Gait Phase Detection Based on Muscle Deformation with Static Standing-Based Calibration

Tamon Miyake, Shintaro Yamamoto, Satoshi Hosono, Satoshi Funabashi, Zhengxue Cheng, Cheng Zhang, Emi Tamaki, Shigeki SuganoSensors 2021 · DOI

TL;DR

This paper introduces a novel gait phase detection system that uses muscle deformation sensors for calibration, eliminating the need for users to walk during the setup process. By leveraging static standing postures and a logistic regression model adjusted with gyroscope data, the system achieves a median accuracy of 90% for detecting foot-contact and foot-off states, offering a more convenient and accessible solution for gait monitoring.

What Problem Does This Paper Address?

Monitoring human gait is crucial for various applications, including robotic assistance and health assessment, especially for individuals experiencing age-related declines in mobility. Traditional gait phase detection methods often rely on wearable sensors that measure inertial, electromyography (EMG), or force myography data. However, many existing machine learning-based approaches require extensive manual calibration data collected while the user is actively walking, which can be time-consuming and burdensome. This paper addresses the challenge of creating a gait phase detection system that is both accurate and user-friendly by developing a calibration method that doesn't require walking data, thereby simplifying the setup process and improving overall usability.

Methods

The core of this research lies in a novel gait phase detection system that utilizes muscle deformation information. Instead of relying on walking data for calibration, the system employs a static standing-based calibration approach.

Here's a breakdown of the key methods:

  • Muscle Deformation Sensing: The system uses the FirstVR wearable device, equipped with 14 muscle deformation sensors. These sensors, based on near-infrared light and photodiodes, measure skin deformation caused by muscle bulging. The sensors are attached to the calf muscles, which are closely related to ankle and foot movement.
  • Static Standing Calibration: The calibration process involves subjects assuming a set of discrete static standing postures. These postures are designed to represent different foot positions (forward/backward) and states (foot-flat, heel-off, foot-off, heel-contact). This allows the system to learn the muscle deformation patterns associated with these key gait phases without requiring the user to walk.
  • Machine Learning Algorithm: A logistic regression (LR) algorithm is employed for binary classification of gait phases (stance vs. swing). LR was chosen for its ability to output class probabilities and its relative simplicity compared to deep learning methods, making it suitable for limited training data and faster calibration.
  • Probability Adjustment: To improve accuracy and account for potential discrepancies between standing and walking data, the LR's probability output is adjusted using angular velocity data from an integrated gyroscope. Specifically, the system uses the angular velocity related to the anterior direction of the body to refine the decision boundary. This "LR-a" (LR with adjustment) approach aims to better reflect the dynamic nature of walking.
  • Data Processing: The system captures muscle deformation data and angular velocity at specific sampling frequencies (50 Hz for FirstVR). The data is then processed to extract relevant features for the LR algorithm.

Datasets

  • Training Data: The system was trained using muscle deformation data collected from subjects assuming seven specific static standing postures (as illustrated in Figure 4). Various combinations of these postures were explored as experimental conditions for calibration.
  • Evaluation Data: Walking data was collected from subjects on a treadmill. Videos were recorded simultaneously to serve as the ground truth for gait phase detection. Approximately 19,651 frames were used for evaluation.

Key Insights and Ideas

  • Calibration Without Walking is Feasible: The most significant insight is that accurate gait phase detection can be achieved by calibrating the system using only static standing postures. This bypasses the user burden of collecting walking data for calibration, a major step towards practical usability.
  • Muscle Deformation as a Proxy for Gait: The study demonstrates that calf muscle deformation patterns, captured by simple wearable sensors, are sufficiently informative to distinguish between stance and swing phases. This suggests that muscle activity changes predictably with body weight distribution and joint movement during gait.
  • Gyroscope Data Enhances Accuracy: Incorporating angular velocity data from a gyroscope to adjust the probability output of the logistic regression model significantly improved detection accuracy. This highlights the benefit of fusing information from different sensor modalities, even simple ones, to refine classification.
  • Individual Differences Matter: The research acknowledges notable individual differences in detection accuracy, particularly between male and female subjects. This suggests that sensor placement, muscle mass, and individual biomechanics play a role and may require personalized calibration strategies or sensor designs.

Key Results

  • Median Accuracy of 90%: The proposed static standing-based calibration method achieved a median accuracy of approximately 90% for gait phase detection during walking.
  • LR-a Outperforms Standard LR and SVM: The adjusted logistic regression (LR-a) algorithm consistently outperformed standard logistic regression and Support Vector Machine (SVM) in terms of accuracy, especially when considering the practical constraints of calibration.
  • Minimal Impact of Training Data Amount: The study found that the amount of training data (from 500 to 2000 samples) did not significantly affect the detection accuracy, suggesting that a relatively short calibration period (e.g., 60 seconds) is sufficient.
  • High Accuracy for Male Subjects: For six male subjects with more prominent muscle mass, the system achieved higher accuracy (around 94% for stance and 86% for swing phases) with an error of approximately 0.05 seconds.

Limitations and Future Work

  • Individual Differences: The paper acknowledges significant individual differences in detection accuracy, with female subjects generally showing lower accuracy. This is attributed to factors like muscle mass and sensor placement.
  • Future Work: Increasing the number of photo reflectors in the sensor could help capture muscle deformation more effectively across diverse individuals. Further research into understanding how slight variations in foot positioning during standing affect muscle deformation could also improve robustness.
  • Sensor Usability and Posture Arbitrariness: The current sensor's usability is somewhat dependent on individual muscle mass. The arbitrary nature of postures, even within the defined set, can also introduce variability.
  • Future Work: Developing more sophisticated methods to define and utilize postures for calibration is suggested.
  • Low Sample Frequency: The current system's low sample frequency limited detection accuracy.
  • Future Work: Constructing a system with shorter processing times and faster communication speeds is recommended to achieve higher detection accuracy.
  • Directional Extraction: The current model did not require directional extraction of movement because the anterior direction was constant. For real-world applications where people walk in any direction, this will be necessary.
  • Future Work: Incorporating directional extraction calculations for more general gait phase detection.
  • Fusion with Other Sensors: While the study focused on muscle deformation, the authors suggest that combining this with other sensing modalities (e.g., IMU-based rules) could further enhance accuracy.

Summary updated

BibTeX
@article{miyake2021gait,
  author = {Tamon Miyake and Shintaro Yamamoto and Satoshi Hosono and Satoshi Funabashi and Zhengxue Cheng and Cheng Zhang and Emi Tamaki and Shigeki Sugano},
  title = {{Gait Phase Detection Based on Muscle Deformation with Static Standing-Based Calibration}},
  journal = {Sensors},
  year = {2021},
  doi = {10.3390/s21041081},
}
Figure