DECODINGLIB::DecodingCsp Class Reference

Common Spatial Patterns decoder for two-class discriminative spatial filtering. More...

#include <decoding_csp.h>

Public Types
enum class	TransformMode { AveragePower , CspSpace }

Public Member Functions
	DecodingCsp (int nComponents=4, TransformMode transformInto=TransformMode::AveragePower, bool useLog=true)
void	fit (const std::vector< Eigen::MatrixXd > &epochs, const Eigen::VectorXi &y)
Eigen::MatrixXd	transform (const std::vector< Eigen::MatrixXd > &epochs) const
Eigen::MatrixXd	fitTransform (const std::vector< Eigen::MatrixXd > &epochs, const Eigen::VectorXi &y)
Eigen::MatrixXd	inverseTransform (const Eigen::MatrixXd &X) const
const Eigen::MatrixXd &	filters () const
const Eigen::MatrixXd &	patterns () const
const Eigen::VectorXd &	mean () const
const Eigen::VectorXd &	stddev () const
bool	isFitted () const

Detailed Description

Common Spatial Patterns decoder for two-class discriminative spatial filtering.

Estimates the joint diagonaliser of two class-conditional covariance matrices estimated from labelled, band-passed epochs and exposes the top and bottom n_components eigenvectors as a bank of spatial filters; the corresponding patterns (the columns of the pseudoinverse of the filter matrix) describe the sensor-space activity that each filter is sensitive to and are what should be plotted as topographies for neurophysiological interpretation, as discussed by Haufe et al.

1. After fit the class works as a deterministic feature extractor: transform reduces an epoch tensor either to one log-power (or z-scored) value per component (AveragePower, the standard input to an LDA / logistic-regression classifier) or to the full time-resolved projection (CspSpace, useful for downstream Riemannian or deep-learning stages).

The implementation expects exactly two unique class labels in y and epochs already restricted to the discriminative frequency band (e.g. 8–30 Hz for sensorimotor rhythms); no regularisation is applied to the covariance estimate, so callers should ensure enough trials per class to avoid the ill-conditioned regime that motivates the regularised variants of Lotte & Guan 2011. inverseTransform provides the closed-form back-projection of a power-feature vector into sensor space, which is what the application layer uses to render CSP topographies side-by-side with the discriminative scores.

See also

DECODINGLIB, mne.decoding.CSP

Definition at line 98 of file decoding_csp.h.

Member Enumeration Documentation

TransformMode

enum class DECODINGLIB::DecodingCsp::TransformMode

strong

Transform mode for the CSP output.

Enumerator
AveragePower	Return average band power per component (n_epochs × n_components).
CspSpace	Return data projected into CSP space (n_epochs × n_components × n_times).

Definition at line 105 of file decoding_csp.h.

Constructor & Destructor Documentation

DecodingCsp()

DecodingCsp::DecodingCsp ( int nComponents = 4, TransformMode transformInto = TransformMode::AveragePower, bool useLog = true )

explicit

Constructs a CSP decoder.

Parameters

[in]	nComponents	Number of CSP components (split between classes). Default: 4.
[in]	transformInto	Feature extraction mode. Default: AveragePower.
[in]	useLog	If true (default) and transformInto == AveragePower, apply log transform; otherwise z-score features using mean_ and std_.

Definition at line 62 of file decoding_csp.cpp.

Member Function Documentation

filters()

const MatrixXd & DecodingCsp::filters

(

)

const

Returns

Spatial filters (n_components × n_channels).

Definition at line 258 of file decoding_csp.cpp.

fit()

void DecodingCsp::fit	(	const std::vector< Eigen::MatrixXd > &	epochs,
		const Eigen::VectorXi &	y )

Fit CSP from labelled epoch data (binary classification).

Parameters

[in]	epochs	Vector of epoch matrices, each (n_channels × n_times).
[in]	y	Class label for each epoch (must contain exactly 2 unique values).

Definition at line 73 of file decoding_csp.cpp.

fitTransform()

MatrixXd DecodingCsp::fitTransform	(	const std::vector< Eigen::MatrixXd > &	epochs,
		const Eigen::VectorXi &	y )

Fit and transform in one step.

Parameters

[in]	epochs	Epoch data.
[in]	y	Class labels.

Returns

Feature matrix (same as transform output).

Definition at line 226 of file decoding_csp.cpp.

inverseTransform()

MatrixXd DecodingCsp::inverseTransform

(

const Eigen::MatrixXd &

X

)

const

Project CSP power features back to sensor space.

Only valid when transformInto == AveragePower.

Parameters

[in]

X

Feature matrix (n_epochs × n_components).

Returns

Sensor-space projection (n_epochs × n_channels × n_components). Stored as (n_epochs, n_channels * n_components) flattened.

Definition at line 235 of file decoding_csp.cpp.

isFitted()

bool DecodingCsp::isFitted

(

)

const

Returns

True if the model has been fitted.

Definition at line 298 of file decoding_csp.cpp.

mean()

const VectorXd & DecodingCsp::mean

(

)

const

Returns

Mean band power per component (computed during fit).

Definition at line 278 of file decoding_csp.cpp.

patterns()

const MatrixXd & DecodingCsp::patterns

(

)

const

Returns

Spatial patterns (n_channels × n_components).

Definition at line 268 of file decoding_csp.cpp.

stddev()

const VectorXd & DecodingCsp::stddev

(

)

const

Returns

Standard deviation of band power per component (computed during fit).

Definition at line 288 of file decoding_csp.cpp.

transform()

MatrixXd DecodingCsp::transform

(

const std::vector< Eigen::MatrixXd > &

epochs

)

const

Transform epoch data using the fitted CSP filters.

When transformInto == AveragePower, returns (n_epochs × n_components) with log-transformed or z-scored mean band power.

When transformInto == CspSpace, returns a matrix where each n_components rows correspond to one epoch's CSP-space projection. The shape is (n_epochs * n_components, n_times).

Parameters

[in]

epochs

Vector of epoch matrices, each (n_channels × n_times).

Returns

Feature matrix.

Definition at line 187 of file decoding_csp.cpp.

---

The documentation for this class was generated from the following files:

• src/libraries/decoding/decoding_csp.h
• src/libraries/decoding/decoding_csp.cpp