A permutation importance and ensemble learning based feature selection approach for muscular intent decoding.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Muscle signals exhibit inherent indeterminism and significant inter-subject variability, posing challenges for accurate movement identification.
• Existing methods may struggle with the complexity and variability of electromyography (EMG) data.
• Developing robust and efficient methods for analyzing muscle signals is crucial for applications like prosthetics and human-computer interfaces.

Purpose of the Study:

• To propose and evaluate a subject-specific feature selection approach for accurate hand movement identification from muscle signals.
• To assess the impact of feature reduction on classification accuracy, F1 score, and computational efficiency.
• To determine the optimal subset of features required for reliable movement prediction.

Main Methods:

• A subject-specific feature selection strategy was developed using permutation importance-based weight calculation.
• An ensemble-based classifier was employed to identify different hand movements.
• Performance was rigorously evaluated using metrics including accuracy, F1 score, and computational time.

Main Results:

• The study found that utilizing only 25% of the original features is sufficient for accurate hand movement prediction.
• A notable accuracy and F1 score increment of approximately 3-5% was observed with the reduced feature set.
• Feature reduction led to a significant decrease in training and validation time, by nearly 40%.

Conclusions:

• A subject-specific feature selection approach effectively addresses the indeterminism and inter-subject variations in muscle signals.
• Significant feature reduction is achievable without compromising, and even improving, classification performance.
• The proposed method offers a computationally efficient solution for real-time hand movement identification from EMG data.