ICLabel-style ICA Component Classification

Classifies independent components (from eegICA()) into seven classes - brain, muscle, eye, heart, line_noise, channel_noise, and other - and returns a calibrated probability for each class per component together with the argmax label. The approach follows the ICLabel framework: interpretable spatial, spectral, and temporal features are extracted from each component and mapped to class probabilities by a lightweight multinomial-logistic (softmax) head.

Usage

eegICLabel(
  x,
  ica_assay = "ica_components",
  assay_name = NULL,
  line_freq = 50,
  backend = c("heuristic", "iclabel")
)

Arguments

x: A PhysioExperiment object with ICA results (from eegICA()).
ica_assay: Metadata name holding the component activations (default: "ica_components").
assay_name: Input assay used only to resolve channel labels (default: first assay).
line_freq: Mains line frequency in Hz used for the line-noise features (default: 50). Only used by the heuristic backend.
backend: Which classifier to use: "heuristic" (default) is the self-contained pure-R classifier (real features, hand-set multinomial weights); "iclabel" runs the genuine trained ICLabel CNN (Pion-Tonachini et al. 2019) by delegating to mne-icalabel via reticulate (needs a Python env with mne and mne-icalabel; channel labels must match a standard 10-20/10-10 montage for the scalp-map).

Value

A data.frame with one row per component: component (integer index), one numeric column per class (brain, muscle, eye, heart, line_noise, channel_noise, other) holding probabilities that sum to 1 across the classes, and label (character argmax class).

Details

Features per component:

Spatial: frontal energy (eye/blink topography), focality (single channel dominance), topography kurtosis.
Spectral: 1/f slope, high- versus low-frequency band ratio, low-frequency fraction, a genuine alpha-peak measure (8-12 Hz power above its theta/low-beta neighbours), and a line-noise power ratio and fraction.
Temporal: lag-1 autocorrelation, activation kurtosis (spiky blink or ECG signatures), and a roughly 1 Hz periodicity measure for cardiac components.

The softmax weights are read from inst/extdata/iclabel_weights.csv; if that file is unavailable an identical built-in weight table is used, so the classifier always works.

References

Pion-Tonachini, L., Kreutz-Delgado, K., & Makeig, S. (2019). ICLabel: An automated electroencephalographic independent component classifier, dataset, and website. NeuroImage, 198, 181-197.

Winkler, I., Haufe, S., & Tangermann, M. (2011). Automatic classification of artifactual ICA-components for artifact removal in EEG signals. Behavioral and Brain Functions, 7, 30.

Examples

if (FALSE) { # \dontrun{
pe <- make_eeg(n_time = 5000, sr = 500)
pe <- eegICA(pe, n_components = 10, method = "fastica")
probs <- eegICLabel(pe)
} # }