decoding_ssd.cpp

Go to the documentation of this file.

//=============================================================================================================
 
//=============================================================================================================
// INCLUDES
//=============================================================================================================
 
#include "decoding_ssd.h"
 
//=============================================================================================================
// EIGEN INCLUDES
//=============================================================================================================
 
#include <Eigen/Eigenvalues>
 
//=============================================================================================================
// STL INCLUDES
//=============================================================================================================
 
#include <algorithm>
#include <cmath>
#include <stdexcept>
 
//=============================================================================================================
// USED NAMESPACES
//=============================================================================================================
 
using namespace DECODINGLIB;
using namespace Eigen;
 
//=============================================================================================================
 
DecodingSsd::DecodingSsd(int nComponents, double regParam)
: m_nComponents(nComponents)
, m_regParam(regParam)
{
}
 
//=============================================================================================================
 
void DecodingSsd::fit(const Ref<const MatrixXd>& data,
                      double sfreq,
                      double signalLow, double signalHigh,
                      double noiseLow, double noiseHigh)
{
    if (noiseLow > signalLow || signalHigh > noiseHigh) {
        throw std::invalid_argument(
            "DecodingSsd::fit: signal band must be within noise band");
    }
 
    const auto n_ch = data.rows();
    const auto n_times = data.cols();
 
    if (n_ch < 2 || n_times < 2) {
        throw std::invalid_argument(
            "DecodingSsd::fit: data must have >= 2 channels and >= 2 time points");
    }
 
    int n_comp = std::min(m_nComponents, static_cast<int>(n_ch));
 
    // Mean-center
    MatrixXd centered = data.colwise() - data.rowwise().mean();
 
    // Bandpass filter for signal and noise bands
    MatrixXd data_signal = bandpassFilter(centered, sfreq, signalLow, signalHigh);
    MatrixXd data_noise = bandpassFilter(centered, sfreq, noiseLow, noiseHigh);
 
    // Covariance matrices
    MatrixXd cov_signal = (data_signal * data_signal.transpose()) / static_cast<double>(n_times - 1);
    MatrixXd cov_noise = (data_noise * data_noise.transpose()) / static_cast<double>(n_times - 1);
 
    // Regularize noise covariance
    cov_noise += m_regParam * cov_noise.trace() / static_cast<double>(n_ch) * MatrixXd::Identity(n_ch, n_ch);
 
    // Generalized eigenvalue problem: C_signal w = λ C_noise w
    GeneralizedSelfAdjointEigenSolver<MatrixXd> ges(cov_signal, cov_noise);
    if (ges.info() != Eigen::Success) {
        throw std::runtime_error(
            "DecodingSsd::fit: eigenvalue decomposition failed");
    }
 
    // Eigenvalues ascending → reverse for descending
    VectorXd evals = ges.eigenvalues().reverse();
    MatrixXd evecs = ges.eigenvectors().rowwise().reverse();
 
    m_eigenvalues = evals.head(n_comp);
    m_filters = evecs.leftCols(n_comp).transpose(); // (n_comp × n_ch)
 
    // Patterns: A = C W inv(W^T C W)
    MatrixXd cov_full = (centered * centered.transpose()) / static_cast<double>(n_times - 1);
    MatrixXd WtCW = m_filters * cov_full * m_filters.transpose();
    m_patterns = (cov_full * m_filters.transpose() * WtCW.inverse()); // (n_ch × n_comp)
 
    m_fitted = true;
}
 
//=============================================================================================================
 
MatrixXd DecodingSsd::transform(const Ref<const MatrixXd>& data) const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSsd::transform: not fitted");
    }
    return m_filters * data;
}
 
//=============================================================================================================
 
MatrixXd DecodingSsd::fitTransform(const Ref<const MatrixXd>& data,
                                   double sfreq,
                                   double signalLow, double signalHigh,
                                   double noiseLow, double noiseHigh)
{
    fit(data, sfreq, signalLow, signalHigh, noiseLow, noiseHigh);
    return transform(data);
}
 
