Turn Intent Detection For Control of a Lower Limb Prosthesis.

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Machine Learning in Prosthetics

Background:

• Adaptable lower limb prostheses require real-time control for variable stiffness, particularly during turning.
• Predicting amputee turning intent is crucial for timely prosthetic adjustments during the swing phase.

Purpose of the Study:

• To identify classification algorithms capable of accurately predicting turning using inertial measurement unit (IMU) signals.
• To determine if these algorithms provide sufficient time for prosthetic stiffness adjustment during the swing phase.

Main Methods:

• Developed classification models (SVM, KNN, Ensemble) using simulated IMU data from the prosthesis shank.
• Trained models on activities including 90° spin turns, 90° step turns, 180° turns, and straight walking.
• Evaluated classifier performance for predicting imminent turns and transition steps.

Main Results:

• Individualized training yielded better results than pooled user data.
• SVM, KNN, and Ensemble classifiers achieved high accuracy (96%, 93%, 91%) in predicting turns 400 ± 70 ms before heel strike.
• Classification accuracy for straight walking transition steps varied (85%, 82%, 97%), with the Ensemble model showing significant differences.

Conclusions:

• The Ensemble model demonstrated the best overall performance.
• SVM and KNN remain viable options depending on specific application priorities (turn vs. transition detection) and computational constraints.
• Findings offer a classifier strategy for lower limb devices aiming to predict turning intent.