Deep Convolutional and LSTM Networks on Multi-Channel Time Series Data for Gait Phase Recognition.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Ageing populations experience a rise in gait disorders, significantly impacting mobility and quality of life.
• Falls associated with gait disorders increase morbidity and mortality, highlighting the need for early diagnosis.
• Current gait analysis systems are often expensive, limiting patient access to advanced diagnostics and therapies.

Purpose of the Study:

• To develop a reliable, low-cost system for detecting human gait phases using inertial measurement units (IMUs).
• To establish the foundation for a new medical device for gait analysis.
• To leverage machine learning for automatic, real-time motion analysis in gait diagnostics.

Main Methods:

• Utilized inertial measurement units (IMUs) for motion data acquisition.
• Developed a machine learning model combining deep 2D-convolutional and Long Short-Term Memory (LSTM) networks.
• Trained and evaluated the model on its ability to classify different human gait phases.

Main Results:

• Achieved over 92% accuracy in predicting gait phases on unseen subjects.
• Successfully differentiated between five distinct gait phases.
• Explored and evaluated various optimization strategies to enhance model performance.

Conclusions:

• The proposed IMU-based, machine learning approach offers a promising solution for accurate gait phase detection.
• This method can form the basis of a new, accessible medical device for gait analysis.
• Early and systematic diagnosis of gait disorders can be improved, potentially reducing suffering and healthcare costs.