Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Forecasting of Pathological Gait via IMU Navigation: A Few-Shot and Generative Learning Framework for Wearable Devices

Wenwen Zhang, Hao Zhang, Zenan Jiang, Amir Servati, Peyman Servati

TL;DR

This work tackles the scarcity and privacy constraints of gait datasets by introducing GaitMotion, a few-shot multitask framework using wearable IMU data for real-time pathological gait analysis. It combines ground-truth labeled sequences, a CNN-based stride-length estimator, and a generative time-series augmentation module to synthesize rare gait patterns, achieving a $65\%$ stride-length accuracy gain over ZUPT and good transferability to real patient data. The approach is validated with transfer learning on the eGait dataset, showing that pre-training on GaitMotion enables accurate gait parameter estimation with minimal patient-labeled data ($RMSE\approx0.201\,\mathrm{m}$). Overall, GaitMotion offers a data-efficient, scalable platform for clinical gait assessment and personalized digital health services, with implications for federated deployment across institutions.

Abstract

Current gait analysis faces challenges in various aspects, including limited and poorly labeled data within existing wearable electronics databases, difficulties in collecting patient data due to privacy concerns, and the inadequacy of the Zero-Velocity Update Technique (ZUPT) in accurately analyzing pathological gait patterns. To address these limitations, we introduce GaitMotion, a novel machine-learning framework that employs few-shot learning on a multitask dataset collected via wearable IMU sensors for real-time pathological gait analysis. GaitMotion enhances data quality through detailed, ground-truth-labeled sequences and achieves accurate step and stride segmentation and stride length estimation, which are essential for diagnosing neurological disorders. We incorporate a generative augmentation component, which synthesizes rare or underrepresented pathological gait patterns. GaitMotion achieves a 65\% increase in stride length estimation accuracy compared to ZUPT. In addition, its application to real patient datasets via transfer learning confirms its robust predictive capability. By integrating generative AI into wearable gait analysis, GaitMotion not only refines the precision of pathological gait forecasting but also demonstrates a scalable framework for leveraging synthetic data in biomechanical pattern recognition, paving the way for more personalized and data-efficient digital health services.

Real-Time Forecasting of Pathological Gait via IMU Navigation: A Few-Shot and Generative Learning Framework for Wearable Devices

TL;DR

This work tackles the scarcity and privacy constraints of gait datasets by introducing GaitMotion, a few-shot multitask framework using wearable IMU data for real-time pathological gait analysis. It combines ground-truth labeled sequences, a CNN-based stride-length estimator, and a generative time-series augmentation module to synthesize rare gait patterns, achieving a stride-length accuracy gain over ZUPT and good transferability to real patient data. The approach is validated with transfer learning on the eGait dataset, showing that pre-training on GaitMotion enables accurate gait parameter estimation with minimal patient-labeled data (). Overall, GaitMotion offers a data-efficient, scalable platform for clinical gait assessment and personalized digital health services, with implications for federated deployment across institutions.

Abstract

Current gait analysis faces challenges in various aspects, including limited and poorly labeled data within existing wearable electronics databases, difficulties in collecting patient data due to privacy concerns, and the inadequacy of the Zero-Velocity Update Technique (ZUPT) in accurately analyzing pathological gait patterns. To address these limitations, we introduce GaitMotion, a novel machine-learning framework that employs few-shot learning on a multitask dataset collected via wearable IMU sensors for real-time pathological gait analysis. GaitMotion enhances data quality through detailed, ground-truth-labeled sequences and achieves accurate step and stride segmentation and stride length estimation, which are essential for diagnosing neurological disorders. We incorporate a generative augmentation component, which synthesizes rare or underrepresented pathological gait patterns. GaitMotion achieves a 65\% increase in stride length estimation accuracy compared to ZUPT. In addition, its application to real patient datasets via transfer learning confirms its robust predictive capability. By integrating generative AI into wearable gait analysis, GaitMotion not only refines the precision of pathological gait forecasting but also demonstrates a scalable framework for leveraging synthetic data in biomechanical pattern recognition, paving the way for more personalized and data-efficient digital health services.
Paper Structure (17 sections, 9 figures, 5 tables)

This paper contains 17 sections, 9 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Wearable Sensors and Gait Analysis
  3. Gait Analysis with Few-shot Machine Learning
  4. Methods
  5. ML Framework Implementation
  6. GaitMotion Dataset
  7. Data Acquisition
  8. Data augmentation via generative model
  9. Transfer Learning
  10. Ethical Considerations
  11. Results and Discussion
  12. GaitMotion Analysis
  13. Transfer Learning Validate
  14. Gait Parameter Prediction Outcomes
  15. Asymmetry and Outlier Analysis
  16. ...and 2 more sections

Figures (9)

  • Figure 1: (a). Experimental setup for data collection: IMUs are attached to the feet, and data is recorded while participants walk over a 1-meter-long Gaitrite mat. (b)-(d). Data type with different walking patterns and ground truth values after sensor fusion and gravity subtraction. The purple lines are the ZUPT segmentation results with missed and overcounted steps. (b). Normal walking usually displays consistent stride length and fixed cadence. (c). Parkinson's gaits have alternating small and shuffling gaits. (d). Stroke gaits exhibit asymmetric gait status where the normal foot has stronger power in moving control, while less motion is observed in the dragging foot over the accelerometer and gyroscope. (f). A detailed comparison zooming in on both normal and dragging feet.
  • Figure 2: Schematic process used by GaitMotion for multiple tasks learning. (a) The walking process is segmented into gait cycles, encompassing both a stance phase and a swing phase. For stride segmentation, the emphasis is placed on determining the heel strike and toe-off events, which are pivotal points in all patterns of gaits. (b) The segmentation process takes accelerometer and gyroscope data and learns to identify the events as shown in (c). Stance time is the time between the heel strike and toe-off, and swing time is the time between toe-off and the next heel strike. (d) Gait parameters such as stride length are calculated using a CNN network that takes in raw accelerometer and gyroscope data for each step, as illustrated in (e).
  • Figure 3: Comparison of stride length estimation and ground truth with subject ID 1-10. The results include performance for Parkinson's, Stroke, and Normal gaits separately.
  • Figure 4: Boxplot errors to separately present the summary of stride length in both pathological (Parkinson's and Stroke) and normal gaits for individual participants for comparison.
  • Figure 5: Comparison of stride length variance and ground truth calculated from Gaitrite system with subject ID 1-10. The stride-to-stride gait variance results are obvious that Parkinson's gaits have the most unstable while normal and Stroke gaits have comparable levels.
  • ...and 4 more figures