Referencing

Common average and median reference

At every sample, subtract the mean (CAR) or the median (CMR) over channels. The median survives a bad channel, where the mean would spread it to all the others.

Python

y = dspk.ops.spatial.cmr(x)       # or car

Rust

pipeline.add(PipelineStage::common_median_reference());   // or common_average_reference()

Spec

{"op": "cmr"}

App

Reference → CMR or CAR.

Surface Laplacian

Each electrode minus the mean of its neighbours, on a grid. It removes what spreads across the whole array and keeps what is local: useful for HD-EMG and ECoG. Inside the grid a channel has 4 neighbours, on an edge 3, and in a corner 2. Channels are row-major: channel row * cols + col.

Python

lap = dspk.ops.spatial.Laplacian.from_grid(8, 8)
y = lap.apply(x)

Rust

pipeline.add(PipelineStage::SurfaceLaplacian(SurfaceLaplacian::from_grid_2d(8, 8)));

Spec

{"op": "laplacian", "grid": [8, 8]}

App

Reference → Laplacian, with presets for 4×8, 8×4, 8×8 and 16×4 grids.

A grid of another size than the chain’s channels is refused, with the reason. Python also has a distance-weighted version from electrode positions, Laplacian.from_positions.

Whitening

Decorrelates the channels, as sorters do before matching: the result has uncorrelated channels of unit variance. It is a matrix fitted on the signal.

Python

w = dspk.ops.spatial.Whitening.fit_zca(x)
y = w.apply(x)

Rust

let white = SpatialWhitening::fit_zca::<f32>(&client, &data, channels, samples, epsilon);
pipeline.add(PipelineStage::SpatialWhitening(white));

Spec

The fit runs before the pipeline, on the signal of the nodes before it: fit_s seconds, in chunks pieces spread over the recording.

{"op": "whiten"}

App

Reference → Whiten. The fit runs in the background, and the raw signal shows until it’s done.

Python also has Whitening.fit_local_knn, which whitens each channel over its nearest channels, as Kilosort4 does.

Worked examples: Referencing, Spatial.