Gait episode identification based on wavelet feature clustering of spectrogram images

Area of Science:

• Biomechanics
• Signal Processing
• Machine Learning

Background:

• Gait parameter measurement is crucial for assessing health and safety.
• Automatic gait analysis using kinematic sensors is an emerging research field.

Purpose of the Study:

• To develop and evaluate a novel approach for detecting gait episodes.
• To utilize a neural network and wavelet-based image clustering for gait event detection.

Main Methods:

• Processing signals from a chest-worn inertial measurement unit (IMU) using Explicit Complementary Filter (ECF) to track torso angle.
• Applying wavelet decomposition to spectrogram images of sensor data.
• Classifying walking episodes using an Augmented Radial Basis Neural Network (ARBF).
• Optimizing ARBF cluster centroids with Rapid Cluster Estimation (RCE).

Main Results:

• The proposed method achieved up to 85.71% sensitivity in distinguishing walking from non-walking activities.
• The approach demonstrated up to 91.34% specificity in classifying gait events.
• A pilot study with 11 participants validated the method's effectiveness.

Conclusions:

• The developed approach shows promise for accurate automatic gait episode detection.
• The combination of wavelet decomposition, ARBF, and RCE offers a robust method for gait analysis.
• This technology has potential applications in health monitoring and safety systems.