An electroencephalogram captures electrical activity generated by neurons through electrodes placed on the scalp. But the signal that actually reaches those electrodes is rarely clean.
Beneath all of that noise, the actual neural signal researchers care about is often small and easily buried. Signal processing is the essential scientific discipline that enables the translation of physical data into interpretable formats for modern analysis and technology.
What is Signal Processing?
The discipline of signal processing encompasses the transformation, analysis, and interpretation of physical events as measurable data. By applying mathematical frameworks to capture data strings, researchers can effectively filter out ambient interference and focus on the parameters of primary interest.
This methodology is central to modern scientific inquiry across diverse fields, including the study of complex human biological functions.
How to Define Digital Signal Processing
Digital signal processing represents the application of discrete-time computations to signals that have been converted from their original analog states.
This refinement occurs through the use of specific algorithms and processors designed to manage computational loads efficiently. Scientists utilize this approach to isolate biological signals, which is vital for fields involving neuroscience where data integrity is paramount for research accuracy.
The Core Goal of Signal Processing
At the peak of its utility, signal processing strives to extract the maximum amount of meaningful information from measured data while minimizing the impact of noise. By optimizing signal-to-noise ratios, researchers recover reliable data points that would otherwise be obscured.
This objective drives the development of sophisticated diagnostic tools used in modern EEG practice.
Types of Signal Processing
Signal processing is broadly categorized based on the nature of the input signal and the methodology applied for its manipulation. The physical world exists in a continuous state, while electronic devices frequently rely on quantized chunks of data to perform operations.
Distinguishing between these categories is fundamental to selecting the appropriate mathematical lens for any given experimental setup.
Analog vs. Digital Signal Processing
Analog systems manipulate continuous electrical waveforms directly through circuit components, maintaining a high degree of fidelity without the overhead of quantization. While analog systems offer instantaneous responses, they lack the flexibility and programmability characteristic of digital architectures.
As standard 10-20 system protocols have transitioned, digital processing has become the primary standard for maintaining comparable data across various research settings.
What Is Digital Signal Processing? Linear vs. Non-Linear Systems
Digital signal processing employs algorithms to refine data into formats suitable for computational modeling, offering a controlled environment for signal adjustment.
Understanding whether a system behaves linearly is critical for predicting how the output will respond to varying inputs, as the superposition principle only holds true for linear operations. Non-linear systems, conversely, introduce complex transformations that can capture subtle dynamics often ignored by simpler models.
System Type | Input-Output Relationship | Primary Characteristic |
|---|---|---|
Linear | Constant Proportionality | Superposition principle applies |
Non-Linear | Variable Dynamics | Higher-order harmonic generation |
Time-Invariant | Fixed Structural Response | Temporal stability of parameters |
By characterizing these behaviors, technicians ensure that the mathematical model accurately reflects the actual dynamics of the input signal before applying diagnostic filters.
Key Concepts and Techniques in Signal Processing
Technical mastery in this field requires a deep understanding of data preparation and mathematical translation. Each step in the processing pipeline purposefully degrades or enhances specific signal components to prepare the final output for human interpretation or machine learning algorithms.
Sampling and Quantization
Sampling involves the capture of data points at specific intervals forming the basis for discrete representation. Quantization follows this step, mapping those discrete time points to specific numerical values representing the amplitude of the signal.
If these steps are not precisely calibrated, aliasing can occur, leading to significant distortions in the analyzed waveforms.
Filtering
Filtering serves the critical role of removing unwanted frequency components that do not belong to the intended signal. Depending on the band-pass requirements, filters might surgically remove high-frequency muscle noise or low-frequency drift.
This process essentially isolates the neural activity of interest from the broader environmental background interference, resulting in a cleaner visual trace.
Transformations (e.g., Fourier Transform)
Transformations shift data between different domains to reveal periodic structures hidden within temporal noise. Common methods employed in this process include:
Fourier Transforms to decompose signals into their constituent frequency components.
Wavelet analysis for localizing transient events in both time and frequency simultaneously.
Hilbert transforms for determining the instantaneous envelope and frequency of complex signals.
Z-transforms for studying discrete-time systems and stability.
