ICA

Namespace: RTPROCESSINGLIB  ·  Library: DSP Library

Python equivalent

mne.preprocessing.ICA (fastica) in MNE-Python.

#include <dsp/ica.h>

class UTILSLIB::ICA

Independent Component Analysis using the FastICA algorithm (deflationary, logcosh nonlinearity).

src/examples/ex_dsp_preprocessing/main.cpp

const IcaResult ica = ICA::run(mixed, 3);                         // FastICA, logcosh, fixed seed
const MatrixXd cleaned = ICA::excludeComponents(mixed, ica, {0}); // drop component 0, back-project the rest

---

Static Methods

run(matData, nComponents, maxIter, tol, randomSeed)

Fit FastICA on the given data matrix.

Parameters:

• matData : const Eigen::MatrixXd & Input data (n_channels x n_samples). Each row is one sensor channel.

• nComponents : int Number of independent components to extract. Pass -1 (default) to use all channels.

• maxIter : int Maximum iterations per component (default 200).

• tol : double Convergence tolerance on |w_new · w_old| - 1 (default 1e-4).

• randomSeed : int Seed for the random weight initialisation (default 42).

Returns:

• IcaResult — IcaResult containing mixing/unmixing matrices and source time series.

---

applyUnmixing(matData, result)

Project new data through a previously fitted unmixing matrix.

The same mean centering as during fitting is applied.

Parameters:

• matData : const Eigen::MatrixXd & New data (n_channels x n_samples), same channel order as training data.

• result : const IcaResult & Result from a previous ICA::run() call.

Returns:

• Eigen::MatrixXd — Source matrix (n_components x n_samples).

---

excludeComponents(matData, result, excludedComponents)

Reconstruct sensor-space data after suppressing selected components.

The excluded component(s) are zeroed in source space before back-projection.

Parameters:

• matData : const Eigen::MatrixXd & Original sensor data (n_channels x n_samples).

• result : const IcaResult & Result from ICA::run() on matData (or a representative segment).

• excludedComponents : const QVector< int > & 0-based indices of components to suppress (e.g. ECG, EOG artifacts).

Returns:

• Eigen::MatrixXd — Cleaned sensor data (n_channels x n_samples).

---

Example

• Full source: src/examples/ex_dsp_preprocessing/main.cpp
• Build target: ex_dsp_preprocessing (runs under CTest)

Used in examples

ex_dsp_preprocessing

Preprocessing tools of the dsp library on synthetic data with a known ground truth.

src/examples/ex_dsp_preprocessing/main.cpp, line 122

const IcaResult ica = ICA::run(mixed, 3);                         // FastICA, logcosh, fixed seed
const MatrixXd cleaned = ICA::excludeComponents(mixed, ica, {0}); // drop component 0, back-project the rest
ok &= expect(worstRecovery(ica.matSources, sources) > 0.99, QString("FastICA recovers all sources (worst |r| %1)").arg(worstRecovery(ica.matSources, sources)));
ok &= expect((mixed - cleaned - ica.matMixing.col(0) * ica.matSources.row(0)).cwiseAbs().maxCoeff() < 1e-9,
             "excluding a component removes exactly its back-projection");

Authors of this file

• Christoph Dinh <christoph.dinh@mne-cpp.org>