Abstract
Recent advances in functional magnetic resonance imaging (fMRI) have been used to reconstruct cognitive states based on brain activity evoked by sensory or cognitive stimuli. To date, such decoding paradigms were mostly used for visual modalities. On the other hand, reconstructing functional brain activity in motor areas was primarily achieved through more invasive electrophysiological techniques. Here, we investigated whether non-invasive fMRI responses from human motor cortex can also be used to predict individual arm movements. To this end, we conducted fMRI studies in which participants moved their arm from a center position to one of eight target directions. Our results suggest that arm movement directions can be distinguished from the multivoxel patterns of fMRI responses in motor cortex. Furthermore, compared to multivoxel pattern analysis, encoding models were able to also reconstruct unknown movement directions from the predicted brain activity. We conclude for our study that non-invasive fMRI signal can be utilized to predict directional motor movements in human motor cortex.
Introduction
Most high-performance Brain-Computer Interfaces (BCIs) that control robotic arms rely on invasive electrophysiology—electrodes implanted directly into the brain. While precise, these carry surgical risks and potential for inflammation. Non-invasive alternatives like fMRI have been widely used to reconstruct visual images or "read" thoughts, but applying them to the motor cortex is challenging due to spatial resolution limits and motion artifacts.
This paper investigates whether a combination of multivoxel pattern analysis (MVPA) and voxel-based encoding models can overcome these hurdles to reconstruct arm movements from the primary motor cortex (M1). The study specifically addresses whether the "pooled" signal from thousands of neurons in a single voxel can still reveal specific directional tuning.
Methods
The researchers recruited eight healthy, right-handed subjects for a "center-out" reaching task. Participants moved their right arm toward one of eight target directions (spaced 45° apart) while inside a 3T MRI scanner.
The study compared two main computational approaches:
- Multivoxel Pattern Analysis (MVPA): Using a Linear Support Vector Machine (SVM) to classify brain activity into one of the eight discrete target directions.
- Encoding Model: Modeling each voxel’s response as a linear combination of six basis functions—half-wave rectified sinusoids representing idealized directional tuning curves. This allows for continuous reconstruction on a 0–360° scale.
- Voxel Selection: Sparse Multinomial Logistic Regression (SMLR) was used to identify informative voxels in the left M1, preventing the model from overfitting to noise.
- Control Analysis: To ensure the model wasn't simply decoding head motion, the same analysis was performed on the ipsilateral (right) M1, which is not primarily involved in right-arm movement.
Key Results and Insights
- Classification Accuracy: The MVPA (SVM) achieved an average accuracy of 41.8%, and the combined Encoding+SVM model reached 42.8%, both significantly above the 12.5% chance level.
- Reconstruction of Unknown Directions: By modeling the underlying tuning of voxels, the researchers could predict movement directions even for targets the model had not seen during training. Reconstructed patterns for these "unknown" directions were highly correlated (ρ = 0.92) with ground truth models.
- Precision through Averaging: While single-trial fMRI is noisy, averaging trials for the same direction significantly reduced the Mean Absolute Error (MAE), dropping it from ~50° to 26.6° for known directions and 31.6° for unknown directions.
- Validating Neural Origin: The control test on the right M1 (ipsilateral) resulted in near-chance performance. This proved that the signal in the left M1 was truly related to neural motor processing rather than global head motion artifacts.
Discussion and Future Work
While this study demonstrates the feasibility of non-invasive motor decoding, several limitations remain. The temporal lag of the hemodynamic response (fMRI) makes it unsuitable for real-time "twitch-response" BCI control in its current form. Furthermore, the current encoding model only considers direction, whereas real-world movements involve velocity, force, and joint position.
Future work may explore incorporating these additional features or utilizing motor imagery (imagining the movement) to bypass physical motion artifacts entirely while still leveraging the motor cortex's directional tuning properties.
Summary updated