Identification of Gait Events in Healthy Subjects and With Parkinson's Disease Using Inertial Sensors: An Adaptive...

Area of Science:

• Robotics and Biomechanics
• Machine Learning in Healthcare
• Wearable Technology

Background:

• Accurate gait event identification is crucial for assistive robotic devices.
• Existing methods often struggle with real-time application and gait impairments.
• Machine learning offers personalized gait pattern recognition but typically requires extensive training.

Purpose of the Study:

• To develop unsupervised, adaptive algorithms for real-time gait event detection.
• To eliminate the need for subject-specific training stages.
• To enable gait event detection using minimal sensor data.

Main Methods:

• Developed two adaptive unsupervised algorithms for detecting four key gait events.
• Utilized data from two single-Inertial Measurement Unit (IMU) foot-mounted wearable devices.
• Evaluated algorithms on healthy adults and individuals with Parkinson's disease (PD) during overground and treadmill walking.

Main Results:

• Achieved high accuracy (F1-score ≥ 0.95) for both healthy and PD groups.
• Demonstrated strong timing agreement with force-sensitive resistors (mean absolute differences of 66 ± 53 msec for healthy, 58 ± 63 msec for PD).
• Algorithms adapted parameters within walking trials, showing potential for personalized optimization.

Conclusions:

• The proposed adaptive unsupervised algorithms effectively detect gait events in real-time.
• These methods reduce reliance on external sensors, labeling, and lengthy training phases.
• The technology holds promise for enhancing control of assistive robotic devices for diverse populations.