Estimation of Ankle Joint Power during Walking Using Two Inertial Sensors

Xianta Jiang, Mohsen Gholami, Mahta Khoshnam, Janice J. Eng, Carlo MenonSensors 2019 · OpenReview

TL;DR

This paper introduces a novel machine learning approach to estimate ankle joint power during walking using just two inertial measurement units (IMUs) placed on the foot and shank. This method offers a more practical and portable alternative to traditional lab-based inverse dynamics techniques, paving the way for real-time gait monitoring in everyday settings.

What Problem Does This Paper Address?

Accurately measuring ankle joint power is crucial for diagnosing gait impairments and developing effective lower limb prosthetics. However, current methods, primarily based on inverse dynamics, rely on complex biomechanical analyses and highly instrumented laboratory environments (e.g., motion capture systems and force plates). This setup is cumbersome, expensive, and not suitable for real-world applications like continuous gait monitoring outside of a clinical setting. The paper aims to bridge this gap by developing a simpler, more accessible method for estimating ankle joint power.

Methods

The researchers employed a machine learning approach, specifically a Random Forest regressor, to estimate ankle joint power. The core idea is to train a model that can predict ankle joint power from the raw sensor data collected by two IMUs.

Here's a breakdown of their methodology:

  • Data Collection: Nine healthy adult male participants walked on a force-plate-instrumented treadmill at five different speeds. During these trials, two IMUs (one on the shank, one on the foot) recorded acceleration and angular velocity. Simultaneously, a Vicon motion capture system tracked marker kinematics, and the treadmill's force plates recorded ground reaction forces.
  • Reference Data Generation: Ankle joint power was calculated using the standard inverse dynamics approach with the motion capture and force plate data. This served as the "ground truth" for training and evaluating the ML model.
  • IMU Data Processing: The IMU data was first smoothed using a median filter. Then, a feature extraction process was applied using a 110-ms window, generating various signal features (e.g., RMS, MAV, variance). These features were normalized to a range between 0 and 1.
  • Machine Learning Model: A Random Forest regressor was chosen for its ability to handle complex relationships and provide robust predictions. The model was trained using the extracted IMU features to predict the reference ankle joint power.
  • Addressing Underestimation: Preliminary results showed the model tended to underestimate peak power. To address this, a quantile regression algorithm was incorporated. This allowed for adjusting the prediction based on specific quantiles, effectively compensating for the underestimation by selecting higher quantiles for peak power estimation.
  • Normalization: A specific normalization technique was used for the testing data, employing the maximum and minimum power values from the first half of each subject's data to normalize the second half.
  • Validation: The model's performance was evaluated using two approaches:
  • Intra-subject: Training and testing were performed on data from the same participant, ensuring the model was specific to individual movement patterns.
  • Inter-subject: The model was trained on data from eight participants and tested on the remaining one, and this process was repeated for all participants. This approach assesses the model's generalizability to new individuals, crucial for a portable device.

Datasets

The study utilized data collected from nine healthy adult male participants. This data comprised:

  • IMU data: Acceleration and angular velocity from two IMUs (foot and shank).
  • Motion capture data: Kinematics from a Vicon system.
  • Force plate data: Ground reaction forces from a split-belt treadmill.

This collected data was used to generate the reference ankle joint power values, which then served as the target for the machine learning model. The study did not introduce a new public benchmark dataset but rather used its own collected data for validation.

Key Insights and Ideas

  • Feasibility of Two IMUs: The most significant insight is the demonstration that ankle joint power can be estimated with high accuracy using only two IMUs. This drastically simplifies the required hardware compared to traditional multi-sensor setups.
  • Machine Learning for Biomechanics: The paper highlights the power of machine learning in translating raw sensor data into meaningful biomechanical parameters, moving beyond purely physics-based models.
  • Quantile Regression for Accuracy: The use of quantile regression to address the underestimation of peak power is a clever adaptation, showing how ML techniques can be fine-tuned for specific prediction challenges.
  • Portability and Real-World Application: The success of the inter-subject testing is a critical step towards developing a practical, wearable system for gait analysis outside of controlled lab environments.

Key Results

The proposed method demonstrated strong performance in estimating ankle joint power:

  • Intra-subject Accuracy:
    • Correlation Coefficient (R): 0.98
    • Root Mean Squared Error (RMSE): 0.06 w/kg
    • Normalized Root Mean Squared Error (NRMSE): 1.05%
  • Inter-subject Accuracy:
    • Correlation Coefficient (R): 0.92
    • Root Mean Squared Error (RMSE): 0.13 w/kg
    • Normalized Root Mean Squared Error (NRMSE): 2.37%
  • Peak Power Estimation:
    • Intra-subject: Predicted vs. true peak power difference of 0.01 w/kg (0.6% of range).
  • Inter-subject: Predicted vs. true peak power difference of 0.14 w/kg (6.8% of range).
  • Timing Accuracy: The occurrence of peak power was detected with a negligible delay of 0.4% of the gait cycle duration for both intra- and inter-subject testing.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Lab Setting: The study was conducted in a controlled lab environment. Future work will involve testing the system in outdoor settings and with more dynamic activities like running.
  • Participant Demographics: The study included only male participants within a limited range of height, weight, and age. This homogeneity might have boosted the model's prediction accuracy. Future research should include a more diverse group to ensure broader applicability.
  • Speed Range: The study focused on walking speeds. Extending the method to higher speeds and running scenarios is a logical next step.
  • Device Integration: While the method is promising for portable devices, further work is needed to integrate it into a user-friendly system for real-time monitoring and reporting.

Summary updated

Abstract

1) Background: Ankle joint power, as an indicator of the ability to control lower limbs, is of great relevance for clinical diagnosis of gait impairment and control of lower limb prosthesis. However, the majority of available techniques for estimating joint power are based on inverse dynamics methods, which require performing a biomechanical analysis of the foot and using a highly instrumented environment to tune the parameters of the resulting biomechanical model. Such techniques are not generally applicable to real-world scenarios in which gait monitoring outside of the clinical setting is desired. This paper proposes a viable alternative to such techniques by using machine learning algorithms to estimate ankle joint power from data collected by two miniature inertial measurement units (IMUs) on the foot and shank, (2) Methods: Nine participants walked on a force-plate-instrumented treadmill wearing two IMUs. The data from the IMUs were processed to train and test a random forest model to estimate ankle joint power. The performance of the model was then evaluated by comparing the estimated power values to the reference values provided by the motion tracking system and the force-plate-instrumented treadmill. (3) Results: The proposed method achieved a high accuracy with the correlation coefficient, root mean square error, and normalized root mean square error of 0.98, 0.06 w/kg, and 1.05% in the intra-subject test, and 0.92, 0.13 w/kg, and 2.37% in inter-subject test, respectively. The difference between the predicted and true peak power values was 0.01 w/kg and 0.14 w/kg with a delay of 0.4% and 0.4% of gait cycle duration for the intra- and inter-subject testing, respectively. (4) Conclusions: The results of this study demonstrate the feasibility of using only two IMUs to estimate ankle joint power. The proposed technique provides a basis for developing a portable and compact gait monitoring system that can potentially offer monitoring and reporting on ankle joint power in real-time during activities of daily living.
BibTeX
@inproceedings{jiang2019estimation,
  author = {Xianta Jiang and Mohsen Gholami and Mahta Khoshnam and Janice J. Eng and Carlo Menon},
  title = {{Estimation of Ankle Joint Power during Walking Using Two Inertial Sensors}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=3gWc6-zdRQ},
}
Figure