scale

name	scale
description	Scale/normalize signal amplitude — robust, standard, or numeric factor
layer	L3
group	channel
metadata	{"tags":["operator","scale","normalize","standardize","robust"],"modalities":["eeg","seeg","ecog","meg","fnirs"],"step_string":"scale","analysis_goal_allowed":["classification","feature_extraction","exploratory","generic","online_inference"],"analysis_goal_forbidden":["source_localization"]}

Scale / Normalize

Function

Normalizes signal amplitude per channel. Three modes: robust scaling (median + IQR), standard scaling (mean + std), or multiplication by a numeric factor.

Parameter Format

scale:{method}

Examples:

• scale:robust — RobustScaler (median-centered, IQR-normalized)
• scale:standard — StandardScaler (zero-mean, unit-variance)
• scale:1e6 — Multiply by 1,000,000 (volts → microvolts)

Parameters

Parameter	Type	Options	Description
method	string or float	robust, standard, or numeric	Scaling method

Method Details

Method	Formula	Use Case
robust	(x - median) / IQR	Outlier-resistant; best for EEG with artifacts
standard	(x - mean) / std	When Gaussian assumption holds; ML default
{number}	x * factor	Unit conversion (e.g., V → uV)

Scaling is applied per channel (fit across time dimension).

When to Use

• Before feeding data to machine learning models (normalization required)
• When combining data from different amplifiers (different gain settings)
• Unit conversion for display or compatibility
• After all filtering steps (scaling before filter can cause numerical issues)

When NOT to Use

• Before ICA (ICA handles scale internally)
• If downstream analysis is scale-invariant (e.g., correlation-based methods)
• For spike rate data (already in meaningful units: spikes/s)

Ordering

• Apply as one of the last steps (after all filtering, resampling)
• Apply AFTER: notch, bandpass, ICA, resample
• Apply BEFORE: clip (clip operates on scaled values)

Recommended Parameters

Paradigm	Method	Rationale
Motor Imagery (CSP)	robust	CSP sensitive to outliers; robust handles artifact trials
P300/ERP	standard	ERP averaging benefits from zero-mean channels
Deep learning input	standard	Most DL models expect ~N(0,1) input
General	robust	Safer default for EEG data with potential outliers

Reference Code

Robust Scaling

from sklearn.preprocessing import RobustScaler
import numpy as np

# data shape: (n_channels, n_samples)
data = RobustScaler().fit_transform(data.T).T.astype(np.float32)

Standard Scaling

from sklearn.preprocessing import StandardScaler
import numpy as np

data = StandardScaler().fit_transform(data.T).T.astype(np.float32)

Factor Multiplication

# Convert volts to microvolts
data = data * 1e6

Key API

• RobustScaler() — uses median and IQR, resistant to outliers
• StandardScaler() — uses mean and std, assumes Gaussian
• .fit_transform(data.T).T — scale per channel (fit across time axis)