Building Effective Machine Learning Models for Ankle Joint Power Estimation During Walking Using FMG Sensors
Oliver Heeb1, Arnab Barua2, Carlo Menon1,3
1Biomedical and Mobile Health Technology Laboratory, ETH Zurich, Zurich, Switzerland.
Frontiers in Neurorobotics
|April 18, 2022
Summary
Machine learning models effectively estimate ankle joint power using force myography (FMG) sensors, eliminating the need for complex equipment. This study achieved a high correlation coefficient (R=0.91) for predicting ankle power during walking.
Area of Science:
- Biomechanics
- Machine Learning
- Wearable Sensors
Background:
- Traditional ankle joint power measurement relies on complex equipment and biomechanical models.
- Force myography (FMG) offers a potential alternative for non-invasive power estimation.
Purpose of the Study:
- To develop and validate machine learning (ML) and deep learning (DL) models for estimating ankle joint power using FMG signals.
- To compare the performance of different ML models and feature sets for this estimation task.
Main Methods:
- FMG signals were collected from nine participants walking at five different velocities using an 8-sensor FMG strap.
- Raw, time-domain, and frequency-domain features were extracted from FMG signals.
- Cat Boost Regressor, LSTM, and CNN models were trained and tested on the extracted features.
Main Results:
- The study achieved a correlation coefficient of R = 0.91 ± 0.07 for intrasubject ankle joint power prediction.
- Convolutional Neural Network (CNN) on raw FMG data and Long-Short Term Memory (LSTM) on time-domain features showed superior performance.
- A performance difference was noted between slowest and fastest walking speeds.
Conclusions:
- ML and DL models, particularly CNN and LSTM, can effectively estimate ankle joint power from FMG signals.
- The choice of feature set and ML model significantly impacts prediction accuracy.
- FMG sensors provide a viable, less complex alternative for ankle joint power assessment.
