9 Computational EEG Analysis for Brain-Computer Interfaces
211
9.4 Other Considerations
9.4.1 Artifact Removal
When needed, BCIs implement many of the same artifact removal techniques discussed in Chap. 2. However, for certain BCI applications, it may not be worth the additional effort and computational resources to design and implement artifact removal
for rare or inconsequential artifacts such as eye blinks, for instance. It is common to
design processing and classification stages to be tuned to the control signal of interest
and inherently immune to artifacts. For many of the aforementioned BCI paradigms,
the control signals have a distinct spatial and spectral distribution from common
artifacts, which can be sufficiently attenuated via standard spatial or spectral filtering without the need for specialized artifact characterization or processing. Without
careful design, an artifact removal technique may also further distort the control
signal and increase computational resources such that there is ultimately little or no
practical benefit over excluding artifact removal from the processing chain.
9.4.2 Real-Time Processing
With the considerable advances available in modern computing technology, the realtime processing requirements for BCIs are no longer as constraining as they once
were. All of the methods presented in this chapter are fully able to be implemented
with minimal delays due to signal processing.
Most standard signal preprocessing steps such as spatial and spectral filtering can
be efficiently implemented to achieve real-time feedback. For instance, spatial filter
parameters can be computed offline with online implementations that require negligible computational resources. Online computation of spatial filters such as online
ICA requires significant computational resources for real-time feedback and are not
commonly implemented for BCIs. For spectral filtering, infinite impulse response
(IIR) filter structures are preferred over finite impulse response (FIR) structures. For
causal online filtering, symmetric FIR filters introduce a delay equivalent to half
the filter length, while IIR filters with comparable frequency characteristics can be
designed having significantly shorter latencies. The trade-off for IIR filters compared
to FIR is that they introduce phase distortion in signals. Depending on the application, this may be tolerable or included in the classifier design without detriment. If
phase distortion is not tolerable, zero-phase filtering can be implemented to eliminate the phase distortion while effectively doubling the latency compared to using
the equivalent filter without a zero-phase implementation [28].
However, it should still be noted that BCIs based on continuous responses are
often far more restrictive than those based on transient responses. For instance, consider a continuous BCI using SSVEP. The processing should occur within a single
data window in order to prevent a lagged control feedback to the user, which could
211
9.4 Other Considerations
9.4.1 Artifact Removal
When needed, BCIs implement many of the same artifact removal techniques discussed in Chap. 2. However, for certain BCI applications, it may not be worth the additional effort and computational resources to design and implement artifact removal
for rare or inconsequential artifacts such as eye blinks, for instance. It is common to
design processing and classification stages to be tuned to the control signal of interest
and inherently immune to artifacts. For many of the aforementioned BCI paradigms,
the control signals have a distinct spatial and spectral distribution from common
artifacts, which can be sufficiently attenuated via standard spatial or spectral filtering without the need for specialized artifact characterization or processing. Without
careful design, an artifact removal technique may also further distort the control
signal and increase computational resources such that there is ultimately little or no
practical benefit over excluding artifact removal from the processing chain.
9.4.2 Real-Time Processing
With the considerable advances available in modern computing technology, the realtime processing requirements for BCIs are no longer as constraining as they once
were. All of the methods presented in this chapter are fully able to be implemented
with minimal delays due to signal processing.
Most standard signal preprocessing steps such as spatial and spectral filtering can
be efficiently implemented to achieve real-time feedback. For instance, spatial filter
parameters can be computed offline with online implementations that require negligible computational resources. Online computation of spatial filters such as online
ICA requires significant computational resources for real-time feedback and are not
commonly implemented for BCIs. For spectral filtering, infinite impulse response
(IIR) filter structures are preferred over finite impulse response (FIR) structures. For
causal online filtering, symmetric FIR filters introduce a delay equivalent to half
the filter length, while IIR filters with comparable frequency characteristics can be
designed having significantly shorter latencies. The trade-off for IIR filters compared
to FIR is that they introduce phase distortion in signals. Depending on the application, this may be tolerable or included in the classifier design without detriment. If
phase distortion is not tolerable, zero-phase filtering can be implemented to eliminate the phase distortion while effectively doubling the latency compared to using
the equivalent filter without a zero-phase implementation [28].
However, it should still be noted that BCIs based on continuous responses are
often far more restrictive than those based on transient responses. For instance, consider a continuous BCI using SSVEP. The processing should occur within a single
data window in order to prevent a lagged control feedback to the user, which could