Limits...
The role of ECoG magnitude and phase in decoding position, velocity, and acceleration during continuous motor behavior.

Hammer J, Fischer J, Ruescher J, Schulze-Bonhage A, Aertsen A, Ball T - Front Neurosci (2013)

Bottom Line: The decoding accuracy based only on phase information was substantially (at least 2 fold) and significantly higher than that based only on magnitudes for position, velocity and acceleration.There was also no indication of additional magnitude-based mechanisms encoding movement information in the LFC range.Thus, our study contributes to elucidating the nature of the informative LFC of motor cortical population activity and may hence contribute to improve decoding strategies and BMI performance.

View Article: PubMed Central - PubMed

Affiliation: Bernstein Center Freiburg, University of Freiburg Freiburg, Germany.

ABSTRACT
In neuronal population signals, including the electroencephalogram (EEG) and electrocorticogram (ECoG), the low-frequency component (LFC) is particularly informative about motor behavior and can be used for decoding movement parameters for brain-machine interface (BMI) applications. An idea previously expressed, but as of yet not quantitatively tested, is that it is the LFC phase that is the main source of decodable information. To test this issue, we analyzed human ECoG recorded during a game-like, one-dimensional, continuous motor task with a novel decoding method suitable for unfolding magnitude and phase explicitly into a complex-valued, time-frequency signal representation, enabling quantification of the decodable information within the temporal, spatial and frequency domains and allowing disambiguation of the phase contribution from that of the spectral magnitude. The decoding accuracy based only on phase information was substantially (at least 2 fold) and significantly higher than that based only on magnitudes for position, velocity and acceleration. The frequency profile of movement-related information in the ECoG data matched well with the frequency profile expected when assuming a close time-domain correlate of movement velocity in the ECoG, e.g., a (noisy) "copy" of hand velocity. No such match was observed with the frequency profiles expected when assuming a copy of either hand position or acceleration. There was also no indication of additional magnitude-based mechanisms encoding movement information in the LFC range. Thus, our study contributes to elucidating the nature of the informative LFC of motor cortical population activity and may hence contribute to improve decoding strategies and BMI performance.

No MeSH data available.


Related in: MedlinePlus

Time-frequency resolved decoding from “kinematic copy + white noise” models and from real ECoG. Correlation coefficients between predicted and actual frequency components (CCFC) computed at different carrier frequencies f and time offsets τ. The kinematic copy (std = 1) was summed up with white noise (std = 20). (A) “position + noise” model predicting position, (B) “velocity + noise” predicting velocity and (C) “acceleration + noise” model predicting acceleration. (D) The CCFC for real ECoG data predicting trajectory derivatives, i.e., position, velocity, and acceleration (averaged over all subjects and also all trajectory derivatives, see section Frequency Resolved Decoding of Position, Velocity, and Acceleration). (E) “position + noise” model, CCFC frequency profile (purple curve, left y-axis) for the time-frequency magnitude and phase features at time offset τ = 0 s and with higher frequency resolution (window size = 5 s) plotted against the power spectral density of the time course of position along the trajectory (PSD, cyan, right y-axis). (F) Same as (E) for “velocity + noise” model (green curve, left y-axis). (G) Same as (E) for “acceleration + noise” model (brown curve, left y-axis). (H) The CCFC obtained from decoding the real ECoG data (red) plotted on top of the CCFC of the 3 models in (E–G). There is only a close match between the ECoG CCFC and the frequency profile resulting from the “velocity + noise” model (green).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3814578&req=5

Figure 9: Time-frequency resolved decoding from “kinematic copy + white noise” models and from real ECoG. Correlation coefficients between predicted and actual frequency components (CCFC) computed at different carrier frequencies f and time offsets τ. The kinematic copy (std = 1) was summed up with white noise (std = 20). (A) “position + noise” model predicting position, (B) “velocity + noise” predicting velocity and (C) “acceleration + noise” model predicting acceleration. (D) The CCFC for real ECoG data predicting trajectory derivatives, i.e., position, velocity, and acceleration (averaged over all subjects and also all trajectory derivatives, see section Frequency Resolved Decoding of Position, Velocity, and Acceleration). (E) “position + noise” model, CCFC frequency profile (purple curve, left y-axis) for the time-frequency magnitude and phase features at time offset τ = 0 s and with higher frequency resolution (window size = 5 s) plotted against the power spectral density of the time course of position along the trajectory (PSD, cyan, right y-axis). (F) Same as (E) for “velocity + noise” model (green curve, left y-axis). (G) Same as (E) for “acceleration + noise” model (brown curve, left y-axis). (H) The CCFC obtained from decoding the real ECoG data (red) plotted on top of the CCFC of the 3 models in (E–G). There is only a close match between the ECoG CCFC and the frequency profile resulting from the “velocity + noise” model (green).

Mentions: This observation led us to the hypothesis that the LFC could be understood as a close time domain copy of a kinematic variable. A somewhat more realistic scenario was constructed when white noise was added to the kinematic “copy” [thus, “copy + noise” model, kinematic signal (with std = 1) + white noise (with std = 20)]. In this case, expectedly, the FCs coinciding with the maximal power spectral density (PSD) of the movement position (Figures 9A,E), velocity (Figures 9B,F) and acceleration (Figures 9C,G), thereby having maximal signal-to-noise ratio, were also best decodable. Notably, the frequency profiles of the CCFC were quite different for each kinematic “copy + noise” model (average over all trajectories of all subjects).


