Decoding index finger position from EEG using random forests

Sebastian Weichwald, Timm Meyer, Bernhard Schölkopf, Tonio Ball, Moritz Grosse-Wentrup

Introduction

Non-invasive Brain-Computer Interfaces (BCIs) have traditionally struggled with granularity. While invasive methods like Electrocorticography (ECoG) can decode fine-grained motor details, electroencephalography (EEG) is often considered too "blurry" for complex tasks. Most EEG-based motor studies focus on binary or low-dimensional tasks, such as distinguishing between left and right hand movements.

This study investigates whether it is possible to decode the specific static position of a single finger—the index finger—on a 3x3 numerical keypad using only non-invasive EEG signals. This task is particularly challenging because the neural populations in the primary motor cortex controlling finger movements overlap significantly, and the mapping of muscle activations to joint positions does not follow a simple somatotopic organization.

Experimental Methods and Data

To address this challenge, researchers recorded 121-channel EEG data from 20 healthy subjects. The experimental paradigm required subjects to use their right index finger to press keys on a 3x3 numerical keyboard. To isolate the movement of the index finger, the subjects' thumbs were fixed in a specialized tube.

Experimental setup showing the hand on a 3x3 keypad with the thumb restrained.

The signal processing pipeline involved several robust steps:

  • Spatial Filtering: Laplacian spatial filtering was applied to the 106 common channels to improve the signal-to-noise ratio.
  • Feature Extraction: Log-bandpower was calculated for the $\mu$-band (8–15 Hz) and the $\beta$-band (20–30 Hz). Using 41 sliding windows (1 second each) across each trial, the team generated a high-dimensional feature space of 8,904 features per trial.
  • Classification: A Random Forest (RF) classifier was chosen for its ability to handle high-dimensional data without extensive hyperparameter tuning.
  • Validation: The study employed a "leave-one-subject-out" cross-validation approach. This rigorous method tests the model on a completely new individual, determining if neural patterns for finger position are consistent enough to generalize across different brains.

Key Insights and Results

The researchers found that the most predictive information was contained within the high $\beta$-band (20–30 Hz). Feature importance scores peaked over the C3 electrode, which corresponds to the contralateral (left-side) sensorimotor cortex controlling the right hand. Interestingly, the most informative period was the first second immediately after the finger reached the target, suggesting that the "beta-rebound" phenomenon—a burst of activity following movement completion—plays a significant role in decoding.

Confusion matrix showing the percentage of trials for target vs. predicted keys.

The model achieved an overall classification accuracy of 12.29%. While this may appear low, it is statistically highly significant ($p < 8 \cdot 10^{-10}$) given the 11.11% chance level for a 9-key task across over 26,000 trials. As shown in the confusion matrix above, performance varied by key; Target 5 (the center key) and Target 9 were the most accurately predicted, reaching 16.71% and 15.94% accuracy, respectively.

Discussion and Future Work

The success of the leave-one-subject-out method implies that human brains represent finger positions in a remarkably similar way. However, several questions remain. It is unclear whether the EEG is capturing the static position of the finger, the tonic muscle activity required to hold the key, or the sensory feedback from the fingertip.

While a 12.29% accuracy is a significant scientific proof-of-concept, it is not yet sufficient for a reliable real-world "mind-controlled" keyboard. Future work may improve these results through hyperparameter optimization of the Random Forest or by exploring whether non-linear models are strictly necessary compared to simpler linear alternatives. This research marks a meaningful step toward high-granularity, non-invasive motor decoding.

Summary updated

Abstract

While invasively recorded brain activity is known to provide detailed information on motor commands, it is an open question at what level of detail information about positions of body parts can be decoded from non-invasively acquired signals. In this work it is shown that index finger positions can be differentiated from non-invasive electroencephalographic (EEG) recordings in healthy human subjects. Using a leave-one-subject-out cross-validation procedure, a random forest distinguished different index finger positions on a numerical keyboard above chance-level accuracy. Among the different spectral features investigated, high β-power (20–30 Hz) over contralateral sensorimotor cortex carried most information about finger position. Thus, these findings indicate that finger position is in principle decodable from non-invasive features of brain activity that generalize across individuals.
BibTeX
@inproceedings{weichwald2014decoding,
  author = {Sebastian Weichwald and Timm Meyer and Bernhard Schölkopf and Tonio Ball and Moritz Grosse-Wentrup},
  title = {{Decoding index finger position from EEG using random forests}},
  booktitle = {CIP},
  year = {2014},
  doi = {10.1109/cip.2014.6844513},
}
Figure