Estimating Ground Reaction Forces from Gait Kinematics in Cerebral Palsy: A Convolutional Neural Network Approach

Area of Science:

• Biomechanics
• Neurology
• Machine Learning

Background:

• Gait analysis is crucial for diagnosing neuromotor disorders like cerebral palsy (CP).
• Accurate ground reaction force (GRF) measurement during natural walking is challenging due to gait variability.
• Existing machine learning studies lack specific focus on CP patients for GRF prediction.

Purpose of the Study:

• To predict GRF in CP patients using joint angles from marker-based motion capture.
• To develop a protocol for gait analysis in CP without relying on force plates.
• To address the gap in machine learning applications for CP-specific GRF prediction.

Main Methods:

• Utilized a dataset of 132 typically developed (TD) subjects and 622 CP patients.
• Collected lower limb kinematic data (joint angles) and GRF data (three axes).
• Employed a 1D convolutional neural network for feature extraction and GRF prediction.

Main Results:

• CP patient GRFs predicted with nRMSE < 20.13% and PCC > 0.84.
• TD subjects showed higher prediction accuracy: nRMSE < 12.65% and PCC > 0.94.
• Predicted GRF patterns closely matched experimental data.

Conclusions:

• Machine learning shows promise for predicting GRF in CP patients.
• The developed method offers a potential force plate-free gait analysis protocol.
• Current prediction accuracy limits immediate clinical application, requiring further model refinement.