Evaluating Machine Learning-Based Classification of Human Locomotor Activities for Exoskeleton Control Using Inertial...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Robotics

Background:

• Accurate classification of human locomotor activities is crucial for controlling wearable robotic exoskeletons.
• Wearable sensors, including inertial measurement units (IMUs) and e-textile pressure sensors, offer a non-invasive method for data collection.
• Machine learning models can analyze complex sensor data to infer activity type, speed, and incline.

Purpose of the Study:

• To evaluate three machine learning models (Logistic Regression, Random Forest, Light Gradient-Boosting Machine) for classifying human locomotor activities.
• To assess the models' ability to differentiate between walking, running, and jumping, as well as predict speed and surface incline.
• To determine the optimal sensor placement and data types for accurate activity prediction.

Main Methods:

• Collected data from 16 healthy participants using lower limb and pelvic IMUs and insole pressure sensors during treadmill locomotion.
• Trained and validated Logistic Regression, Random Forest, and Light Gradient-Boosting Machine models on sensor data.
• Evaluated model performance on classifying activity type, speed, and incline using a separate test dataset.

Main Results:

• The Light Gradient-Boosting Machine (LGBM) model demonstrated superior performance in classifying activity type and predicting speed compared to Logistic Regression and Random Forest.
• LGBM showed robustness with reduced IMU sensor input and identified gyroscope data as most critical for performance.
• Incline classification accuracy was not significantly improved by the evaluated models, and speed prediction was challenged by transitional gait patterns.

Conclusions:

• The LGBM model shows significant potential for real-time locomotor activity prediction using data from lower-limb-mounted IMUs.
• This technology can enhance the control capabilities of wearable robotic exoskeletons by accurately identifying user intentions.
• Further research is needed to improve incline prediction and address challenges in distinguishing between fast walking and slow running.