Preprocessing decisions#

Preprocessing is where most EEG decoding results are silently won or lost. The defaults you accept — high-pass cutoff, line-noise notch, montage, reference, channel set, artifact policy — encode strong assumptions about the signal you intend to study. They also interact: a 0.5 Hz high-pass plus an average reference plus ICA-based blink removal is a different pipeline from a 1 Hz high-pass plus a mastoid reference plus no ICA, and the two can disagree about a class boundary.

This page does not prescribe a recipe. EEGDash is preprocessing-agnostic and forwards your data to MNE-Python [Gramfort et al., 2013] [1] and braindecode for transformation. What this page does is name the choices that matter, describe what flips when you change them, and point at the tutorial where each one is exercised. The framing follows Tips 4–7 of Cisotto & Chicco (2024) [2], which give the same advice in a clinical setting.

A short audit before you start#

Before you run a single filter() call, write down the answers to four questions. They constrain the rest of the pipeline.

What frequency band carries the signal you care about? ERPs live below 30 Hz; sensorimotor rhythms peak at 8–30 Hz; high-gamma coupling lives above 60 Hz. There is no universal default.
What is the dominant noise? 50/60 Hz mains, slow drift, sweat-driven 0.1 Hz baseline, EMG bursts, blink saccades. Each one has its own remedy.
What is the unit of analysis? Single trial, averaged ERP, time- frequency map, spectral feature. Different analyses tolerate different amounts of damage from filtering.
What is the comparison set? If you are reporting numbers next to another paper, your filter, reference, and montage should be either identical or explicitly justified as a deliberate change.

Filtering: high-pass and low-pass#

Filtering is necessary and lossy. High-pass filters remove slow drift but also distort late, slow ERP components if the cutoff is too high. Low-pass filters remove muscle and noise but also smear sharp transients.

A few rules of thumb:

For ERP work, a high-pass between 0.1 Hz and 0.5 Hz with a slow, zero-phase FIR is the safest default. Pushing to 1 Hz or above introduces measurable distortion in late slow components.
For oscillation work, a high-pass at 1 Hz is often acceptable and sometimes desirable; the exact value should be reported.
A notch filter at the line frequency (50 Hz in the EU, 60 Hz in the US) is almost always called for. Verify that the notch is not inside your analysis band.
A low-pass at 40–45 Hz is a sensible default for ERP and slow oscillation analyses; a low-pass at 100 Hz is appropriate for high-gamma work and is what HBN-style data are typically delivered at.

Tutorial How do I preprocess EEG and create model-ready windows? walks through this on a real recording, including the choice of filter order and zero-phase vs. causal filtering.

Resampling#

If your downstream model only consumes data up to, say, 40 Hz, there is no benefit to feeding it 1024 Hz samples. Downsampling to 256 Hz or even 128 Hz reduces memory and compute by a large factor. Always low-pass before downsampling to avoid aliasing; MNE’s raw.resample does this for you.

The flip side: never upsample EEG to mimic a higher rate. You are not adding information.

Re-referencing#

EEG is fundamentally relative. The reference electrode is part of the signal, and changing it can flip the polarity of certain components and re-distribute power across channels. Common choices:

Common average reference (CAR). Re-references each channel to the average of all channels. Sensible when you have a dense montage and care about spatial topographies. Sensitive to outlier channels; always run after bad-channel detection.
Mastoid or linked-mastoid reference. Common in ERP work because a stable, near-zero reference at the mastoid emphasises peri-vertex signals. Disqualifies CAR-based pipeline comparisons.
REST reference. A cortical-source-based virtual reference that attempts to approximate an infinity reference. Useful when comparing across labs that used different physical references.

The choice is part of your pipeline, not a preprocessing afterthought. Pernet et al. (2019) [3] require the reference to be declared in the BIDS sidecar; respect that.

Montage and channel selection#

A montage tells MNE the 3D position of every electrode. Without one, topographic plots, source localisation, and certain spatial filters (CSP, Laplacian) cannot run. EEGDash datasets ship a montage when the underlying BIDS sidecar provides one; otherwise you must set it manually with raw.set_montage("standard_1020") or a vendor-specific file.

Channel selection is the other side of the same coin. Dropping a flaky channel is fine; dropping half the array because “it makes the model faster” is not — your reference, montage, and any spatial filter will silently change. If you must reduce channels, do it once, document it, and apply it identically to train and test.

Artifact policy#

Eye blinks, saccades, jaw EMG, and movement spikes are present in every real EEG recording. Three families of mitigations exist, in roughly decreasing severity:

ICA decomposition. Decompose into independent components and remove the eye-blink and EMG components. Most expensive, most transparent.
Threshold-based rejection. Drop windows whose amplitude exceeds a threshold. Cheap; loses data.
Robust models. Train models that tolerate residual artifacts (e.g., temporal augmentation, dropout). Cheapest, weakest signal.

A well-known pattern is to combine (1) for blinks with (2) for residual muscle bursts. This is the default in many MNE-based pipelines and is what the EEGDash tutorial gallery demonstrates. The choice matters because aggressive rejection can remove the signal you wanted to study — for instance, motor-imagery training data with high muscle artifact.

Default parameters are not neutral#

Every default in this list is a position, not the absence of one. A half-sentence in the methods that says “standard preprocessing was applied” is unverifiable; a sentence that says “0.5–40 Hz FIR filter, common average reference, ICA-based blink removal, 2-second windows with 50% overlap” is reproducible.

When you write your tutorial, paper, or report, list these decisions explicitly. When you read someone else’s, look for them. If a result is sensitive to those defaults — and many EEG decoding results are — the list is the most important paragraph in the paper.