//=============================================================================================================
 
MatrixXd DecodingSsd::apply(const Ref<const MatrixXd>& data) const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSsd::apply: not fitted");
    }
 
    // Denoise: project to SSD space, then reconstruct using patterns
    // patterns_ is (n_ch × n_comp), ssdSources is (n_comp × n_times)
    MatrixXd ssdSources = transform(data);
    return m_patterns * ssdSources;
}
 
//=============================================================================================================
 
const MatrixXd& DecodingSsd::filters() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSsd::filters: not fitted");
    }
    return m_filters;
}
 
//=============================================================================================================
 
const MatrixXd& DecodingSsd::patterns() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSsd::patterns: not fitted");
    }
    return m_patterns;
}
 
//=============================================================================================================
 
const VectorXd& DecodingSsd::eigenvalues() const
{
    if (!m_fitted) {
        throw std::runtime_error("DecodingSsd::eigenvalues: not fitted");
    }
    return m_eigenvalues;
}
 
//=============================================================================================================
 
bool DecodingSsd::isFitted() const
{
    return m_fitted;
}
 
//=============================================================================================================
 
MatrixXd DecodingSsd::bandpassFilter(const MatrixXd& data,
                                     double sfreq,
                                     double lowFreq,
                                     double highFreq)
{
    const auto n_ch = data.rows();
    const auto n_times = data.cols();
 
    int filter_order = std::min(
        static_cast<int>(std::round(3.0 * sfreq / lowFreq)),
        static_cast<int>(n_times) - 1);
    if (filter_order % 2 != 0)
        ++filter_order;
    filter_order = std::min(filter_order, 128);
 
    const int half = filter_order / 2;
    const double pi = M_PI;
 
    // Windowed-sinc coefficients
    VectorXd h(filter_order + 1);
    double w_low = 2.0 * pi * lowFreq / sfreq;
    double w_high = 2.0 * pi * highFreq / sfreq;
 
    for (int i = 0; i <= filter_order; ++i) {
        int n = i - half;
        if (n == 0) {
            h(i) = (w_high - w_low) / pi;
        } else {
            h(i) = (std::sin(w_high * static_cast<double>(n)) - std::sin(w_low * static_cast<double>(n))) / (pi * static_cast<double>(n));
        }
        // Hamming window
        h(i) *= 0.54 - 0.46 * std::cos(2.0 * pi * static_cast<double>(i) / static_cast<double>(filter_order));
    }
 
    // Normalize to unit gain at center frequency
    double center = (lowFreq + highFreq) / 2.0;
    double w_center = 2.0 * pi * center / sfreq;
    double gain = 0.0;
    for (int i = 0; i <= filter_order; ++i) {
        gain += h(i) * std::cos(w_center * static_cast<double>(i - half));
    }
    if (std::abs(gain) > 1e-12)
        h /= std::abs(gain);
 
    // Forward + reverse convolution (zero-phase)
    MatrixXd result(n_ch, n_times);
 
    for (Index ch = 0; ch < n_ch; ++ch) {
        VectorXd forward(n_times);
        for (Index t = 0; t < n_times; ++t) {
            double sum = 0.0;
            for (int k = 0; k <= filter_order; ++k) {
                auto idx = t - k;
                if (idx >= 0 && idx < n_times)
                    sum += h(k) * data(ch, idx);
            }
            forward(t) = sum;
        }
 
        VectorXd backward(n_times);
        for (Index t = n_times - 1; t >= 0; --t) {
            double sum = 0.0;
            for (int k = 0; k <= filter_order; ++k) {
                auto idx = t + k;
                if (idx >= 0 && idx < n_times)
                    sum += h(k) * forward(idx);
            }
            backward(t) = sum;
        }
 
        result.row(ch) = backward.transpose();
    }
 
    return result;
}