Gains:
- Ability to create artificial intelligence workflow in artifact removal, event detection and feature extraction in EEG recordings
- Ability to recognize the impact of EEG's low signal-to-noise ratio and high variability on verification
- Ability to explain that neurosignal AI outputs should be tested with the approval of a clinical neurophysiologist
Electroencephalography (EEG; English electroencephalography) is the recording of the electrical activity of the brain with electrodes placed on the scalp. EEG is used in a wide range of areas, from epilepsy diagnosis to sleep staging, from intensive care monitoring to brain-computer interfaces. But it is a much more challenging signal than an ECG: its amplitude is on the order of microvolts (one millionth of a volt), meaning it is almost as small as noise. This low signal-to-noise ratio (SNR for short; the relative strength of the useful signal to the noise) makes EEG both attractive and dangerous for artificial intelligence. In this unit, we will see the EEG workflow, the problem of artifacts, and why neurophysiologist approval is indispensable.
Let's state from the beginning: an EEG model may say "seizure is possible"; But epilepsy diagnosis and seizure interpretation belong to the clinical neurophysiologist. Artificial intelligence does not diagnose.
EEG Processing Chain
Step 1 — Registration and assembly. The EEG is multichannel (typically 19-256 electrodes). The placement of electrodes is done according to an international standard order (10-20 system). Signals are read relative to reference electrodes; Mounting choice affects interpretation.
Step 2 — Artifact extraction. Sources of EEG contaminants are abundant: blinking and eye movement, muscle activity (especially jaw and neck), heartbeat, sweating, electrode contact, and grid interference. These artifacts can mimic real brain waves. Artificial intelligence helps in cleaning up artifactual segments through methods such as flagging and independent component analysis.
Step 3 — Feature extraction. EEG is usually divided into frequency bands: delta (slow, deep sleep), theta, alpha (rest), beta (awake attention), and gamma. Band strengths, connectivity measures, and event-related responses are used as features.
Step 4 — Incident detection/classification. The model suggests events such as seizure activity, spike-wave discharge, or sleep phase. The output is timestamped marks and tags.
Step 5 — Expert evaluation. Signs are presented to the clinical neurophysiologist; The expert confirms, corrects or rejects. Comment is for the expert.
Why Is EEG Verification So Difficult?
Three reasons make the EEG particularly fragile. First, low SNR: the real signal is hidden in the noise, the model can easily mistake the noise for a pattern. Second, high variability: even the same person's EEG varies with alertness, medication, and age; The difference between individuals is even greater. Third, reference uncertainty: even experts do not always agree on the “gold standard” marking of events such as seizures (interobserver variability). When these three are combined, an EEG pattern that looks impressive in the laboratory easily collapses in the real clinic.
Three Mini Cases: By the Numbers
Case 1 — Blinking mimicked a seizure. A seizure detection model mistook rhythmic eye blinks for spikes and waves in frontal regions in a child patient, producing several false signals per minute. The specialist saw that the marks overlapped with the eye canals and separated the artifact. The model produced the sign, the expert extracted the reality.
Case 2 — Time savings in sleep staging. A sleep laboratory manually phased recordings in 30-second segments throughout the night; Approximately 2 hours per recording. AI pre-staging reduced this time to 35 minutes; the technician examined only the slices and transitions that the model left unclear. The specialist approved the final staging report.
Case 3 — Decline in external validation. A seizure detection model showed 90% sensitivity at the center where it was developed; It decreased to 68% in the data collected by a different hospital with a different electrode and recording device. The number of electrodes, sampling rate, and patient profile differed. The model could not be considered reliable until it had been revalidated in each environment in which it would be used.
Weak Prompt / Strong Prompt
Note: The following prompts are helpful in editing the EEG report/feature text; Raw EEG interpretation is the work of verified software and neurophysiologist.
Weak prompt:
Is there epilepsy in this EEG?[features]
Powerful prompt:
Your role: You are a neurosignal analysis assistant (YOU ARE NOT DIAGNOSIS). Evaluate the following anonymous EEG summary features: - Table the band strengths and event signals, interpret them, but do not diagnose. - For each event sign, "could there be artifact?" add column (eye/muscle/grid).- Write ambiguities clearly; add low SNR warning.- Finally list 3 points the neurophysiologist should confirm in the raw recording.Anonymous features:[band strengths, event timestamps, channel information]
Four Copiable Templates
1) Band strength summary:
Table the following EEG band powers (delta, theta, alpha, beta) and evaluate their plausibility with the awake state. Don't diagnose. Data: [values]
2) Artifact possibility check:
For each of these signs of occurrence, mark the likelihood of eye/muscle/network artifact as low/medium/high and write your rationale. Signs: [list]
3) Sleep staging QC:
Mark sudden/unlikely transitions (e.g. jump from REM to deep sleep) in the AI sleep staging output below; List slices that require technician control.Output: [staging]
4) External verification plan:
Write a draft external validation plan for an EEG event detection model: different number of devices/electrodes, interobserver agreement (kappa), SNR, and artifact tests. Mark individual validation points.
Model's Role: By Task Type
Task type
AI contribution
criticality
verification
Artifact marking
high
low
Expert review
Band/feature extraction
high
medium
physiological plausibility
Sleep staging preliminary work
high
medium
Expert approval
Pre-seizure detection
medium
high
Neurophysiologist + raw recording
diagnostic decision
limited
very high
Full expert approval
Tip: Before relying on a pattern signature in the EEG, check the temporal overlap of that signal with the eye and muscle channels. Much of the activity that appears to be a seizure is actually biological artifact.
Caution: When EEG models are trained on data collected from a small number of patients, they overfit (memorize) individual-specific patterns and crash in new patients. High laboratory performance is no guarantee of true clinical success; external validation is essential.
Common mistakes
- Mistaking the artifact for a brain signal. Eyeblinking and muscle activity mimic a seizure; Channel overlap should be checked.
- Ignoring low SNR. A weak signal in noise easily misleads the model.
- Verification in one center. Different device and electrode layout reduces performance.
- Ignoring interobserver variability. Model metrics are misleading if the reference itself is ambiguous.
- Eliminating the expert. Signs are support; Seizure and diagnosis interpretation belongs to the specialist.
In summary
- EEG is a multi-channel, microvolt signal with low SNR; It is extremely susceptible to artifacts.
- The workflow consists of assembly, artifact extraction, feature extraction, and event detection.
- Low SNR, high variability, and reference uncertainty make verification particularly difficult.
- Models should be revalidated in each device and population in which they will be used.
- Seizure and diagnostic interpretation belong to the clinical neurophysiologist; AI does not diagnose.
Application task
Select a hypothetical EEG event detection scenario (e.g. seizure pre-detection). List five event signatures that the model produces and ask for each one, “could this be an artifact?” fill in the column (eye/muscle/grid/real). Then write three items of an external validation plan for this model and clearly mark the point at which the expert makes the final decision.
checklist
- [ ] I know that EEG is a low SNR, multi-channel, and artifact-prone signal.
- [ ] I understood the assembly, artifact extraction, feature and event detection steps of the workflow.
- [ ] I can see that artifacts mimic events and apply channel overlap control.
- [ ] I understand why external validation is essential and I understand interobserver variability.
- [ ] I have internalized that the interpretation of seizures and diagnosis remains with the specialist.