These mathematical maps allow engineers to visualize data from perspectives that identify specific pathologies or patterns that remain entirely invisible in the standard time-domain representation.
What Is an EEG Preprocessing Pipeline?
A preprocessing pipeline is simply a standardized sequence of operations applied to continuous raw EEG data before any analysis begins. While the exact order and specific methods vary between labs and software packages, most pipelines share a common backbone:
Filtering out unwanted frequencies
Re-referencing the signal to a neutral electrical baseline
Identifying and handling artifacts
Slicing the continuous recording into short segments tied to specific events, a process called epoching
The order matters less than the fact that every single decision in this chain can shift the final result. A 2020 benchmark study led by Robbins et al. found that even studies claiming to follow established best-practice guidelines showed considerable variation in how those guidelines were actually implemented in code.
Similarly, a 2021 study by Bertazzoli et al. demonstrated that in transcranial magnetic stimulation combined with EEG (TMS-EEG) research, the choice of preprocessing pipeline had a marked effect on the resulting brain response waveforms, and even changed how reliable those results were when the same subjects were tested twice.
Filtering: How to Isolate Real Brain Signals
Raw EEG contains two broad categories of unwanted frequency content.
At the slow end, voltage drifts creep into the signal from sweat building up under an electrode or from the electrode itself shifting slightly against the scalp. At the fast end, muscle tension and electrical interference from nearby equipment introduce high-frequency noise that has nothing to do with neurons firing.
Filtering works by removing these frequency ranges, keeping only the band where genuine brain-generated signals are strongest.
The most basic and widely applied step is high-pass filtering, which removes the slow drifts below roughly 0.5 to 1 Hz. Depending on the type of analysis, researchers may also apply a low-pass filter or combine both into a band-pass filter, commonly restricting the data to the 1 to 40 Hz range when studying event-related potentials (ERPs), the brain's characteristic voltage responses to specific stimuli.
What makes filtering particularly notable is how consistently it helps. In 2023, Arnoud Delorme tested a wide range of automated preprocessing corrections and found that most either had no measurable effect or actively decreased data quality. High-pass filtering was one of only two steps that reliably improved results across the board.
Researchers from Harvard similarly treat filtering as a foundational, non-negotiable component of its HAPPE pipeline for high-artifact developmental data. Filtering, in short, is close to a safe default. The same cannot be said for every step that follows.
Re-Referencing: Setting the Electrical Zero Point
EEG never measures an absolute voltage. It measures a difference between two points, an active electrode and a reference point.
Raw recordings are typically stored relative to whatever reference electrode was physically used during collection, often a single site like the mastoid bone behind the ear or the Cz electrode at the top of the head, positions defined by the 10-20 System used to standardize electrode placement. That original reference point is somewhat arbitrary, so re-referencing mathematically recalculates every channel's voltage relative to a new, theoretically more neutral point, such as the average of all electrodes or the combined signal from both mastoids.
The importance of re-referencing stems from volume conduction, which causes the brain's electrical activity to diffuse through surrounding tissue and fluid, ensuring that a single neural event registers across nearly every scalp electrode. Because of this spread, changing the reference point can change the apparent shape, amplitude, and even statistical significance of the recorded activity, even though nothing about the underlying brain activity has changed.
This is a core concept underlying the design choices behind EEG montages, including approaches like the Common Average Reference or a Referential Montage, each of which frames the same underlying brain signal differently.
Bad Channel Detection and Interpolation
Across any multi-electrode recording, a handful of channels will inevitably produce poor signal. A channel might go flat due to a loose connection, or it might become excessively noisy from poor scalp contact. Left uncorrected, these bad channels can distort analyses that depend on comparing signal patterns across the full scalp.
The standard fix involves identifying these problem channels, temporarily removing them from the dataset, and then reconstructing their missing data through interpolation, essentially estimating what that electrode likely would have recorded based on the activity of its healthy neighbors. This restores a complete scalp map for later analysis steps without letting the contaminated channel skew the results.
This is the second step research Arnaud Delorme identified as consistently beneficial. Alongside high-pass filtering, bad channel interpolation was one of the only corrections tested that did not decrease data quality.
