Signal Acquisition & Processing
Analog to digital, sampling, filtering, and SNR
From analog ink to digital samples
The original electroencephalograph wrote a continuous analog trace in ink, and filtering was performed by physical resistor-capacitor circuits before the pen ever moved. Modern systems are digital: the conditioned analog signal is converted to a stream of numbers by an analog-to-digital converter (ADC), after which all montage reformatting, filtering, and display are mathematical operations on stored samples. This is a profound advantage - the same record can be re-referenced, re-filtered, and re-montaged after the fact - but it introduces two irreversible decisions made at acquisition: the sampling rate and the anti-alias filter. Everything downstream is reversible; these two are not. The shift from analog to digital therefore did not abolish the need to get acquisition right - it concentrated the unforgiving decisions into the moment of digitization.
Two parameters define the fidelity of the conversion. Sampling rate (samples per second, Hz) sets the temporal and, by the theorem below, the spectral resolution. Amplitude resolution is set by the ADC's bit depth, which divides the input voltage range into discrete steps; modern EEG uses high-resolution converters (commonly 16 to 24 bits) so that quantization error sits far below the noise floor and the full dynamic range from low-amplitude fast activity to high-amplitude slow waves is captured without clipping. Clipping - flat-topping a waveform that exceeds the converter's range - is an irreversible distortion that destroys morphology and injects spurious high-frequency content, and it is one reason the input range and gain must be set to accommodate the largest expected transients.
The Nyquist theorem and aliasing
The Nyquist-Shannon sampling theorem states that a signal can be faithfully reconstructed only if it is sampled at more than twice the highest frequency it contains. The threshold equal to half the sampling rate is the Nyquist frequency. Sample a 70 Hz signal at 200 Hz (Nyquist = 100 Hz) and you are safe; sample it at 100 Hz (Nyquist = 50 Hz) and you are not. Practically, clinical EEG is commonly digitized at 256, 512, or 1024 Hz, comfortably above the cerebral frequencies of interest and leaving headroom for high-frequency analyses. The push toward studying high-frequency oscillations (HFOs) - ripples and fast ripples in the 80 to 500 Hz range, of intense interest as biomarkers of epileptogenic tissue - has driven intracranial sampling rates to 2000 Hz and beyond, because resolving a 500 Hz ripple demands a sampling rate above 1000 Hz at the absolute minimum and comfortably higher in practice.
When a frequency above Nyquist is sampled, it does not simply vanish - it is aliased, masquerading as a spurious lower frequency that is mathematically indistinguishable from a genuine slow rhythm. A 130 Hz component sampled at 200 Hz folds back to appear as a 70 Hz wave, and once aliased it can never be removed, because the false low frequency is now baked into the data. This is the cardinal, irreversible failure mode of digital acquisition: aliasing manufactures believable artifact in the very band you care about. The danger is insidious precisely because the aliased component looks entirely physiological - there is no telltale signature in the sampled data to flag it as false, which is why prevention at acquisition is the only defense.
Aliasing cannot be fixed in software after sampling. A fixed analog low-pass (anti-alias) filter must remove energy above the Nyquist frequency before digitization. This filter is built into the hardware, tied to the chosen sampling rate, and is the one filter you cannot relocate or undo later. Trust it; you cannot replace it. This is also why you cannot simply 'turn up' the displayed frequency range beyond what the original sampling rate and its anti-alias filter permit.
The clinical filter chain
After digitization, three adjustable filters shape the displayed trace. The high-pass filter (often labeled the low-frequency filter, set by a time constant) attenuates slow drifts - sweat artifact, electrode polarization, respiration - below a chosen corner, typically around 0.5 to 1 Hz for routine review. The low-pass filter (high-frequency filter), typically set near 70 Hz, attenuates muscle and high-frequency noise above the cerebral range. The notch filter is a narrow band-reject filter centered on the power-line frequency (50 or 60 Hz) used to suppress residual mains interference. Each of these is now a digital operation applied to stored samples, which means it is reversible and can be re-applied at different settings - a freedom that is both the great strength and the great temptation of digital review.
| Filter | Typical routine setting | Removes | Distortion risk if misused |
|---|---|---|---|
| High-pass (low-frequency) | ~0.5-1 Hz | Slow drift, sweat, electrode polarization | Too high a corner attenuates real delta and distorts slow transients (e.g., flattens triphasic/periodic morphology), and can introduce artifactual sharp deflections |
| Low-pass (high-frequency) | ~70 Hz | Muscle (EMG), environmental high-frequency noise | Too low a corner blunts spikes, rounding sharp transients and hiding fast activity and HFOs |
| Notch | 50 or 60 Hz | Power-line interference | Can attenuate or distort genuine activity near the notch (including some spike harmonics), ring, and is a crutch for fixable impedance problems |
Aggressive filtering does not reveal truth; it sculpts appearance. A high-pass set too aggressively can erase real slowing and even create artifactual sharpness; a low-pass set too low can blunt an epileptiform spike into a benign-looking bump. Always know the active filter settings before declaring a finding, and re-review at standard settings when morphology is in question. The notch filter in particular should be a last resort after impedance is corrected, not a substitute for clean electrode application.
Phase distortion and the artifacts filters create
Beyond attenuating amplitude, filters alter timing and can manufacture features that were never in the signal - a subtlety experts must understand to avoid being deceived. Conventional causal filters (those that use only past samples, as analog filters must) introduce a frequency-dependent phase shift or group delay, smearing the temporal relationships between frequency components. This matters acutely for any analysis that depends on precise latency or on the phase relationship between channels or bands; a causal high-pass filter can shift and distort the apparent onset of a transient. Offline, zero-phase filtering (applying the filter forward and then backward) eliminates net phase shift, but at a price: it is non-causal, so it spreads a sharp event symmetrically in both directions in time, and can therefore create the illusion of activity preceding the true event - a serious confound when timing onset.
