eeg-signal-processing

Overview

eeg-signal-processing covers the full EEG analysis pipeline: from raw multichannel recordings to interpretable neural measures. It emphasizes principled preprocessing (filtering, re-referencing, artifact rejection), the major analysis families (ERPs, time-frequency, connectivity), and the statistical discipline that noisy, high-dimensional EEG data demands.

EEG has millisecond temporal resolution but poor spatial resolution and a terrible signal-to-noise ratio — most of the work is removing everything that is not brain signal while proving you did not remove the brain signal too.

When to use

• Designing an EEG experiment: sampling rate, electrode montage, reference choice, trial counts.
• Preprocessing: filtering, bad-channel handling, re-referencing, ICA-based artifact removal.
• ERP analysis: epoching, baseline correction, component measurement (N170, P300, N400, etc.).
• Time-frequency analysis: ERSP, inter-trial coherence, band-power (alpha, beta, gamma) dynamics.
• Connectivity: coherence, phase-locking value, Granger causality — with volume-conduction caveats.
• Group statistics: cluster-based permutation tests for the multiple-comparisons problem.
• BCI or real-time pipelines: low-latency preprocessing trade-offs.

Core concepts

• Sampling and filtering. Sample at ≥2x your highest frequency of interest (practically 250-1000 Hz). High-pass at 0.1-1 Hz (higher risks distorting slow ERPs), low-pass at 30-100 Hz depending on the question. Use zero-phase (filtfilt) filters; filter continuous data before epoching to avoid edge artifacts.
• Referencing. The reference defines every voltage you measure. Average reference is common for high-density caps; mastoid/linked-ears for ERPs; REST/rREST for reference-free estimates. Never compare amplitudes across studies with different references.
• Artifacts and ICA. Eye blinks, saccades, muscle, line noise, heartbeats. ICA separates statistically independent sources — remove components that are clearly artifactual (frontal topography + blink time course), but verify removal did not flatten genuine ERPs.
• Bad channels. Interpolate (spherical spline) rather than delete when possible; >10-15% bad channels means the recording is suspect. Report interpolation counts per subject.
• ERPs. Time-locked averaging: epoch (-200 to 800 ms typical), baseline-correct (pre-stimulus), average by condition. Components are defined by polarity, latency, and topography — a "P300" that peaks at 200 ms at Oz is probably not a P300.
• Time-frequency. Morlet wavelets or multitapers: power changes (ERD/ERS) and phase consistency (ITC/PLV). Baseline-normalize (dB or percent change); choose cycles to trade time vs frequency resolution.
• The multiple-comparisons problem. Thousands of time points × channels × frequencies. Use cluster-based permutation tests (Maris & Oostenveld) rather than uncorrected t-tests at every sample — the latter guarantees false positives.
• Volume conduction. Nearby electrodes share signal through the skull/scalp. Connectivity measures must use phase-lag-based metrics (PLI, wPLI, imaginary coherence) or source-space analysis; raw coherence between neighbors is mostly conduction.

Practical workflow

1. Inspect raw data. Plot continuous traces; note bad channels, drifts, line noise, movement bursts. Decide exclusion criteria before preprocessing.
2. Filter. High-pass 0.5 Hz (ICA benefits from 1 Hz), low-pass 40-100 Hz, notch at line frequency only if needed (notch filters ring; prefer CleanLine-style approaches).
3. Clean. Interpolate bad channels → re-reference → run ICA (on 1-Hz-filtered copy, apply weights to 0.5-Hz data) → remove artifact components → epoch → reject residual bad epochs (amplitude/step criteria).
4. Analyze. ERPs: average and measure mean amplitude in a priori windows. Time-frequency: wavelets, baseline-normalized. Keep analysis choices preregistered where possible.
5. Statistics. Cluster-based permutation tests across the full time×channel×frequency space; report cluster p-values, not cherry-picked electrodes.
6. Report. Preprocessing steps with parameters, trial counts retained per condition, reference scheme, and exact statistical tests. Share code and (anonymized) data when possible.

Example (MNE-Python sketch):

raw.filter(l_freq=0.5, h_freq=40)
ica = mne.preprocessing.ICA(n_components=20).fit(raw)
ica.exclude = [0, 3]  # identified blink/muscle components
epochs = mne.Epochs(raw, events, tmin=-0.2, tmax=0.8, baseline=(-0.2, 0))
evoked = epochs["target"].average()

Common pitfalls

• Filtering epoched data (edge artifacts) instead of continuous data.
• Double-dipping: selecting electrodes/time windows from the same data you then test.
• Reporting uncorrected p-values across hundreds of tests.
• Confusing muscle artifact (broadband, high-frequency) with gamma-band "cognition."
• Over-aggressive ICA removing genuine neural components along with blinks.
• Baseline-correcting with a baseline that contains condition differences.
• Comparing ERP amplitudes across different reference schemes.