✦ Day 07 · Automated Component Labeling

ICLabel
Automated Classification

Replace subjective manual component inspection with a multi-branch deep convolutional neural network trained on over 200,000 components to classify ICs into 7 physiological categories with calibrated probabilities.

🤖 Deep Learning CNN 🏷 7-Class Taxonomy 🎓 SCCN Diagnostic Rules 📊 Calibrated Probabilities 🐍 mne-icalabel
1

ICA Applied to EEG & Component Anatomy
Forward Scalp Topographies vs. Activation Patterns

On Day 6, we discovered how ICA separates mixed electrode signals into independent sources $\mathbf{s}(t) = \mathbf{W}\,\mathbf{x}(t)$. Why does this model work so exceptionally well on human EEG?

1. Instantaneous Volume Conduction: Electromagnetic fields travel through brain tissue and the skull at the speed of light. Across scalp distances ($<20$ cm), transmission delay is under $0.0000003$ ms, satisfying the instantaneous linear mixing model: $\mathbf{x}(t) = \mathbf{A}\,\mathbf{s}(t)$.

2. Spatially Fixed Generators: Cortical dipole patches and artifact sources (ocular dipoles, temporalis muscles, heartbeat) remain at fixed spatial locations during recording. Each column of the mixing matrix $\mathbf{A}$ defines the fixed scalp topography (spatial projection) of that component.

Interactive Component Anatomy Explorer

Select any of the 7 physiological component types to explore its distinctive Scalp Map Topography, Time-Series Activation waveform, and Power Spectral Density (PSD).

7-Class Taxonomy
🧠 Brain
💪 Muscle
👁 Eye Blink
❤ Heart (ECG)
⚡ Line Noise
🔌 Chan Noise
❓ Other
FOCAL DIPOLAR MAP

🧠 Neural Brain Component

True cortical source generated by synchronized postsynaptic potentials across pyramidal neurons in cerebral cortex.

  • Spatial Topography: Dipolar or focal scalp distribution conforming to a single equivalent current dipole inside brain volume.
  • Power Spectrum: Characteristic 1/f spectral decay with distinct physiological oscillation peaks (Alpha 8–12 Hz, Theta 4–8 Hz).
  • Time Course: Continuous rhythmic oscillations or event-locked fluctuations without sudden high-voltage spikes.
2

ICLabel: Automated Deep Learning Classifier
Replacing Subjectivity with Calibrated Probability Estimates

Manual classification of ICA components is notoriously slow, fatiguing, and subjective: two expert electrophysiologists disagree on up to 20–30% of ambiguous components.

ICLabel (Pion-Tonachini et al., 2019, NeuroImage) solves this by utilizing a multi-branch deep convolutional neural network trained on over 200,000 expert-labeled components across 6,000+ EEG recordings.

The 6 Input Feature Streams

ICLabel extracts six complementary representations from every independent component before classification:

1. Scalp Topography Image

Interpolated 32×32 pixel head mesh capturing spatial gradient and dipole location.

2. Power Spectrum (PSD)

Normalized log-power vs log-frequency (1–100 Hz) capturing $1/f$ vs muscle slopes.

3. Autocorrelation (ACF)

Temporal self-similarity across time lags to detect persistent oscillatory rhythms.

4. Dipole Fit Residual

Residual variance (RV) of an equivalent single current dipole inside the brain volume.

5. Spatial Smoothness

Focal channel contribution ratio to detect bad single-channel noise.

6. Time-Domain Moments

Kurtosis and zero-crossing rates capturing bursty muscle contractions.

ICLabel Probability & Decision Boundary Playground

Adjust the predicted class probabilities $\mathbf{p} \in [0,1]^7$ (sum locked to 100%) to observe automated classification labels, confidence grading, and decision advice across research pipelines.

Classifier Sandbox
Lab Presets:
Predicted Class: Brain
High Confidence (82%)
🔬 Pipeline Rejection Strategy

Balanced Rule (Default ICLabel): Marks components for rejection if top predicted class is an artifact with $p > 0.50$. Suitable for standard cognitive research and general spectral analysis.

Pipeline Decision Verdict:
Keep Component: Dominant Brain probability ($>70\%$) with negligible artifact contamination. Safe for ERP and spectral analyses.

3

SCCN Expert Diagnostic Heuristics
Visual Triangulation & Edge-Case Classification (UCSD SCCN Guidelines)

Before deep neural networks, electrophysiologists at the Swartz Center for Computational Neuroscience (SCCN, UC San Diego) established rigorous visual heuristics to classify independent components. By triangulating four complementary representations—Scalp Topography, Equivalent Dipole (ECD) Fit, Power Spectrum (PSD), and ERP Images—experts achieve reliable distinction across ambiguous waveforms.