A particularly treacherous phenomenon is filter ringing (the Gibbs phenomenon): a sharp transient passed through a steep filter, especially a narrow notch, can sprout oscillatory wiggles before and after the event that did not exist in the original. A steep notch filter applied to a spike can surround it with rhythmic-appearing oscillations that mimic a discharge or a fast rhythm. The lesson is that an over-engineered filter does not merely remove unwanted content; it can imprint its own dynamics onto sharp features. When a suspicious oscillation appears adjacent to a transient under heavy filtering, the disciplined response is to reduce the filtering and inspect the raw data - because the filter, not the brain, may have authored the finding.
When morphology, timing, or the reality of a finding is in question, return to the raw, minimally filtered signal at standard display settings. Filters reshape amplitude, shift phase, and can ring; automated displays can mislead. The unfiltered trace is the arbiter, and the expert's reflex is to consult it before committing to any interpretation built on heavily processed data.
The shaded −3 dB line marks each cutoff. Watch how the over-filtered band removes real slow and fast activity, while turning filters fully off lets drift and mains noise swamp the trace.
Signal-to-noise ratio and source separation
Every acquisition decision ultimately serves the signal-to-noise ratio (SNR) - the ratio of cerebral signal power to the power of everything else. SNR is improved at the source (low, balanced impedances; a quiet electrical environment; a cooperative patient state) far more cheaply than by post-hoc processing, because filtering that removes noise also removes overlapping signal. The most powerful SNR tool for time-locked phenomena is signal averaging: because brain responses are time-locked to a stimulus while noise is random, averaging N repetitions improves SNR by roughly the square root of N - the principle underlying evoked potentials and event-related responses extracted from a sea of background EEG. The square-root scaling is also a sobering constraint: quadrupling the number of trials only doubles the SNR, so very small responses demand very many repetitions.
Artifact rejection spans a hierarchy of methods. Simple approaches reject or interpolate contaminated epochs and apply targeted filtering. More sophisticated, source-separation methods - principal component analysis and especially independent component analysis (ICA) - decompose the multichannel record into statistically independent components, allowing stereotyped artifacts such as eye blinks, lateral eye movements, ECG, and some muscle activity to be identified and removed while the cerebral signal in other components is preserved. ICA works because these artifacts are statistically independent of brain activity and have distinctive, reproducible scalp topographies; the eye-blink component, for instance, has an unmistakable frontal field. The expert caution is that no algorithm is free: every rejection or correction step trades some genuine signal for cleanliness, ICA can blend brain and artifact into the same component when data are limited or noisy, and an over-cleaned record can lose the very transient that mattered or, worse, leave behind a processed artifact that mimics pathology.
The 2026 landscape adds a further layer: machine-learning and deep-learning tools for automated artifact removal, spike detection, and seizure detection are now widely deployed and genuinely useful for triage and for flagging events in long recordings. But they carry their own failure modes. They can hallucinate or miss findings on data unlike their training distribution, they often provide little insight into why they flagged a segment, and their false-positive and false-negative profiles must be understood rather than assumed. A network trained largely on adult scalp data may falter on neonatal or intracranial recordings; a seizure detector tuned for sensitivity will flag rhythmic artifact as ictal. These tools augment the reader; they do not replace the obligation to confirm against the raw trace.
The cognitive failure modes and Bayesian discipline
Beyond instrumentation, the most dangerous distortions in interpretation are cognitive. Automation bias is the tendency to over-trust an automated artifact-rejection, spike-detection, or seizure-detection output and to under-scrutinize the raw data behind it - a hazard that grows precisely as the algorithms get good enough to be trusted most of the time. Confirmation bias is the tendency to filter, re-montage, and re-review until the record shows the finding one already expected - a genuine danger given how dramatically settings reshape morphology, since with enough processing one can coax almost any record toward a desired appearance. These are not character flaws; they are predictable products of how expert pattern recognition interacts with flexible, powerful processing tools, and they must be countered by explicit discipline: inspect raw data, document settings, predefine what counts as a finding, and let the unfiltered signal arbitrate.
The corrective frame is Bayesian. Any finding must be weighed as a likelihood against a prior probability set by the clinical context. The same sharp transient carries very different posterior odds of being epileptiform in a young patient with stereotyped focal seizures than in an elderly patient referred for syncope, and a benign variant that mimics epileptiform activity should be favored when the prior for epilepsy is low. Acquisition and processing feed directly into this reasoning, because a heavily filtered or algorithm-flagged finding has a lower likelihood ratio than the same finding confirmed on raw data with a clean field and an unambiguous physiological context. Over-reading - assigning pathological significance to ambiguous or artifactual features - is a leading source of EEG misdiagnosis, with real consequences: a patient mislabeled as epileptic on the strength of an over-processed or over-interpreted trace may carry that diagnosis, and its treatment and life restrictions, for years. The processing chain of this module is, in the end, in service of the diagnostic reasoning that protects against exactly that error.
Treat automated outputs as hypotheses to confirm, not conclusions to accept. Anchor every interpretation in the pretest probability set by the clinical question. When a finding would change management, demand that it survive on the raw, standardly filtered trace - because the cost of a false-positive epilepsy diagnosis is measured in years of unnecessary treatment and altered lives, and no algorithm shoulders that responsibility for you.
1. A muscle component at 120 Hz is present in the analog signal, but the system samples at 200 Hz with no adequate anti-alias filter. What happens?
2. To extract a small, stimulus-locked evoked potential buried in background EEG, the most effective strategy is:
3. An automated seizure-detection algorithm flags a 30-second segment as ictal in an ICU recording. Before escalating treatment, the expert's first action is to: