eegdash.features.feature_bank.csp#

Common Spatial Pattern Features Extraction#

This module provides the Common Spatial Pattern (CSP) feature extractor for signal classification.

Data Shape Convention#

This module follows a Time-Last convention:

Input: (..., time)
Output: (...,)

All functions collapse the last dimension (time), returning an ndarray of features corresponding to the leading dimensions (e.g., subjects, channels).

Classes

CommonSpatialPattern()

Common Spatial Pattern (CSP) for binary signal classification.

class eegdash.features.feature_bank.csp.CommonSpatialPattern[source]

Bases: TrainableFeature

Common Spatial Pattern (CSP) for binary signal classification.

CSP finds spatial filters that maximize the variance for one class while minimizing it for the other. It transforms multi-channel signals into a subspace where the differences between two conditions are most prominent.

_weights

The spatial filter matrix.

Type:: ndarray

_eigvals

The eigenvalues representing the variance ratio for class 0.

Type:: ndarray

_means

The class-wise means used for centering.

Type:: ndarray

_covs

The class-wise covariance matrices.

Type:: ndarray

Notes

This implementation supports online learning through partial_fit, allowing the model to be updated with new batches.

For a theoretical overview of Common Spatial Patterns, see the Wikipedia entry.

clear()[source]: Reset the internal state of the feature extractor.

See also

clear()

partial_fit(x, y=None)[source]

Incrementally update class-wise mean and covariance statistics.

Parameters:

x (ndarray) – Input array of shape (n_epochs, n_channels, n_times).
y (ndarray) – Class labels for each epoch (must contain exactly two classes).

Raises:

AssertionError – If more than two unique labels are detected across all partial fits.

static transform_input(x)[source]

Reshape and transpose epoch data for matrix operations.

Converts 3D epoch data into a 2D format suitable for covariance estimation and spatial filtering. The temporal dimension is collapsed into the samples dimension.

Parameters:: x (ndarray) – Input array of shape (n_epochs, n_channels, n_times).
Returns:: Reshaped array of shape (n_epochs * n_times, n_channels).
Return type:: ndarray

fit()[source]

Solve the generalized eigenvalue problem to find spatial filters.

Calculates the filters $W$ such that the ratio of variances between the two classes is maximized. Filters are sorted by their discriminative power (distance from 0.5 eigenvalue).