University of Tokyo · 2Centre for Automation and Robotics (CAR), Spanish National Research Council
Ground truth (GT) and reconstructions with pose-only (P) and clinical-aware (C) fine-tuning, using six IMUs or the head-free five-IMU set. Qualitative example from virtual IMUs, original playback speed.
Take-home message
- Clinical-aware fine-tuning reduced joint-velocity and leg-swing errors vs. pose-only fine-tuning.
- The benefit held with a head-free five-IMU set, updating only 0.48% of parameters.
Main result
| C − P (absolute log-ratio error) | Six IMUs | Five IMUs |
|---|---|---|
| Joint velocity | −0.025 | −0.037 |
| Sagittal leg swing | −0.027 | −0.022 |
Negative values mean lower error with clinical-aware tuning. All four 95% confidence intervals exclude zero. Virtual IMUs from 781 walks of 23 people with Parkinson’s disease; subject-wise cross-validation.
Cite
K. Yang, H. Sun, H. Hamada, K. Kikuchi, A. Yamashita, D. Torricelli, and Q. An, “Preserving Clinically Relevant Gait Features in Wearable-Sensor Pose Estimation for Parkinson’s Disease,” International Conference on NeuroRehabilitation (ICNR 2026), Seoul, Poster P-113.
Contact
Kai Yang — yangkai@robot.t.u-tokyo.ac.jp