Automatic real-time gait event detection in children using deep neural networks.

Area of Science:

• Biomechanics
• Computational Neuroscience
• Pediatric Medicine

Background:

• Accurate annotation of foot-contact and foot-off events is crucial for quantitative gait analysis.
• Manual annotation is time-consuming and subjective, while existing automatic methods struggle with pathological gait and assistive devices.
• There is a need for reliable, data-driven automatic tools for gait event detection in diverse pediatric populations.

Purpose of the Study:

• To develop and validate a data-driven model for accurate prediction of foot-contact and foot-off events in children.
• To compare the performance of the proposed model against heuristic-based approaches.
• To enable real-time applications for gait analysis and assistive device control.

Main Methods:

• Utilized a dataset of 9092 gait cycle measurements from children with normal and pathological gait.
• Employed Long Short-Term Memory (LSTM) artificial neural networks for time-series prediction of gait events.
• Trained and evaluated the predictive model using kinematic and marker data.

Main Results:

• The best-performing LSTM model achieved average errors of 10 ms for foot-contact and 13 ms for foot-off events.
• The data-driven approach significantly outperformed popular heuristic-based methods.
• The model demonstrated sufficient accuracy for clinical and research applications in pediatric gait analysis.

Conclusions:

• The developed LSTM-based model provides accurate and reliable automatic annotation of gait events in children.
• The approach generalizes well to both normal and pathological gait patterns, including those with assistive devices.
• The real-time prediction capability opens avenues for advanced control of active assistive technologies.