Insole-Based Estimation of Vertical Ground Reaction Force Using One-Step Learning With Probabilistic Regression and...

Area of Science:

• Biomechanics
• Medical Engineering
• Machine Learning

Background:

• Insole-based vertical ground reaction force (vGRF) estimation offers a cost-effective alternative to force plates for pathological gait analysis.
• Current machine learning methods for vGRF estimation often require force plates and multiple walking steps, posing risks and physical strain for patients.

Purpose of the Study:

• To develop an accessible and efficient learning scheme for insole-based vGRF estimation.
• To overcome limitations of force plate dependency and the need for extensive data collection in pathological gait evaluation.

Main Methods:

• Utilized a low-cost scale as a substitute for expensive force plates.
• Employed Gaussian Process Regression (GPR) for robust vGRF estimation from small, noisy datasets.
• Introduced a "one-step learning" scheme with probabilistic data augmentation to simulate data from a single walking step.

Main Results:

• GPR models trained on just two walking steps achieved mean vGRF estimation errors of 8% or less.
• The "one-step learning" approach with probabilistic augmentation demonstrated enhanced estimation accuracy compared to traditional methods.

Conclusions:

• The proposed system provides an efficient and accessible method for insole-based vGRF estimation.
• The novel learning scheme effectively addresses overfitting and data scarcity issues in pathological gait analysis, enhancing accuracy and patient safety.