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

Subash Padmanaban, Justin Baker, Bradley GregerFrontiers in Neuroscience 2018 · DOI

Abstract

Objective: The performance of machine learning algorithms used for neural decoding of dexterous tasks may be impeded by high-dimensional data. This study compares four feature selection techniques—Wilcoxon signed-rank test, Relative Importance, Principal Component Analysis (PCA), and Mutual Information Maximization (MIM)—on SVM classification performance for a dexterous decoding task in a non-human primate (NHP). Approach: An NHP was trained to perform coordinated finger movements while action potentials were recorded via a microelectrode array in the motor cortex. We evaluated robustness to simulated failure and longevity of classification using both isolated-neuron (single-unit) and multi-unit firing rates. Main results: MIM achieved the highest average decoding accuracy (97.65% for single-units). Notably, using only 10% of features selected via MIM maintained 90.79% accuracy for single-units, demonstrating significant robustness compared to other methods. Significance: Optimal feature selection improves decoding accuracy and robustness, which is critical for the clinical longevity of brain-machine interfaces (BMIs).

Introduction: The Challenge of High-Dimensional Neural Data

Brain-Machine Interfaces (BMIs) translate neural activity into control signals for external devices. While basic reaching tasks have been mastered, decoding dexterous finger movements remains challenging due to the "Curse of Dimensionality."

The Curse of Dimensionality: As the number of dimensions increases, the percentage of data points enclosed by a standard deviation region asymptotes to zero.

In high-dimensional spaces, data points become increasingly sparse. As shown in the figure above, the volume of the space increases so rapidly that the available data becomes insufficient for reliable statistical modeling. This often leads to "overfitting," where a model performs exceptionally well on training data but fails to generalize to new, real-world signals.

PDF of Univariate (A) and Bivariate (B) Gaussian Distributions showing data points enclosed by mean ± 1 SD.

As the dimensionality increases from 1D (A) to 2D (B) and beyond, the proportion of data points within the central region of the distribution decreases significantly. For BMIs, which must operate reliably over months or years despite physical electrode degradation, overcoming this sparsity through intelligent feature selection is essential.

Methods and Experimental Design

The study involved a male rhesus macaque trained to perform cued flexions of the thumb, index, and middle fingers. Neural activity was captured using a 96-channel microelectrode array implanted in the hand area of the motor cortex.

The researchers compared four distinct feature selection strategies to identify the most informative neural signals for a Support Vector Machine (SVM) classifier:

  1. Wilcoxon Signed-Rank Test: A non-parametric statistical test identifying electrodes with significant activity changes during movement.
  2. Relative Importance: Ranking features based on firing rate variance across different movement types.
  3. Principal Component Analysis (PCA): A linear transformation that captures the directions of maximum variance.
  4. Mutual Information Maximization (MIM): An information-theoretic approach that selects features sharing the most information with the target movement labels.

The analysis spanned 47 recording sessions over 142 days, comparing single-unit (isolated neurons) and multi-unit (aggregate electrode activity) features.

Results: The Superiority of Mutual Information Maximization

MIM emerged as the most effective method for maintaining high decoding accuracy. While PCA focuses on variance, MIM focuses on task-relevance, which proved crucial for dexterous decoding.

Peak Performance

The average decoding accuracy using MIM was 96.74 ± 3.5% for multi-unit features and 97.65 ± 3.36% for single-unit features. This high baseline indicates that the SVM could accurately distinguish between complex finger movements when provided with optimized inputs.

Robustness to Signal Loss

A critical test for any BMI is its ability to function when electrodes fail. The researchers simulated this by reducing the available features to just 10% of the original set.

  • Single-Units: Accuracy dropped only slightly, from 95.32% to 90.79% (a 4.75% reduction).
  • Multi-Units: Accuracy plummeted from 93.7% to 51.09% (a 45.56% reduction).

This suggests that single-unit data contains more distinct, robust information than multi-unit data, which may aggregate too much noise when the number of channels is limited.

Discussion and Clinical Significance

The study demonstrates that smart feature selection, particularly MIM, can significantly extend the functional life of a BMI. By identifying a small subset of highly informative neurons, the system remains robust even as the physical hardware degrades over time.

While single-unit features offer superior robustness, they require spike sorting, which adds computational complexity. However, the trade-off is often justified by the gain in longevity and accuracy. Even a small percentage increase in decoding performance can dramatically improve the user experience for individuals relying on BMIs for daily tasks. Future research will focus on generalizing these models across subjects and further optimizing real-time implementation.

Summary updated

Abstract

<b>Objective:</b> The performance of machine learning algorithms used for neural decoding of dexterous tasks may be impeded due to problems arising when dealing with high-dimensional data. The objective of feature selection algorithms is to choose a near-optimal subset of features from the original feature space to improve the performance of the decoding algorithm. The aim of our study was to compare the effects of four feature selection techniques, Wilcoxon signed-rank test, Relative Importance, Principal Component Analysis (PCA), and Mutual Information Maximization on SVM classification performance for a dexterous decoding task. <b>Approach:</b> A nonhuman primate (NHP) was trained to perform small coordinated movements-similar to typing. An array of microelectrodes was implanted in the hand area of the motor cortex of the NHP and used to record action potentials (AP) during finger movements. A Support Vector Machine (SVM) was used to classify which finger movement the NHP was making based upon AP firing rates. We used the SVM classification to examine the functional parameters of (i) robustness to simulated failure and (ii) longevity of classification. We also compared the effect of using isolated-neuron and multi-unit firing rates as the feature vector supplied to the SVM. <b>Main results:</b> The average decoding accuracy for multi-unit features and single-unit features using Mutual Information Maximization (MIM) across 47 sessions was 96.74 ± 3.5% and 97.65 ± 3.36% respectively. The reduction in decoding accuracy between using 100% of the features and 10% of features based on MIM was 45.56% (from 93.7 to 51.09%) and 4.75% (from 95.32 to 90.79%) for multi-unit and single-unit features respectively. MIM had best performance compared to other feature selection methods. <b>Significance:</b> These results suggest improved decoding performance can be achieved by using optimally selected features. The results based on clinically relevant performance metrics also suggest that the decoding algorithm can be made robust by using optimal features and feature selection algorithms. We believe that even a few percent increase in performance is important and improves the decoding accuracy of the machine learning algorithm potentially increasing the ease of use of a brain machine interface.
BibTeX
@inproceedings{padmanaban2018feature,
  author = {Subash Padmanaban and Justin Baker and Bradley Greger},
  title = {{Feature Selection Methods for Robust Decoding of Finger Movements in a Non-human Primate}},
  booktitle = {Frontiers in Neuroscience},
  year = {2018},
  doi = {10.3389/fnins.2018.00022},
}
Figure