State-Dependent Decoding Algorithms Improve the Performance of a Bidirectional BMI in Anesthetized Rats

Vito De Feo, Fabio Boi, Houman Safaai, Arno Onken, Stefano Panzeri, Alessandro VatoFrontiers in Neuroscience 2017 · DOI

Abstract

Brain-machine interfaces (BMIs) promise to improve the quality of life of patients suffering from sensory and motor disabilities by creating a direct communication channel between the brain and the external world. Yet, their performance is currently limited by the relatively small amount of information that can be decoded from neural activity recorded form the brain. We have recently proposed that such decoding performance may be improved when using state-dependent decoding algorithms that predict and discount the large component of the trial-to-trial variability of neural activity which is due to the dependence of neural responses on the network's current internal state. Here we tested this idea by using a bidirectional BMI to investigate the gain in performance arising from using a state-dependent decoding algorithm. This BMI, implemented in anesthetized rats, controlled the movement of a dynamical system using neural activity decoded from motor cortex and fed back to the brain the dynamical system's position by electrically microstimulating somatosensory cortex. We found that using state-dependent algorithms that tracked the dynamics of ongoing activity led to an increase in the amount of information extracted form neural activity by 22%, with a consequently increase in all of the indices measuring the BMI's performance in controlling the dynamical system. This suggests that state-dependent decoding algorithms may be used to enhance BMIs at moderate computational cost.

Introduction: The Challenge of Trial-to-Trial Variability

A major hurdle in neurotechnology is "trial-to-trial variability." If the brain is provided with the exact same sensory input twice, the resulting neural response will look different each time. Traditionally, BMI decoders treat this variation as random noise to be averaged away. However, much of this "noise" is actually the result of the brain’s current internal state—ongoing fluctuations in excitability, attention, or spontaneous activity.

This research investigates whether integrating the pre-stimulus "state" of the brain can improve the real-world performance of a closed-loop BMI. By accounting for the network's internal state just before a command is issued, decoders can better interpret the resulting signal, effectively "cleaning up" the post-stimulus response.

Methods: A Bidirectional BMI Framework

The researchers implemented a bidirectional BMI in anesthetized rats, creating a closed loop between the brain and a simulated dynamical system (a point mass moving in a viscous medium). The system utilized two primary interfaces:

  1. The Encoder (Input): The system communicated the position of the point mass to the brain via intracortical microstimulation (ICMS) delivered to the somatosensory cortex (S1).
  2. The Decoder (Output): Neural activity was recorded from the motor cortex (M1) to determine the intended direction of movement for the point mass.
Schematic of the bidirectional BMI loop, showing the encoder (S1 stimulation) and decoder (M1 recording) interface with a dynamical system.

The study utilized data from 5 male Long-Evans rats across 13 experimental sessions. Performance was evaluated using a leave-one-out cross-validation approach on a simulated 36 x 36 cm horizontal plane where the BMI drove the point mass toward a central target.

Decoding the Brain's Internal State

The core innovation of this study was the comparison of three decoding strategies:

  • State-Independent (SI): The standard approach, analyzing only the neural response after the stimulus.
  • State-Dependent Time-Averaged (SD-TA): Considers the post-stimulus response plus the average firing rate of the brain just before the stimulus.
  • State-Dependent Time-Dependent (SD-TD): Considers the post-stimulus response plus the full temporal history of the pre-stimulus activity.
Flowchart of the state-dependent decoding algorithm, illustrating the transition from neural activity to motor commands via dimensionality reduction and state-aware classification.

By using Principal Component Analysis (PCA) for dimensionality reduction and Linear Discriminant Analysis (LDA) for classification, the researchers found that the temporal structure of pre-stimulus activity (SD-TD) was critical. The state itself did not carry information about the target, but knowing the state allowed the decoder to filter the subsequent response more accurately.

Key Results and Performance Gains

The state-dependent approach (SD-TD) showed clear dominance across all performance metrics:

  • Information Extraction: A 21.74% increase in extracted information compared to the standard SI decoder.
  • Accuracy: A reduction in angular error (~7.5%) when translating neural signals into force vectors.
  • Control Efficiency: The convergence rate (reaching the target) increased by 10.49%, and the system required 16.38% fewer steps to reach the goal.
  • Stability: Within-trajectory variance decreased by 26.24%, resulting in significantly smoother control paths.

Discussion and Future Directions

While these results demonstrate a significant performance boost at a moderate computational cost, several challenges remain. The study was conducted under anesthesia; an awake brain presents far more complex, shifting states of attention and multi-tasking. Additionally, the SD-TD model requires observing pre-stimulus activity for several hundred milliseconds, which could create a bottleneck in high-speed real-time applications.

Future work will focus on validating these algorithms in behaving subjects and integrating these mathematical models into programmable hardware or neuromorphic chips. This would allow for low-power, "on-the-fly" state decoding in portable medical devices, bringing more intuitive and stable prosthetic control closer to reality.

Summary updated

Abstract

Brain-machine interfaces (BMIs) promise to improve the quality of life of patients suffering from sensory and motor disabilities by creating a direct communication channel between the brain and the external world. Yet, their performance is currently limited by the relatively small amount of information that can be decoded from neural activity recorded form the brain. We have recently proposed that such decoding performance may be improved when using state-dependent decoding algorithms that predict and discount the large component of the trial-to-trial variability of neural activity which is due to the dependence of neural responses on the network's current internal state. Here we tested this idea by using a bidirectional BMI to investigate the gain in performance arising from using a state-dependent decoding algorithm. This BMI, implemented in anesthetized rats, controlled the movement of a dynamical system using neural activity decoded from motor cortex and fed back to the brain the dynamical system's position by electrically microstimulating somatosensory cortex. We found that using state-dependent algorithms that tracked the dynamics of ongoing activity led to an increase in the amount of information extracted form neural activity by 22%, with a consequently increase in all of the indices measuring the BMI's performance in controlling the dynamical system. This suggests that state-dependent decoding algorithms may be used to enhance BMIs at moderate computational cost.
BibTeX
@inproceedings{feo2017state,
  author = {Vito De Feo and Fabio Boi and Houman Safaai and Arno Onken and Stefano Panzeri and Alessandro Vato},
  title = {{State-Dependent Decoding Algorithms Improve the Performance of a Bidirectional BMI in Anesthetized Rats}},
  booktitle = {Frontiers in Neuroscience},
  year = {2017},
  doi = {10.3389/fnins.2017.00269},
}
Figure