Feature Selection Methods for Robust Decoding of Finger Movements in a Non-human Primate.

Area of Science:

• Neuroscience and Machine Learning
• Brain-Computer Interface (BCI) Development

Background:

• High-dimensional data challenges machine learning performance in neural decoding.
• Feature selection aims to optimize decoding algorithms by identifying crucial features.

Purpose of the Study:

• To compare four feature selection techniques: Wilcoxon signed-rank test, Relative Importance, Principal Component Analysis (PCA), and Mutual Information Maximization (MIM).
• To evaluate their impact on Support Vector Machine (SVM) classification performance for a dexterous neural decoding task.

Main Methods:

• Recorded action potentials (AP) from a nonhuman primate's motor cortex during coordinated finger movements.
• Utilized SVM to classify finger movements based on AP firing rates.
• Compared isolated-neuron and multi-unit firing rates as feature vectors.

Main Results:

• Mutual Information Maximization (MIM) demonstrated superior performance among the tested feature selection methods.
• Average decoding accuracy with MIM reached 97.65% for single-unit features and 96.74% for multi-unit features.
• Feature reduction using MIM showed minimal accuracy loss (4.75% for single-unit, 45.56% for multi-unit) when using 10% of features.

Conclusions:

• Optimally selected features significantly enhance neural decoding performance.
• Feature selection algorithms can improve the robustness and longevity of decoding algorithms.
• Even minor performance improvements are critical for advancing brain-machine interface usability.