ICLabel Tutorial: EEG Independent Component Labeling

Telling Components Apart

Brain Component

Brain components are thought to be generated by patches of cortex across which local field activity becomes spatially coherent. The electrical field of this patch can usually be modeled accurately by an "equivalent current dipole" (ECD), which is like a tiny, electrical version of a bar magnet. This means the electrical field produced by the component would create a positive potential on one side of the ECD and a negative potential on the other. Therefore, fitting a theoretical dipole to the scalp topography should result in a good fit. Some components are better described by 2 dipoles (meaning two distinct patches of cortex are synchronized), and so the one dipole fit shown in the images may have higher residual variance than expected.

While brain activity exists at frequencies that surpass 200 Hz, low frequencies are typically the only ones synchronous enough to show up in EEG. Therefore, brain components tend to have diminishing power at higher frequencies. Additionally, brain components tend to have repeated patterns at certain frequencies, leading to a peak in the power spectrum. These peaks are often found between 5 and 30 Hz, with 10Hz (termed alpha) being the most common.

As far as the component time series go, it's difficult to know what to expect unless the dataset is epoched (i.e. repeated sections of the experiment have been preselected). If the data is epoched, you can look for an event related potential (ERP). You can tell if a component comes from an epoched dataset by reading the label above the ERP Image.The best way to see if there is an ERP is by looking at he ERP-Image plot on the top right of the image. In the lower section of that plot, is an average of all trials.


  • Scalp topography often looks dipolar
  • Residual variance from dipole fit (marked RV on images) should be low. Usually below 15% unless the component is better explained with two dipoles
  • Power spectrum decreases as frequency increased (1/f)
  • Power spectrum usually has peaks between 5 and 30 Hz, most often at 10 Hz
  • Epoched data will likely have a visible ERP


This component is clearly fit by a single dipole located in the brain, both by the scalp topography and by the dipole plot. The power spectrum also suggests a brain source because of the peak at 10 Hz. This is continuous data, so there is no ERP (although there needn't always be one either).

This component is also well fit by a single dipole located in the brain, but is deeper in the brain than would be expected (visible in the dipole plot). Component depth is the least accurate part of dipole location estimates because of the dependence on accurate tissue conductivity values. The power spectrum also suggests a brain source because of the peak at 7 Hz.

This component is not as "clean" as the first two, but is still clearly a brain component as seen from the strong 10 Hz peak, 9% residual variance of the single dipole model, and roughly dipolar scalp topography. As the data is epoched, you can also see a clear event related potential in the ERP image.

Special Case: The scalp topography of this component and high residual variances clearly indicate that this is not a brain component. However, the strong ERP visible in the ERP Image and the powerful 10 Hz peak in the power spectrum are strongly indicative of a brain component. Therefore it could be correct to mark both "Brain" and "Other" for this component as they are both highly likely to apply. Likewise, "Other" alone is also a reasonable label as it is quite odd. As a side note, a possible explanation for this component may be that the channel location information is incorrectly registered in this dataset.

Eye Component

Eye components describe eye motion. Each retina (the part of the eye that registers incoming light) creates an electric field which can be effectively modeled as an "equivalent current dipole" (ECD). Typically, eye motion is split into two components: vertical movement and horizontal movement. Other eye components can also be found, such as diagonal directions, but they are rare and depend on the experiment. For all eye components, the power spectrum will vary due to experiments and people, but generally most of the power will reside at frequencies below 5 Hz as people do not usually move their eye faster than that.

Vertical movement has a scalp topography that can be modeled with two ECDs, on in each eye, that are oriented up down and there for only look positive or only negative as EEG recordings rarely record from the underside of the head. Due to nearness of the eyes and the strength of the ECDs, one ECD can usually fit this type of component very well. The component activity should show clear spikes relatively frequently due to eye-blinks. There can also be some drift due to looking up or down.

Horizontal eye movement has a scalp topography that can be modeled an ECD placed between the eyes and oriented left/right with positive values on one side and negative on the other. The component activity should have intervals of relative stability with occasional and very fast transitions to different values. Such patterns are cause by visual scanning. The periods of stability are times when visual focus is help somewhere and the fast transitions are saccades.


  • Scalp topographies suggest ECDs near eyes
  • Power concentrated at low frequencies (below 5 Hz)
  • Vertical eye movement components will contain blinks in the data
  • Horizontal eye movement components will look like step functions

This components captures the effects of eye blinks. This is most clearly visible in the time series plot but can also be seen in the ERP Image. The scalp topography shows that the component affects the electrodes around the eyes, roughly equally. More evidence for this is the high data variance accounted for and the lack of peaks in the power spectrum. Two dipoles fit the scalp topography significantly better than just one and they are located near the eyes. It is a bit odd that they appear above the eyes, but that could potentially be error in the electrical model.

This components captures the effects of eye blinks. This is primarily visible the time series data, power spectrum, and the scalp topography. The 2-dipole plot shows the dipoles placed above the eyes which could indicate that this is not an eye component, but in this case, more likely means that the electrical model as poorly aligned to the electrical model of the subject's head. Another oddity is the ERP Image which suffers from scaling issues caused by high amplitude artifacts, rendering it useless here. Finally, the offset of ~25 micro Volts in the time series is also odd. Nonetheless, this component can still be confidently classified as "Eye".

This components captures the effects of eye movement. In both the time series plot and the ERP Image, you can see clear evidence of saccades by the discontinuities surrounded by relative stationarity. The scalp topography and dipole plot reinforce this interpretation as they indicate that the source origin is in or near the eyes.

This components captures the effects of horizontal eye movement, although some high frequency power in included from some other source.

Muscle Component

Muscle components describe the electrical fields generated by muscle activity, known as electromyography (EMG).Their activations are powerful relative to EEG but motor unit action potentials (the underlying source of EMG) do not synchronize causing most of the power of EMG to be spread out among higher frequencies. Nonetheless, these components can still look dipolar, although they will seem very shallow as they are not localized within the brain. You can tell a shallow dipole by how concentrated its scalp topography is. The more concentrated, the shallower. That isn't to say that all muscle components will be dipolar.


  • Power concentrated in higher frequencies (20 Hz and above)
  • Can still be dipolar, but will be located outside the skull

Every plot in this image suggests a muscle component. Dipole plots and the scalp topography indicate a shallow source. The power spectrum has broad-band high frequency power. Lastly, the ERP image shows highly non-stationary activity, typical of changing muscle usage based on body position and activity.

This component is most readily recognized by the power spectrum as it has low amounts of low frequency power and high amounts of broad band high frequency power. The dipoles are also located very close to the outside of the skull. Even though the component has high residual variance, the scalp topography a pattern that can easily be interpreted to be a very shallow dipole.

This dataset was not detrended before ICA was applied, leading to the unusual looking ERP Image. Nonetheless, it can be comfortably classified as a muscle component from the power spectrum's high frequency power and the "shallow" scalp topography (also visible in the dipole plot).

As usual, the power spectrum is the leading evidence for a muscle component. In this case, though, the time series shows a district and strong muscle activity burst that is typical of EMG data.

Heart Component

Heart components capture the electrical potentials generated by the heart. This measurement is known as electrocardiography (ECG or EKG). The pattern generated by the heart is very typical and is known as a QRS complex . These should occur at about 1 HZ. because of the distance of the heart, the scalp map will look like that of a very far dipole and so will look almost like a linear gradient.


  • Clear QRS complex in the data at about 1 Hz
  • Near linear gradient scalp topography
  • No peaks in power spectrum

A regular QRS complex is immediately visible in the time series plot. The scalp topography is the a wide, roughly linear, gradient as is typical of heart components. It is not perfectly separated, as visible by the activity between the QRS complexes and the spurious peaks in the power spectrum, but it is resolved enough to comfortably call a heart component.

Linear scalp map gradient and a clear and regular QRS complex indicate a heart component.

Linear scalp map gradient and a clear and regular QRS complex indicate a heart component.

