iCanClean

iCanClean uses canonical correlation between a primary recording and a physical or pseudo-reference recording to remove shared variance. The reference can be built from selected channels or from a supplied reference array [1][2].

Usage

from mne_denoise.icanclean import ICanClean

model = ICanClean(
    sfreq=250.0,
    ref_channels=[6, 7],
)
clean = model.fit_transform(data)  # data: (n_channels, n_times)

Key points

• The primary input and reference must have matching observations. The functional API accepts channel-first continuous arrays; the estimator also supports its documented MNE containers.

• Modes control the cleaning workflow: sliding, global, calibrated, and hybrid are supported. A physical reference uses supplied channels; pseudo-reference construction is enabled separately with pseudo_ref=True and requires a suitable filter_ref.

• Thresholds can use an explicit threshold or the circular-shift null option. null_r2_threshold returns a quantile of the maximum squared CCA correlation across surrogate shifts; a high shared correlation does not by itself identify artifact.

• fit is a compatibility no-op for this transductive estimator; cleaning is estimated during transform/fit_transform.

References

---

[1]

Ryan J. Downey and Daniel P. Ferris. The icanclean algorithm: how to remove artifacts using reference noise recordings. 2022. arXiv:2201.11798, doi:10.48550/arXiv.2201.11798.

[2]

Ryan J. Downey and Daniel P. Ferris. Icanclean removes motion, muscle, eye, and line-noise artifacts from phantom eeg. Sensors, 23(19):8214, 2023. doi:10.3390/s23198214.