Decoding the activity of individual neural cells during natural behaviors is a cornerstone of neuroscience, allowing researchers to study how the nervous system generates and controls movement. Unlike many other neural cells, the activity of spinal alpha motor neurons can be determined non-invasively through the decomposition of electromyographic (EMG) signals. However, traditional decomposition is a slow, offline process.
"I-Spin live" is a new open-source software package that addresses this limitation by performing real-time decoding of spinal motor neurons using a blind-source separation (BSS) approach. By processing multichannel EMG signals from high-density electrode grids or intramuscular arrays, the software achieves sub-millisecond processing speeds, enabling advanced neurofeedback and human-machine interface applications.
The Challenge of Real-Time Decoding
When the brain initiates movement, signals travel down spinal alpha motor neurons to muscle fibers. Electromyography (EMG) records the resulting electrical activity, but the raw signal is a complex, overlapping mixture of hundreds of motor units firing simultaneously.
Isolating individual neurons requires mathematically "unmixing" these signals. While offline algorithms exist, real-time applications like neurofeedback—where a participant learns to control a specific neuron by seeing its activity—require immediate decoding. I-Spin live provides a validated, open-source framework for this task.
Methodology
The MATLAB-based pipeline supports high-density surface EMG grids (up to 64 channels) and intramuscular electrode arrays. The workflow consists of two primary phases:
- Offline Training: During a steady baseline contraction, the software uses a fast Independent Component Analysis (fastICA) algorithm to identify "separation vectors." These filters are unique to each motor unit based on its spatial and temporal signature across the electrode array.
- Real-Time Decoding: During active testing, the software processes EMG data in 125-millisecond chunks. It applies the pre-calculated filters and uses K-means clustering to distinguish true neural spikes from background noise.
The isolated discharge activity is then instantly visualized for the user as a scrolling raster plot or a smoothed firing rate curve.
Experimental Validation
The software was validated with human volunteers performing isometric contractions at various force levels (e.g., 20% and 40% of maximum voluntary contraction).
- Muscles: Gastrocnemius (calf), vastus (thigh), and tibialis anterior (shin).
- Hardware: 64-channel surface grids and 16-channel intramuscular arrays.
Key Insights
- Computational Efficiency: The math required to apply pre-calculated filters is lightweight. Algorithmic processing takes only ~0.37 ms per motor unit, meaning the primary delay is simply the time required to buffer the 125ms data epoch.
- Anatomical Factors: Accuracy varies by muscle. The tibialis anterior (shin) showed the highest performance, likely due to less complex activation patterns compared to the calf or thigh muscles.
- Stationarity Requirement: BSS relies on a stable physical relationship between electrodes and muscle fibers. Significant changes in contraction intensity or joint movement can shift this geometry, breaking the spatial filters.
Performance Results
- High Accuracy: The system achieved a 94% agreement rate for surface recordings of the shin and 97% for intramuscular recordings when compared to gold-standard manual offline editing.
- Precise Biofeedback: The real-time visual feedback closely mirrored ground truth, with a root-mean-square error (RMSE) of less than 2 pulses per second across all tests.
Limitations and Future Directions
While a significant step forward, the current version has limitations. The initial validation was conducted only on male participants, and the system is currently restricted to isometric contractions due to the stationarity constraint of BSS. Additionally, the software is currently optimized for specific commercial hardware (OT Bioelettronica Quattrocento) and requires an initial manual curation step by an expert. Future iterations aim to automate filter selection and adapt to dynamic movements.
Summary updated