Capturing spike train temporal pattern with wavelet average coefficient for brain machine interface

Shixian Wen, Allen Yin, Po-He Tseng, Laurent Itti, Mikhail Lebedev, Miguel A. L. NicolelisScientific 2021 · DOI

Abstract

Motor brain machine interfaces (BMIs) directly link the brain to artificial actuators and have the potential to mitigate severe body paralysis caused by neurological injury or disease. Most BMI systems involve a decoder that analyzes neural spike counts to infer movement intent. However, many classical BMI decoders (1) fail to take advantage of temporal patterns of spike trains, possibly over long time horizons; (2) are insufficient to achieve good BMI performance at high temporal resolution, as the underlying Gaussian assumption of decoders based on spike counts is violated. Here, we propose a new statistical feature that represents temporal patterns or temporal codes of spike events with richer description—wavelet average coefficients (WAC)—to be used as decoder input instead of spike counts. We constructed a wavelet decoder framework by using WAC features with a sliding-window approach, and compared the resulting decoder against classical decoders (Wiener and Kalman family) and new deep learning based decoders (Long Short-Term Memory) using spike count features. We found that the sliding-window approach boosts decoding temporal resolution, and using WAC features significantly improves decoding performance over using spike count features.

Introduction: The Limitations of Rate Coding

Most current Brain-Machine Interfaces (BMIs) rely on "rate coding," which involves counting the number of neural spikes within a specific time bin (e.g., 50ms or 100ms). While effective, this approach suffers from two major flaws. First, it results in a significant loss of information by ignoring the precise timing of spikes. A neuron firing three times at the beginning of a window is treated identically to one firing three times at the end, despite potentially different motor meanings. Second, there is a "resolution paradox": high-speed control requires small time bins (e.g., 5ms), but in such short windows, neural data becomes binary (0 or 1 spikes). This violates the Gaussian assumptions of classical decoders like Kalman filters, leading to a sharp decline in performance.

The Wavelet Framework and WAC Features

To address these issues, researchers developed a Wavelet Framework that captures rich temporal patterns over long horizons while maintaining high-speed output. The framework consists of four primary stages:

  1. Kernel Function: Instead of simple counts, spike trains are converted into a discrete waveform. As shown in the figure below, the signal fluctuates based on the distribution of spikes, effectively capturing the "rhythm" of neural activity.
  2. Discrete Wavelet Transform (DWT): This waveform is decomposed into multiple scales using wavelets. This allows the system to analyze both broad trends and fine details in the neural signal.
  3. Preprocessing (WAC): Wavelet coefficients are averaged to create Wavelet Average Coefficients (WAC). This compact feature set prevents overfitting while retaining temporal information.
  4. Sliding Window Decoder: A sliding window (e.g., 500ms) moves in 5ms increments, providing enough data to maintain Gaussian-like distributions for the decoder while ensuring high-frequency updates.
The Wavelet Framework pipeline, from spike trains to decoded kinematics.
The kernel function converts discrete spike trains into a continuous neural signal waveform.

Experimental Evaluation

The framework was tested using neural recordings from the primary motor cortex (M1) of four monkeys across two distinct tasks:

  • Center-out Reaching: Monkeys used a joystick to reach for targets.
  • Locomotion: Monkeys walked on a treadmill (forward and backward) while their ankle movements were tracked.

The dataset included recordings from 153 to 490 neurons per monkey. The researchers compared WAC features against standard spike counts using classical (Wiener, Kalman) and deep learning (LSTM) decoders.

Key Insights and Results

The study revealed several critical findings:

  • Temporal Coding Superiority: WAC features significantly outperformed standard spike counts across all tasks. This confirms that the rhythm and "quiet periods" of spike trains contain vital information for motor control.
  • Resolution and Physics: The sliding window approach successfully fixed the Gaussian violation, allowing classical filters to operate at high temporal resolutions. Furthermore, the optimal window size correlated with movement physics: position (slow) favored longer windows (~500ms), while velocity (fast) benefited from shorter ones (~350ms).
  • Deep Learning Synergy: Even the LSTM, which is inherently designed for sequences, showed improved performance when fed WAC features compared to raw spike counts at high resolutions.

Conclusion and Future Directions

While the Wavelet Framework offers a significant leap in BMI accuracy, challenges remain. Performance is sensitive to hyperparameters like window size and wavelet levels, requiring manual tuning. Additionally, the computational complexity of real-time DWT calculation must be addressed for clinical implants. Future work will focus on "closed-loop" testing to validate if these offline improvements translate to better real-time control for human patients.

Summary updated

Abstract

Motor brain machine interfaces (BMIs) directly link the brain to artificial actuators and have the potential to mitigate severe body paralysis caused by neurological injury or disease. Most BMI systems involve a decoder that analyzes neural spike counts to infer movement intent. However, many classical BMI decoders (1) fail to take advantage of temporal patterns of spike trains, possibly over long time horizons; (2) are insufficient to achieve good BMI performance at high temporal resolution, as the underlying Gaussian assumption of decoders based on spike counts is violated. Here, we propose a new statistical feature that represents temporal patterns or temporal codes of spike events with richer description-wavelet average coefficients (WAC)-to be used as decoder input instead of spike counts. We constructed a wavelet decoder framework by using WAC features with a sliding-window approach, and compared the resulting decoder against classical decoders (Wiener and Kalman family) and new deep learning based decoders ( Long Short-Term Memory) using spike count features. We found that the sliding-window approach boosts decoding temporal resolution, and using WAC features significantly improves decoding performance over using spike count features.
BibTeX
@inproceedings{wen2021capturing,
  author = {Shixian Wen and Allen Yin and Po-He Tseng and Laurent Itti and Mikhail Lebedev and Miguel A. L. Nicolelis},
  title = {{Capturing spike train temporal pattern with wavelet average coefficient for brain machine interface}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-98578-5},
}
Figure