BCI Classification

Classifies BCI features using Linear Discriminant Analysis (LDA) or shrinkage LDA. Implements Fisher's LDA with optional Ledoit-Wolf shrinkage regularization for robust classification with high-dimensional or small-sample data. Optionally performs k-fold cross-validation.

Usage

eegBCIclassify(
  x,
  features = NULL,
  labels,
  method = c("lda", "shrinkage_lda"),
  cv_folds = NULL,
  assay_name = NULL
)

Arguments

x

A PhysioExperiment object with epoched (3D) EEG data.

features

Optional pre-computed feature matrix (n_trials x n_features). If NULL, features are extracted using eegBCIfeatures with the "bandpower" method.

labels

Character or factor vector of class labels, one per trial. Must contain exactly two unique classes.

method

Classification method: "lda" (Fisher's LDA) or "shrinkage_lda" (LDA with Ledoit-Wolf shrinkage).

cv_folds

Number of cross-validation folds (default: NULL, meaning no cross-validation). If set (e.g., 5), performs k-fold CV and reports out-of-fold predictions. The CV accuracy is stored as attr(result, "cv_accuracy").

assay_name

Input assay name used when extracting features (default: first assay).

Value

A data.frame with columns: trial, predicted_class, confidence, and true_class. When cv_folds is not NULL, predictions are out-of-fold and attr(result, "cv_accuracy") contains the cross-validated accuracy. The trained LDA model (on all data) is stored in metadata(x)$bci_model, containing weights, threshold, classes, method, and class_means.

References

Blankertz, B., et al. (2008). Optimizing spatial filters for robust EEG single-trial analysis. IEEE Signal Processing Magazine, 25(1), 41-56.

See also

eegCSP(), eegBCIfeatures(), eegMotorImagery()

Examples

if (FALSE) { # \dontrun{
pe <- make_eeg_bci(n_trials = 20, n_channels = 8, sr = 256)
labels <- metadata(pe)$labels
features <- eegBCIfeatures(pe, method = "bandpower")
result <- eegBCIclassify(pe, features = features, labels = labels, method = "lda")

# With 5-fold cross-validation
result_cv <- eegBCIclassify(pe, features = features, labels = labels,
                            method = "lda", cv_folds = 5)
attr(result_cv, "cv_accuracy")
} # }