Dual-Branch Deep Learning for Continuous Gait Cycle Estimation with wearable IMU Sensors and Anthropometric Data

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Gait cycle phase prediction is crucial for clinical assessments and assistive device control.
• Wearable inertial measurement units (IMUs) offer a promising avenue for unobtrusive gait monitoring.
• Accurate real-time gait analysis requires robust models that can process complex sensor data.

Purpose of the Study:

• To develop and validate a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model for predicting gait cycle phases.
• To investigate the utility of wearable IMU data and anthropometric features for real-time gait estimation.
• To identify key kinematic features that improve the performance of gait prediction models.

Main Methods:

• A CNN-LSTM model was employed for gait cycle phase prediction.
• A sliding window approach was utilized for real-time gait estimation.
• Gait cycle percentage was calculated using peak detection and lower limb length data.
• Feature selection focused on IMU-derived kinematic features, particularly joint angles.

Main Results:

• The CNN-LSTM model achieved a high accuracy, with an R-squared value of 0.911.
• The model demonstrated robust performance across various walking conditions.
• Feature selection successfully identified key kinematic features that enhanced model performance.

Conclusions:

• The developed CNN-LSTM model provides accurate real-time gait cycle phase prediction using wearable IMU data.
• This approach supports applications in personalized rehabilitation, early detection of abnormal gaits, and prosthetic control.
• The model's reliance on kinematic data makes it suitable for both clinical and home-based gait assessments.