Combined with filtering, it formed the one pipeline configuration in that study that outperformed simple filtering alone. Among the many automated corrections evaluated, this pairing stands out as a comparatively safe and useful procedure.
The Artifact Problem in Raw EEG
Beyond electrode-level problems, EEG recordings are contaminated by biological artifacts generated by the participant's own body:
Eye blinks generate large deflections at frontal electrodes
Eye movements create step-like voltage shifts
Jaw and forehead muscle activity introduces high-frequency noise
Sudden movements cause abrupt, multi-channel voltage offsets
The traditional response to contaminated segments has been to simply throw them away, rejecting any trial or time window where an artifact appears. This works reasonably well when a participant sits still and artifacts are rare. It becomes far less workable in populations where artifacts are the norm rather than the exception.
The HAPPE stresses that as EEG studies move toward higher-density electrode arrays and longer recordings, manually inspecting and rejecting bad segments by hand becomes unsustainable, and manual judgment calls introduce their own subjectivity.
Leach et al. adds a concrete illustration of the stakes involved: pediatric EEG recordings routinely lose large amounts of usable data simply because movement and blink artifacts are so frequent in young children. When data is this precious, throwing away an entire trial because of one blink is an expensive solution.
Independent Component Analysis (ICA) for Artifact Removal
Independent Component Analysis, or ICA, offers a more surgical alternative to wholesale rejection.
Conceptually, ICA is a blind source separation technique. It takes the mixed signal recorded at each scalp electrode and mathematically unpacks it into a set of statistically independent components.
Some of these components correspond to recognizable artifact patterns, such as the signature of an eye blink or muscle tension. Others correspond to patterns of brain activity. Once the artifact-related components are identified, they can be removed and the remaining clean components projected back into channel data, ideally preserving genuine brain signals while discarding the contamination.
The obstacle is that manually inspecting each component to decide whether it reflects an eye blink or a piece of cortex is slow and inherently subjective. This gave rise to automated classifiers such as ADJUST, ICLabel, and MARA, designed to make that judgment call algorithmically and consistently.
The evidence on how well this works is mixed rather than uniformly positive.
Arnaud Delorme found that automated ICA-based rejection of eye and muscle components failed to reliably increase the percentage of statistically significant channels in ERP data, meaning it did not deliver a clear, dependable improvement.
Robbins et al. compared two ICA-based approaches, LARG and MARA, across 17 studies and found that while the overall structure of results looked similar, significant differences still emerged, particularly in low-frequency spectral features and in how much blink-related residue remained in the cleaned data.
On the more encouraging side, Leach’s team showed that a modified version of the ADJUST algorithm, tuned specifically for pediatric geodesic net recordings, improved classification accuracy and retained more usable trials compared to the original ADJUST algorithm or no ICA correction at all, and in some measures outperformed ICLabel for pediatric data specifically.
Taken together, ICA can be a genuinely useful tool for isolating and removing artifacts, but its performance is not guaranteed. It depends heavily on how the classification algorithm is tuned and on the population being studied.
Artifact Subspace Reconstruction (ASR) as an Alternative
Artifact Subspace Reconstruction, or ASR, takes a different route entirely. Rather than decomposing the signal into independent components, ASR works by comparing incoming data against statistics drawn from a clean reference period, automatically flagging and reconstructing portions of the signal that deviate too far from that clean baseline.
Robbins’ team directly compared two variations of ASR against ICA-based pipelines across the same set of 17 studies. The broad conclusion echoed the ICA comparison: overall result structures looked similar across methods, but meaningful differences still appeared, again concentrated in low-frequency spectral features.
ASR is therefore a legitimate alternative to ICA rather than a strictly superior replacement, and the same principle applies here as everywhere else in this pipeline: the specific method chosen shapes the specific outcome obtained.
Should You Reject Bad EEG Segments?
You should avoid rejecting entire segments, as evidence indicates that the resulting loss of statistical power often outweighs the benefits of noise removal.
Simple thresholding is a blunt tool that disproportionately discards usable data; thus, experts prefer more surgical methods like interpolation and component-based correction.
