Skip to content

EEG - Electroencephalography

Electroencephalography (EEG) captures the electrical activity of the cerebral cortex using electrodes placed on the scalp, offering a millisecond-level window into ongoing brain dynamics.

EEG provides multivariate time series data: \(n\) electrode channels each sampled at rate \(SR\) (commonly 250–500 Hz). The number of channels \(n\) typically follows the 10-20 international system, with 32 channels being a common research standard. Consumer headsets use 5–14 channels.

Brain Anatomy Relevant to Cognitive Monitoring

The cerebral cortex is divided into the left hemisphere (logical reasoning) and right hemisphere (intuition/perception). Electrodes with odd suffixes are on the left; even suffixes are on the right; \(z\) suffix electrodes sit on the midline.

Human brain anatomy and lobe regions

Neural Substrate Brain Area Electrodes Role
FL Frontal Lobe Fp1, Fp2, F3, F4 Thinking, judgement, working memory
PFC Prefrontal Cortex Fp1, Fp2 Free-choice decision making
dlPFC Dorsolateral PFC F3, F4 Cognitive control, arithmetic, executive function
vlPFC Ventrolateral PFC F7, F8 Emotion regulation, cognitive reappraisal
MC Motor Cortex C3, C4, Cz Body movement control
IPC Inferior Parietal Cortex P7, P8 Emotion recognition
OL Occipital Lobe O1, O2 Visual processing
TL Temporal Lobe T7, T8 Hearing, memory, emotion
HC Hippocampus T7, T8 Memory storage and retrieval
AM Amygdala P7, P8 Emotion processing
FPN Frontoparietal Network Fp1, Fp2, F3, F4, F7, F8, P3, P4 Active problem-solving
DMN Default Mode Network Fz, Cz, Pz Mind wandering, passive rest

NASA-TLX Stressors and Their Neural Correlates

Because the primary downstream label is the NASA-TLX cognitive workload composite score, each of its six dimensions maps to specific brain regions and electrode clusters:

Stressor How It Works Neural Substrate Key Electrodes
Mental Demand Thinking, deciding, calculating, remembering PFC + Frontal FPN + HC F3, F4, Fp1, Fp2
Physical Demand Controlling, pulling, physical exertion Motor Cortex (MC) C3, C4, Cz
Temporal Demand Time pressure, pace stress vlPFC + HC F7, F8, T7, T8
Effort Combined mental + physical exertion dlPFC + dorsal FPN + MC F3, F4, C3, C4, Cz, P3, P4
Frustration Stress, irritation, insecurity AM + dlPFC + IPC + HC F3, F4, T7, T8, P7, P8
Performance Perceived success in achieving task goal FL/FPN (active) − DMN (deactivated) Fp1, Fp2, F3, F4, F7, F8, P3, P4

Signal Formalisation

Univariate EEG Signal

A single-channel recording \(\mathbf{x} \in \mathbb{R}^T\) is a sequence of \(T\) uniformly-spaced observations:

\[\mathbf{x} = [x_1, x_2, \dots, x_T]^\top\]

Multivariate EEG Time Series

A recording across \(d\) simultaneous channels:

\[X \in \mathbb{R}^{d \times T}, \quad X = \begin{bmatrix} \mathbf{x}^{(1)} \\ \mathbf{x}^{(2)} \\ \vdots \\ \mathbf{x}^{(d)} \end{bmatrix}\]

Tensor shape notation: (C, T) where C = number of channels, T = number of time steps.

EEG Classification Dataset

A labelled dataset \(\mathcal{D}\) consists of \(N\) windowed samples with corresponding labels:

\[\mathcal{X} = \{ X_1, X_2, \dots, X_N \}, \quad X_i \in \mathbb{R}^{d \times T_i}\]
\[\mathbf{y} = [y_1, y_2, \dots, y_N]^\top, \quad y_i \in \mathcal{C} = \{1, \dots, c\}\]

The channel dimension \(d\) is shared across all samples; the sequence length \(T_i\) may vary.

EEG Classifier

A classification model is a function:

\[\mathcal{M}: \mathbb{R}^{d \times T} \rightarrow [0, 1]^c\]

mapping a multivariate window to a class-probability vector. The predicted label is:

\[\hat{y} = \arg\max_{i \in \mathcal{C}} \hat{y}_i\]

Signal Properties

Understanding EEG's inherent properties is essential for designing appropriate preprocessing and modelling strategies:

Property Description Implication for Modelling
Low SNR Cortical signals (~µV) easily masked by noise Artefact rejection and filtering are critical
Apparent Stochasticity Appears random but driven by unmeasured latent factors Large datasets and strong regularisation needed
Nonstationarity Statistical properties shift over time Per-window normalisation; non-stationary models
Nonlinearity Linear models are insufficient Deep neural networks required
High Dimensionality Up to 128 simultaneous channels × hundreds of Hz Dimensionality reduction or patch-based tokenisation

EEG Frequency Bands

Band Frequency Range Primary Cognitive Associations
Delta (δ) 0.5 – 4 Hz Deep sleep; unconscious processing
Theta (θ) 4 – 8 Hz Working memory load; frontal midline theta increases with task difficulty
Alpha (α) 8 – 13 Hz Relaxed wakefulness; suppressed during cognitively demanding tasks
Beta (β) 13 – 30 Hz Active concentration; sensorimotor rhythms
Gamma (γ) 30 – 100 Hz Higher-order processing; perceptual binding

Alpha suppression as a workload indicator

The decrease in alpha power (event-related desynchronisation, ERD) during cognitively demanding tasks is one of the most robust and replicated workload indicators in the EEG literature. Occipital alpha suppression is particularly reliable for visual attention tasks.

Hardware

OpenBCI

The OpenBCI Cyton amplifier is used with flat snap and comb snap AgCl-coated electrodes. Flat and comb snap electrodes require no gel, suit real-time monitoring, and tolerate hair - at the cost of higher impedance and noise. Electrode caps use wet gel electrodes for higher signal quality and are reserved for controlled experimental sessions.

OpenBCI Cyton EEG amplifier

Flat and comb snap AgCl electrodes

Commercial Headsets

Headset Channels Sampling Rate Key Electrode Positions
Emotiv Epoc X 14 256 Hz AF3/4, F7/8, F3/4, FC5/6, T7/8, P7/8, O1/2
Emotiv Insight 5 128 Hz AF3/4, T7/8, Pz
Neuroelectrics Enobio 8 8 500 Hz FP1/2, AF7/8, P3/4, T9/10
Muse S Headband 5 256 Hz AF7/8, TP9/10, FpZ
Naoyun (in-ear) - - Targets temporal lobe via ear canal

Emotiv Epoc X headset

Muse S headband

Muse S in cognitive monitoring research

The UNIVERSE dataset uses the Muse S headband (5 channels, 256 Hz) for its 315-hour recording, making it one of the largest consumer-grade EEG cognitive datasets. The limited channel count and frontal/temporal coverage (AF7/8, TP9/10, FpZ) tests whether strong cognitive state prediction is achievable with minimal electrode coverage.