🗺️

1. Topography & Dipole Fit

Spatial Model

Synchronized cortical pyramidal patches generate dipolar electromagnetic fields modeled by an Equivalent Current Dipole (ECD) inside the brain volume (residual variance $RV < 15\%$). Peripheral focal edge maps indicate cranial muscle EMG. Far-field linear gradients indicate cardiac ECG.

📊

2. Power Spectrum (PSD)

Frequency Curve

True brain sources show characteristic $1/f$ spectral rolloff with distinct physiological oscillation peaks (Alpha ~10 Hz, Theta ~4–8 Hz). Muscle EMG exhibits high broadband power plateauing or rising above 20 Hz. Line noise shows an ultra-sharp delta spike at 50/60 Hz.

📈

3. Activation Dynamics s(t)

Time Series

Brain components exhibit continuous, rhythmic oscillations without sudden voltage jumps. Ocular blinks produce high-amplitude solitary pulses every 2–5s. Saccades show step-function intervals of fixation. Muscle shows bursty, high-kurtosis hash.

🖼️

4. ERP Image (Epoched Data)

Trial Consistency

In trial-sorted 2D ERP images, genuine cognitive components demonstrate consistent phase-locking or latency-aligned potential bands across experimental trials, whereas non-brain artifacts show non-stationary horizontal sweeps or random high-voltage outliers.

The 5 Golden Rules & Edge Cases from the SCCN Tutorial

Rule 1: The High IC Index Falloff (IC > N/2)

Because ICA sorts components in decreasing order of explained data variance, higher-numbered components (e.g., IC 70 of 128) have an extremely low probability of being meaningful neural sources. Even if a weak 10 Hz bump appears in the PSD, components with high index numbers and splotchy scalp maps are overwhelmingly classified as Other / unmixed residual noise.

Rule 2: Dual / Bilateral Cortical Dipole Co-activation

When two distinct cortical patches synchronize (e.g. bilateral primary auditory A1 or occipital visual V1 cortices), ICA merges them into a single component. Fitting a single dipole may yield elevated residual variance ($RV > 15\%$), but fitting a symmetric dual-dipole model resolves the source with high accuracy.

Rule 3: Vertical Blinks vs. Horizontal Saccadic Step Functions

Vertical Eye Movement: Monopolar frontal scalp map (affecting Fp1, Fp2, AFz equally) dominated by impulsive solitary spikes.
Horizontal Eye Movement: Bipolar lateral scalp map (positive potential on one hemisphere, negative on the other) characterized by step-like intervals of visual fixation separated by rapid saccadic transitions.

Rule 4: Notch Filter Dips vs. True Line Noise Spikes

Many preprocessing pipelines apply a notch filter at 50 Hz or 60 Hz. This creates a sharp downward notch in the PSD curve, which is a filtering artifact. A true line noise independent component exhibits an ultra-narrow upward delta spike towering 20–50 dB above the spectral floor.

Rule 5: Shallow Peripheral Dipoles = Cranial Musculature (EMG)

Because temporalis, frontalis, and neck muscle motor units reside outside the skull, their fitted equivalent dipoles localize outside the brain volume. The tighter and more concentrated the scalp focal point on the outer montage boundary, the shallower the source.

Interactive SCCN Expert Diagnostic Matcher

Select the visual features observed in an independent component to discover its SCCN classification verdict and biophysical rationale.

SCCN Rule Engine
Predicted SCCN Class

🧠 Neural Brain Component

High Confidence

Dipolar scalp topography with low residual variance ($RV < 15\%$), characteristic 1/f spectral decay with an alpha peak, and trial-locked ERP activity indicate a true synchronous cortical source.

4

End-to-End Pipeline & Code Reference
MNE-Python (mne-icalabel)

Here is the standard industry workflow for applying ICA + ICLabel in research pipelines: 

import mne
from mne.preprocessing import ICA
from mne_icalabel import label_components

# 1. Load continuous raw EEG
raw = mne.io.read_raw_edf('sub-01_task-rest_eeg.edf', preload=True)

# 2. Filter data: 1.0 Hz highpass is critical for clean ICA separation
raw_filt = raw.copy().filter(l_freq=1.0, h_freq=100.0)
raw_filt.set_eeg_reference('average')

# 3. Determine rank and fit FastICA
rank = mne.compute_rank(raw_filt, rank='info')['eeg']
ica = ICA(n_components=rank - 1, method='fastica', random_state=42)
ica.fit(raw_filt, picks='eeg')