Epoching: Focusing on Events of Interest
Most EEG analyses, whether measuring ERPs or time-frequency patterns, are anchored to specific moments in time: the presentation of a stimulus, or a participant's response to it. Epoching is the step that cuts the long, continuous recording into short, equal-length windows surrounding each of these events, typically including a brief period before the event itself known as the baseline.
That pre-event baseline period plays a correcting role, helping account for slow voltage shifts that might otherwise be mistaken for a genuine response to the stimulus. But this correction step is not immune to the same caution that applies to re-referencing.
The aforementioned study by Arnaud Delorme found that advanced baseline removal methods significantly decreased performance in terms of ERP significance. The safer approach, based on this evidence, is a simpler baseline correction, typically just subtracting the average voltage measured during the pre-stimulus window, rather than reaching for more elaborate correction schemes.
Why Pipeline Choice Changes Your Results
Placed side by side, the comparative studies reviewed here make one point unmistakably clear: preprocessing pipelines are not interchangeable, even when they draw from the same general toolkit of methods.
For instance the Delorme’s study found that of all the pipeline variations tested across leading EEG software platforms, EEGLAB, FieldTrip, MNE, and Brainstorm, only one configuration, combining high-pass filtering with bad channel interpolation, significantly outperformed simple high-pass filtering by itself. Many other commonly used steps actively reduced data quality rather than improving it.
Meanwhile, Bertazzoli’s team found that in TMS-EEG research specifically, waveform amplitude and global mean field power varied considerably across four different pipelines applied to the identical dataset. Topographical patterns were strongly correlated across pipelines only at latencies beyond 100 milliseconds after the stimulus; at earlier latencies, correlation values ranged widely, from 0.2 to 0.9, depending on which pipeline was used. Test-retest reliability itself depended on which preprocessing approach was chosen.
Lastly, the Harvard study demonstrated a more encouraging pattern for a specific case: automated pipelines built for high-artifact, developmental data, like HAPPE, outperformed seven widely-used alternative processing approaches, removing more artifact content while simultaneously preserving equal or greater amounts of genuine EEG signal in nearly every comparison.
The consistent thread running through these studies is that no single pipeline configuration is universally optimal across every dataset, population, or research question. Understanding the reasoning behind each stage, rather than memorizing a fixed recipe, is what allows a researcher to make deliberate, defensible choices and to recognize when a given pipeline might be poorly suited to their specific data.
Why Preprocessing Decisions Shape What You See in EEG Data
Raw brain recordings arrive cluttered with blinks, muscle noise, and electrical hum that can easily bury the quieter signals researchers want to study. Cleaning this data is not a neutral cleanup step—the exact path you follow from filtering to artifact correction can meaningfully shift the final picture of brain activity. The evidence shows that only a handful of moves, like high-pass filtering and patching broken channels, reliably reveal clearer neural patterns, while many common steps can unexpectedly weaken the results you’re chasing.
Therefore, watching what each preprocessing choice does is far more powerful than following a one-size-fits-all recipe. Since different reasonable methods can steer studies in different directions, openly reporting your process becomes as important as the data itself.
Tuning your approach to match the specific group you’re working with, whether a sitting adult or a moving child, often outperforms using a general-purpose tool without adjustment. Building that kind of thoughtful, transparent pipeline lets you move from trusting software defaults to truly understanding what your signal is trying to tell you.
References
Robbins, K. A., Touryan, J., Mullen, T., Kothe, C., & Bigdely-Shamlo, N. (2020). How sensitive are EEG results to preprocessing methods: a benchmarking study. IEEE transactions on neural systems and rehabilitation engineering, 28(5), 1081-1090. https://doi.org/10.1109/TNSRE.2020.2980223
Bertazzoli, G., Esposito, R., Mutanen, T. P., Ferrari, C., Ilmoniemi, R. J., Miniussi, C., & Bortoletto, M. (2021). The impact of artifact removal approaches on TMS–EEG signal. NeuroImage, 239, 118272. https://doi.org/10.1016/j.neuroimage.2021.118272
Delorme A. (2023). EEG is better left alone. Scientific reports, 13(1), 2372. https://doi.org/10.1038/s41598-023-27528-0
Gabard-Durnam, L. J., Mendez Leal, A. S., Wilkinson, C. L., & Levin, A. R. (2018). The Harvard Automated Processing Pipeline for Electroencephalography (HAPPE): standardized processing software for developmental and high-artifact data. Frontiers in neuroscience, 12, 97. https://doi.org/10.3389/fnins.2018.00097
Leach, S. C., Morales, S., Bowers, M. E., Buzzell, G. A., Debnath, R., Beall, D., & Fox, N. A. (2020). Adjusting ADJUST: Optimizing the ADJUST algorithm for pediatric data using geodesic nets. Psychophysiology, 57(8), e13566. https://doi.org/10.1111/psyp.13566
Frequently Asked Questions
How does signal processing differ from standard data analysis?
