Ocular correction: ICA and regression

ICA correction is a sequence of training preparation, decomposition, component selection and continuous-data correction. These are separate checkpoints. Changing a component decision does not necessarily require retraining; changing the training sample selection or training filters generally does.

1. Review training-data rejection

ICA uses adjacent artificial training epochs, not the final experimental event-related epochs. Configure their length, training-only filters and automatic channel/epoch rejection criteria.

The ICA training-data rejection dialog lets you inspect automatic proposals and adjust channel/epoch decisions. Accepting training decisions allows the prepared data to be used by Infomax. Whitening and training preparation are reusable only while their settings and input identity remain compatible.

An optional training-epoch baseline is disabled by default. If enabled, it subtracts the channel mean over the whole artificial epoch or the configured inclusive point range before rejection/whitening. It affects training only. Do not enable it merely to make the display look centred.

Rejecting a training preparation and retrying with different settings is different from stopping the entire batch. Read the offered action and retention choice; do not assume that every Cancel/Reject action has the same scope.

2. Inspect convergence

The ICA convergence view shows extended-Infomax progress and the weight-change history. A stopped decomposition may be available for provisional component inspection; provisional does not mean converged or accepted.

Compatible optimizer checkpoints allow exact continuation rather than starting again. Changed training identity can require a new preparation and fit. Check the status and log before treating an interrupted fit as final. Show last ICA convergence in the ERP/ERF tools reopens the available graph.

3. Review components

The ICA component-review dialog combines component scalp maps, activations, spectra, training-trial heatmaps and configured EOG evidence. Saved selections can be restored for review. Automatic suggestions are aids to inspection, not a guarantee that a component is ocular or safe to remove.

Use the displays together; a global EOG correlation alone does not identify a component. Component polarity is arbitrary. The optional average-reference display changes the shown maps, not the stored decomposition or correction. Hiding a trace changes visibility, not its analytical inclusion.

A run ending at Decide ICA component rejection saves the decision without yet writing an accepted corrected recording.

4. Preview the correction

Apply ICA correction subtracts selected components from the broadband continuous source. In review mode, the before/after display compares the original and provisional correction. Inspect both the artefact and nearby signal before accepting.

You can accept, return to component selection, or reject the decomposition using the offered workflow. Returning to selection rebuilds the provisional preview. Rejecting a subject in a batch is distinct from stopping the batch; retention/retraining choices determine which ICA checkpoints remain available.

The final corrected .ica.TAW and its configuration become accepted outputs after the decision. With no selected components, correction is an identity copy. Display baselines and hidden traces do not alter that output.

With reuse enabled, an already accepted compatible correction can be reused. Use the flow-control range and cleanup preview deliberately when revisiting a completed decision.

Export a completed decomposition to EEGLAB

Use Tools > ERP/ERF tools > Third party integration tools > Export completed EMEGS ICA decomposition to EEGLAB and select the .ica.mat file. The export contains the broadband data before component removal and the decomposition. It produces an EEGLAB .set with a companion .fdt; keep both together. No component is marked rejected, allowing a separate correction in EEGLAB.

Regression alternative

The regression option uses configured EOG predictors on extracted epochs. It is not an additional post-epoch correction when continuous ICA has already been applied. Selecting regression changes the available step list and the meaning of the epoch/correction stage. Check the EOG channel definitions and the selected calibration/all-trial option before running it.

For reproducibility, record the chosen correction method, training settings, rejected channels/epochs and component decisions. This Help explains operation; component-selection criteria require scientific review for your recordings.