The role of ECoG magnitude and phase in decoding position, velocity, and acceleration during continuous motor behavior.

Hammer J, Fischer J, Ruescher J, Schulze-Bonhage A, Aertsen A, Ball T - Front Neurosci (2013)

Time-frequency resolved decoding from “kinematic copy + white noise” models and from real ECoG. Correlation coefficients between predicted and actual frequency components (CCFC) computed at different carrier frequencies f and time offsets τ. The kinematic copy (std = 1) was summed up with white noise (std = 20). (A) “position + noise” model predicting position, (B) “velocity + noise” predicting velocity and (C) “acceleration + noise” model predicting acceleration. (D) The CCFC for real ECoG data predicting trajectory derivatives, i.e., position, velocity, and acceleration (averaged over all subjects and also all trajectory derivatives, see section Frequency Resolved Decoding of Position, Velocity, and Acceleration). (E) “position + noise” model, CCFC frequency profile (purple curve, left y-axis) for the time-frequency magnitude and phase features at time offset τ = 0 s and with higher frequency resolution (window size = 5 s) plotted against the power spectral density of the time course of position along the trajectory (PSD, cyan, right y-axis). (F) Same as (E) for “velocity + noise” model (green curve, left y-axis). (G) Same as (E) for “acceleration + noise” model (brown curve, left y-axis). (H) The CCFC obtained from decoding the real ECoG data (red) plotted on top of the CCFC of the 3 models in (E–G). There is only a close match between the ECoG CCFC and the frequency profile resulting from the “velocity + noise” model (green).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3814578&req=5

Figure 9: Time-frequency resolved decoding from “kinematic copy + white noise” models and from real ECoG. Correlation coefficients between predicted and actual frequency components (CCFC) computed at different carrier frequencies f and time offsets τ. The kinematic copy (std = 1) was summed up with white noise (std = 20). (A) “position + noise” model predicting position, (B) “velocity + noise” predicting velocity and (C) “acceleration + noise” model predicting acceleration. (D) The CCFC for real ECoG data predicting trajectory derivatives, i.e., position, velocity, and acceleration (averaged over all subjects and also all trajectory derivatives, see section Frequency Resolved Decoding of Position, Velocity, and Acceleration). (E) “position + noise” model, CCFC frequency profile (purple curve, left y-axis) for the time-frequency magnitude and phase features at time offset τ = 0 s and with higher frequency resolution (window size = 5 s) plotted against the power spectral density of the time course of position along the trajectory (PSD, cyan, right y-axis). (F) Same as (E) for “velocity + noise” model (green curve, left y-axis). (G) Same as (E) for “acceleration + noise” model (brown curve, left y-axis). (H) The CCFC obtained from decoding the real ECoG data (red) plotted on top of the CCFC of the 3 models in (E–G). There is only a close match between the ECoG CCFC and the frequency profile resulting from the “velocity + noise” model (green).
Mentions: This observation led us to the hypothesis that the LFC could be understood as a close time domain copy of a kinematic variable. A somewhat more realistic scenario was constructed when white noise was added to the kinematic “copy” [thus, “copy + noise” model, kinematic signal (with std = 1) + white noise (with std = 20)]. In this case, expectedly, the FCs coinciding with the maximal power spectral density (PSD) of the movement position (Figures 9A,E), velocity (Figures 9B,F) and acceleration (Figures 9C,G), thereby having maximal signal-to-noise ratio, were also best decodable. Notably, the frequency profiles of the CCFC were quite different for each kinematic “copy + noise” model (average over all trajectories of all subjects).

Bottom Line: The decoding accuracy based only on phase information was substantially (at least 2 fold) and significantly higher than that based only on magnitudes for position, velocity and acceleration.There was also no indication of additional magnitude-based mechanisms encoding movement information in the LFC range.Thus, our study contributes to elucidating the nature of the informative LFC of motor cortical population activity and may hence contribute to improve decoding strategies and BMI performance.

View Article: PubMed Central - PubMed

Affiliation: Bernstein Center Freiburg, University of Freiburg Freiburg, Germany.

ABSTRACT
In neuronal population signals, including the electroencephalogram (EEG) and electrocorticogram (ECoG), the low-frequency component (LFC) is particularly informative about motor behavior and can be used for decoding movement parameters for brain-machine interface (BMI) applications. An idea previously expressed, but as of yet not quantitatively tested, is that it is the LFC phase that is the main source of decodable information. To test this issue, we analyzed human ECoG recorded during a game-like, one-dimensional, continuous motor task with a novel decoding method suitable for unfolding magnitude and phase explicitly into a complex-valued, time-frequency signal representation, enabling quantification of the decodable information within the temporal, spatial and frequency domains and allowing disambiguation of the phase contribution from that of the spectral magnitude. The decoding accuracy based only on phase information was substantially (at least 2 fold) and significantly higher than that based only on magnitudes for position, velocity and acceleration. The frequency profile of movement-related information in the ECoG data matched well with the frequency profile expected when assuming a close time-domain correlate of movement velocity in the ECoG, e.g., a (noisy) "copy" of hand velocity. No such match was observed with the frequency profiles expected when assuming a copy of either hand position or acceleration. There was also no indication of additional magnitude-based mechanisms encoding movement information in the LFC range. Thus, our study contributes to elucidating the nature of the informative LFC of motor cortical population activity and may hence contribute to improve decoding strategies and BMI performance.

No MeSH data available.


Related in: MedlinePlus