Signal processing focuses specifically on the temporal and frequency dynamics of continuous inputs, whereas general data analysis often deals with static or independent variables.
Can signal processing remove all forms of noise from a signal?
While filtering can significantly improve signal clarity, perfect noise removal is typically impossible due to the unpredictable overlaps between signal and ambient interference.
Why is the sampling rate important in digital systems?
The sampling rate determines the maximum frequency the system can accurately represent, which is essential for capturing high-frequency neural indicators without distortion.
What is the purpose of a transformation like the Fourier Transform?
Transformations are used to convert data from the time domain, where sequence is important, to the frequency domain, where individual component frequencies are emphasized.
How do linear and non-linear systems affect output precision?
Linear systems offer predictable, proportional responses, while non-linear systems provide complex capabilities that reveal hidden dynamics at the cost of simpler predictability.
Why does raw EEG data require preprocessing before analysis?
Raw EEG recordings contain not only brain activity but also interference from eye blinks, muscle tension, electrical noise, and electrode drift. Preprocessing isolates the genuine neural signal by removing these artifacts, making the data interpretable and usable for scientific study.
What is the one preprocessing step that almost always helps?
High-pass filtering, which removes slow voltage drifts from sweat or electrode movement, is consistently beneficial. Across multiple studies, it was one of the few steps that reliably improved data quality when applied early in the pipeline.
What does re-referencing do to the EEG signal?
Re-referencing mathematically recalculates every channel’s voltage relative to a new neutral point, such as the average of all electrodes, instead of the original recording reference. Since EEG measures voltage differences, this choice can reshape the apparent amplitude and even the statistical significance of brain responses.
Why are bad channels detected and how are they fixed?
Some electrodes inevitably lose good contact or become excessively noisy, distorting scalp-wide analyses. After identifying these bad channels, interpolation reconstructs their signal by estimating what they would have recorded based on the activity of surrounding healthy electrodes.
What is Independent Component Analysis (ICA) used for in EEG?
ICA is a technique that mathematically sorts the mixed signal into independent components, some of which correspond to artifacts like eye blinks or muscle noise. Once artifact-related components are identified and removed, the remaining clean components can be projected back into channel data to preserve genuine brain activity.
How does Artifact Subspace Reconstruction (ASR) differ from ICA?
ASR identifies and corrects noisy segments by comparing incoming data to a clean reference period, automatically flagging portions that deviate too far from that baseline. Unlike ICA, it does not decompose the signal into components but instead rebuilds contaminated sections based on the clean data’s statistics.
Why shouldn’t you simply throw out every segment that contains noise?
Bluntly rejecting noisy trials often discards large amounts of usable data, reducing statistical power and potentially weakening results. Since precious data is lost, methods like interpolation or component-based corrections are generally preferred, especially when recording time is limited.
What is epoching and why does it include a pre-event baseline?
Epoching slices the continuous recording into short time windows aligned to specific events, like a stimulus, to analyze brain responses. The pre-event baseline period provides a reference voltage level, so that slow drifts are not mistaken for a response to the stimulus.
Does the specific preprocessing pipeline actually change the final results?
Yes, different pipeline choices—such as filtering, reference scheme, or artifact correction method—can alter waveform shapes, significance, and even test-retest reliability. No single pipeline works best for all situations, so researchers must understand each step’s impact rather than blindly following a default recipe.
Emotiv is a neurotechnology leader helping advance neuroscience research through accessible EEG and brain data tools.
Christian Burgos
