Skip to main content

DecodingSpoc

Namespace: DECODINGLIB  ·  Library: Decoding Library

Python equivalent

mne.decoding.SPoC in MNE-Python.

#include <decoding/decoding_spoc.h>

class DECODINGLIB::DecodingSpoc

Source Power Comodulation decoder for continuous-target regression on band-power.

Estimates a bank of spatial filters whose log band-power envelope is maximally correlated with a continuous trial-level target. The fit builds the unweighted trial-mean covariance and a target-weighted covariance from centred labels, solves the resulting generalised eigenvalue problem, and keeps the top n_components eigenvectors as filters and the columns of their pseudoinverse as the corresponding activation patterns. As with CSP, the patterns — not the filters — are the quantity that should be plotted as topographies when interpreting which cortical sources drive the regression, in line with Haufe et al. 2014.

After fit the decoder behaves as a deterministic feature extractor. AveragePower returns one (optionally log- or z-scored) band-power value per component and epoch, which plugs directly into a linear regressor; CspSpace returns the time- resolved filtered signal, useful when the downstream model consumes envelopes or instantaneous phase. Inputs must be epoched and already filtered in the relevant frequency band; the target vector zz is a real-valued vector of length n_epochs.

src/examples/ex_decoding/main.cpp
DecodingSpoc spoc(1);
spoc.fit(spocEpochs, target);
const MatrixXd spocFeatures = spoc.transform(spocEpochs); // epochs x 1, log band power

Public Methods​

DecodingSpoc(nComponents, transformInto, useLog)​

Constructs a SPoC decoder.

Parameters:

  • nComponents : int Number of components. Default: 4.

  • transformInto : TransformMode Feature extraction mode. Default: AveragePower.

  • useLog : bool If true and transformInto == AveragePower, apply log transform.

fit(epochs, y)​

Fit SPoC from epoch data and a continuous target variable.

Parameters:

  • epochs : const std::vector< Eigen::MatrixXd > & Vector of epoch matrices, each (n_channels × n_times).

  • y : const Eigen::VectorXd & Continuous target variable (one value per epoch).

transform(epochs)​

Transform epoch data using the fitted SPoC filters.

Parameters:

  • epochs : const std::vector< Eigen::MatrixXd > & Vector of epoch matrices.

Returns:

  • Eigen::MatrixXd — Feature matrix.

fitTransform(epochs, y)​

Fit and transform in one step.

Parameters:

  • epochs : const std::vector< Eigen::MatrixXd > & Vector of epoch matrices, each (n_channels × n_times).

  • y : const Eigen::VectorXd & Continuous target variable (one value per epoch).

Returns:

  • Eigen::MatrixXd — Feature matrix: (n_epochs × n_components) band powers, or stacked (n_epochs·n_components × n_times) projections in CspSpace mode.

filters()​

Returns:

  • const Eigen::MatrixXd & — Spatial filters (n_components × n_channels).

patterns()​

Returns:

  • const Eigen::MatrixXd & — Spatial patterns (n_channels × n_components).

mean()​

Returns:

  • const Eigen::VectorXd & — Mean band power per component (computed during fit).

stddev()​

Returns:

  • const Eigen::VectorXd & — Standard deviation of band power per component.

isFitted()​

Returns:

  • bool — True if the model has been fitted.

Example​

Used in examples​

ex_decoding​

Spatial decoders and ICA labelling on synthetic mixtures with a known ground truth.

src/examples/ex_decoding/main.cpp, line 153
DecodingSpoc spoc(1);
spoc.fit(spocEpochs, target);
const MatrixXd spocFeatures = spoc.transform(spocEpochs); // epochs x 1, log band power
const double spocCorr = correlation(spocFeatures.col(0), target.array().log().matrix());
ok &= expect(std::fabs(spocCorr) > 0.9, QString("SPoC power follows log(target): |r| %1 > 0.9").arg(std::fabs(spocCorr)));

Authors of this file​