# 4. Predict labels using ICLabel neural network
ic_results = label_components(raw_filt, ica, method='iclabel')
labels = ic_results['labels']             # List of predicted strings
probs = ic_results['y_pred_proba']        # (n_components, 7) probability matrix

# 5. Automatically mark non-brain artifact components for rejection
exclude_idx = [
    i for i, (lbl, p) in enumerate(zip(labels, probs))
    if lbl not in ('brain', 'other') and max(p) > 0.50
]
ica.exclude = exclude_idx
print(f"Rejected {len(exclude_idx)} artifact components: {exclude_idx}")

# 6. Apply unmixing & reconstruction to the ORIGINAL unfiltered raw dataset
raw_clean = ica.apply(raw.copy())

💡 Neuroscientific Deep Dive: Why 1.0 Hz Highpass Filtering for ICA?

Many researchers instinctively filter raw EEG at $0.1$ Hz to preserve ultra-slow cortical potentials. However, fitting ICA directly on $0.1$ Hz data severely undermines component separation. Why?

  • ✦ Non-Stationary Baseline Drifts ($>100\,\mu\text{V}$): Slow sweat potentials, galvanic skin impedance shifts, and electrode polarization create massive low-frequency drifts that violate the stationarity assumption of ICA.
  • ✦ Gradient Monopolization: In infomax or FastICA, high-amplitude drifts dominate the likelihood gradient, forcing ICA to waste multiple degrees of freedom modeling baseline wander rather than unmixing cortical and EMG sources.
  • ✦ Empirical Proof (Winkler et al., 2015; Dimigen, 2020): Highpass filtering at $1.0\text{–}1.5$ Hz increases the number of near-dipolar brain ICs extracted by up to 300% compared to $0.1$ Hz.
Golden Rule: $\mathbf{W} \leftarrow \text{Fit ICA on temporary } 1.0\text{ Hz copy} \implies \mathbf{x}_{\text{clean}}(t) = \mathbf{A}_{-\text{ex}} \mathbf{W}_{-\text{ex}} \mathbf{x}_{\text{raw}}(t)$

⚠️ Average Reference & Rank Deficiency Traps

A common mathematical pitfall in automated pipelines is running ICA without accounting for rank deficiency:

  • ✦ Average Referencing ($N - 1$ Rank): Re-referencing to average subtracts the mean across all channels at every time point, introducing an exact linear dependency. A 64-channel montage now has a true matrix rank of 63.
  • ✦ Pre-ICA Bad Channel Interpolation Trap: If you interpolate $K$ bad electrodes before ICA, you introduce $K$ additional linear combinations. Matrix rank drops to $N - 1 - K$. Running ICA with $N$ components on rank-deficient data causes matrix inversion failure, producing phantom high-frequency noise ICs.
  • ✦ Correct Pipeline Order: Mark bad channels $\rightarrow$ Compute true rank $\rightarrow$ Fit ICA with $\text{rank}-1$ components $\rightarrow$ Reject artifacts with ICLabel $\rightarrow$ Reconstruct $\rightarrow$ Interpolate bad channels last!
📚

Further Reading & Resources

  1. ICLabel: An Automated Electroencephalographic Independent Component Classifier, Dataset, and Website (Pion-Tonachini et al., 2019) 📑
    The definitive NeuroImage paper detailing the multi-branch deep convolutional architecture, 6 input feature representations, and crowd-sourced validation on 200,000+ expert labels.
  2. SCCN ICLabel Tutorial: Telling Independent Components Apart (Swartz Center for Computational Neuroscience, UCSD) 🎓
    The definitive visual inspection tutorial by the original creators of ICLabel, illustrating dipole models, spectral signatures, vertical/horizontal eye movements, and high IC index falloff rules.
  3. mne-icalabel: Automated Component Labeling for MNE-Python (Li et al., 2022) 🐍
    Official documentation for mne-icalabel with automated prediction APIs, probability extraction, and MNE-Python pipeline integration guides.
  4. EEGLAB Plugin Documentation & Component Interpretation Guidelines 📖
    Comprehensive reference for interactive GUI inspection, probability threshold selection rules for clinical vs cognitive research, and script-based batch cleaning.
  5. Automated Quality Evaluation of EEG Components (Chaumon, Bishop, & Busch, 2015) 🔬
    Foundational comparative study evaluating automated statistical rules (SASICA) vs visual inspection for preserving genuine Event-Related Potential (ERP) amplitudes.
  6. On the Influence of High-Pass Filtering on ICA-Based Artifact Reduction in EEG (Winkler et al., 2015) 📑
    Groundbreaking experimental validation proving that 1.0–2.0 Hz highpass filtering maximizes the extraction of near-dipolar neural components while avoiding spatial distortion.