ANOVA and linear regression feature selection for GRU-based foot position prediction in powered prostheses

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Powered prostheses aim to restore natural gait for amputees.
• Accurate foot position prediction is crucial for intuitive prosthesis control.
• Optimizing feature selection can enhance model efficiency and performance.

Purpose of the Study:

• To evaluate ANOVA and Linear Regression for feature selection in GRU models for powered prosthesis control.
• To compare these methods against Recursive Feature Elimination for efficiency and accuracy.
• To assess the generalization capability of the optimized models across varied terrains.

Main Methods:

• Kinematic data were collected from healthy participants during walking, stair negotiation, and standing.
• ANOVA and Linear Regression were employed for feature selection.
• Gated Recurrent Unit (GRU) networks were trained using selected features.
• Models were tested on independent subjects to evaluate generalization.

Main Results:

• ANOVA and Linear Regression efficiently selected relevant features, reducing computational load.
• The optimized GRU models achieved comparable performance to Recursive Feature Elimination.
• Root Mean Square Error (RMSE) as low as 0.066 radians demonstrated robust foot position prediction.
• The models showed strong generalization across different activities and terrains.

Conclusions:

• Feature selection using ANOVA and Linear Regression is an effective strategy for optimizing GRU-based powered prosthesis control.
• This method offers computational efficiency without compromising prediction accuracy.
• Further clinical validation with amputee subjects is required to confirm real-world applicability.