Line Noise

Line noise is contamination from the alternating current that is used to power nearly all lighting fixtures and electronics these days. Depending on where you are, the frequency used can be either 50 Hz or 60 Hz. The datasets we use come from all over and so you can expect to see some of both (but never together). While this noise is typically removed in the data cleaning steps before ICA is applied, ICA may separate out line noise if it has enough channels and data to work with. Line noise components are mostly evident from their sharply peaked power spectrum with the peak at either 50 or 60 Hz. These components are not to be confused with other components that simply have some line noise contaminating them.

Note: To avoid issues with this type of noise, many datasets are passed through a notch filter centered at either 50 Hz or 60 Hz. This will cause a sharp dip in the power spectrum at that frequency and is not indicative of any of these categories.


  • Strong peak in power spectrum at either 50Hz or 60Hz

The only real indicator here is that huge peak in the power spectrum at 50 Hz. Of interest, though, is that the source of the line noise is non-stationary as the effects are only present during certain segments of the data as can be seen in the ERP Image. The scalp topography is not informative beyond possibly allowing one to guess the direction of the line-noise source.

This component, instead, has a strong peak at 60 Hz. The line noise power is well above the rest of the power spectrum, even if it is lower than in the other examples.

This component has a peak at 50 Hz and is the most clearly non-stationary of all the examples.

Peak at 50 Hz.

Channel Noise

If a channel gets bumped a lot during a recording or if it has a poor contact, it will often generate large artifacts that do not affect any other channels. ICA will often separate these out into their own component which we call channel noise components. They can be recognized by their scalp topographies which put almost all the weighting on a single channel.

These components can be difficulty to classify as often times they can seem very similar to muscle components. The similarity resides mostly in the scalp topography. Telling them apart is primarily accomplished by looking for typical high frequency power typical of muscle components. Channel noise components, on the other hand, typically have a 1/f spectrum.


  • Very focal scalp topography
  • Large and/or consistent artifacts in the component activations.
  • Easily confused with muscle components, but PSD is different.

The scalp topography is only weighted on a single electrode and the power spectrum is a cleanly decreasing curve. Very typical channel noise component despite a lack of large artifacts in the data.

The scalp topography is only weighted on a single electrode and the power spectrum is a cleanly decreasing curve. Very typical channel noise component despite a lack of large artifacts in the data.

Very typical channel noise component. 1/f power spectrum and extremely focal scalp topography.


Not all components are meaningful. More specifically: most components are not meaningful. ICA assumes that there are as many independent components as there electrodes and that is almost never the case. When a component does not converge onto a meaningful signal, it can either capture a mixture of signals or some noise. In rough terms, anything that does not fit the above categories can be deemed "other". Signs of components being in the "other" category would be anything that stands out as very unusual or uncharacteristic of the other categories. Also, the high the "component number" (and therefore the lower the explained variance in the data), the more likely that a component is in the "Other" category. Usually the last half of components can be called "Other". For instance, even if there seems to be a 10 Hz peak in the power spectrum, the component is not necessarily a brain component. If it has a very non-dipolar scalp topography and there is very little


  • Anything that doesn't fit the above categories.
  • More likely the higher the IC number (as in IC 150 of 220 is very likely to be "Other"
  • Non-dipolar scalp maps
  • Spectrum can still have weak 10 Hz peak as brain signals are likely mixed with other signals

The two main features that stand out for this component are the splotchy scalp topography and that it is IC 50 of 120. Components with numbers this high have a very low chance of being meaningful. Note that this is an example of how ERPs do not always mean the component is meaningful.

The messy scalp topography in conjunction with a lack of peaks in the power spectrum suggest this component is classifiable as "other". The odd step in the power spectrum is likely do to a low pass filter at 25 Hz.

There are few meaningful clues to extract from the information presented asides from the non-dipolar scalp map. Therefore this component can be deemed other. The amount of 50 Hz power is not enough to deem this a purely "Line Noise" component, but it wouldn't wrong to mark both "Other" and "Line Noise" in this case.
Return To Labeling