What a brain-computer interface actually is: from EEG signal to decision
Written by Yasin TANIŞ
A brain-computer interface, or BCI, is a system that measures activity of the nervous system and translates it into an output, such as a command, a selection or a control signal, without going through muscles. That definition is less exciting than most headlines, but far more useful, because it also says what a BCI is not: it does not read thoughts word by word, and it does not understand a mind. It recognises specific, measurable patterns defined in advance and links them to actions that have also been set in advance.
This article follows a single path through such a system: the one based on electroencephalography (EEG), the most common non-invasive method. From a voltage on the scalp to a decision a machine can act on.
Where the signal comes from
Neurons communicate through electrical and chemical signals. A single neuron’s activity is far too small to detect from outside the skull. What EEG picks up is the summed activity of very large populations of neurons in the cortex, especially the slow postsynaptic currents of pyramidal cells that are aligned in parallel and active at the same time. When enough neurons work in synchrony, the resulting field passes through brain tissue, fluid, skull and scalp, and arrives at the surface as a voltage of a few to roughly a hundred microvolts.
Two consequences follow. First, EEG has excellent timing: it can follow changes on the scale of milliseconds. Second, it has poor spatial detail. The skull smears the signal, and each electrode sees a mixture of sources from a broad area. A BCI designer works within both facts.
Recording: electrodes, amplifiers, numbers
Electrodes are placed on the scalp, usually following the international 10–20 system, a standard layout that names positions such as Cz at the top of the head or O1 and O2 over the visual cortex. Wet electrodes use a conductive gel to lower the impedance between skin and sensor; dry electrodes are faster to put on but more sensitive to movement and contact quality.
Because the signal is so small, the amplifier matters as much as the electrode. It measures each channel against a reference, suppresses interference that appears on all channels at once, and passes the result to an analogue-to-digital converter. Typical research systems sample between 250 and 1000 times per second per channel. From this point on, the brain signal is a stream of numbers.
Cleaning: noise comes with the job
Raw EEG is mostly not brain. Eye blinks and eye movements produce large, slow deflections at the front of the head. Jaw and neck muscles add fast, broadband activity. Mains electricity leaves a steady hum at 50 or 60 hertz. Loose electrodes produce drifts and jumps.
Processing therefore starts with cleaning: band-pass filtering to keep the frequency range of interest, notch filtering to remove mains interference, re-referencing, and removing or correcting artefacts. Methods such as independent component analysis can separate some artefact sources, like blinks, from brain activity. Some data is simply discarded. No algorithm later in the chain can recover information that was lost or corrupted here.
Finding a pattern worth using
A BCI does not look for arbitrary thoughts. It relies on brain responses that are known to be reliable and measurable from the scalp. Three main approaches have been studied for decades.
- P300 event-related potential. When a rare, relevant stimulus appears among frequent irrelevant ones, the brain produces a positive deflection roughly 300 milliseconds later. A speller flashes rows and columns of letters; the letter that produces the P300 is the one the person is attending to.
- Steady-state visually evoked potentials (SSVEP). Looking at a light that flickers at a fixed frequency produces activity at that frequency, and its harmonics, over the visual cortex. Several targets flickering at different rates become several possible commands.
- Motor imagery. Imagining a movement of the left or right hand changes the power of rhythms in the mu (about 8 to 12 hertz) and beta (about 13 to 30 hertz) bands over the opposite side of the motor cortex. These changes can be detected without any actual movement.
The broader EEG vocabulary helps here: delta below 4 hertz, theta from 4 to 8, alpha from about 8 to 13, beta from 13 to 30 and gamma above that. These bands are conventions for describing rhythms, not boxes that each hold a single function.
From pattern to decision
Once the data is clean, the signal is cut into epochs, short windows aligned to a stimulus or a cue. From each epoch the system extracts features: the amplitude at certain time points for a P300, the power at the flicker frequencies for SSVEP, or the band power after spatial filtering for motor imagery. Common spatial patterns (CSP) is a classic method for that last case.
Features go to a classifier. Linear discriminant analysis (LDA) remains a common and robust baseline; support vector machines, Riemannian geometry methods and neural networks are also used. The classifier is trained during a calibration session in which the person performs known tasks, so the system can learn what their particular signal looks like. Brain signals vary between people, and they drift within the same person from day to day, so calibration is not a one-time chore.
The output is a decision: this letter, that target, left rather than right. Performance is usually summarised as accuracy and as information transfer rate, measured in bits per minute. For non-invasive BCIs these rates are modest compared with a keyboard. That is an honest constraint, not a failure.
Closing the loop
A BCI becomes useful when its output returns to the person as feedback: a cursor moves, a letter appears, a device responds. People learn to produce clearer patterns when they can see the result, and adaptive systems can update their models as they go. The interface is then not a sensor attached to a passive user but a loop in which the person and the machine adjust to each other.
What BCIs are not
- They are not mind reading. A non-invasive BCI distinguishes a small set of trained patterns under controlled conditions.
- They are not all the same. Invasive systems, using electrodes on the surface of the brain (ECoG) or arrays within the cortex, record far richer signals but require surgery. Non-invasive EEG trades signal quality for safety and accessibility.
- They are not finished. Comfort, setup time, robustness outside the lab and calibration effort are still open engineering problems.
Those open problems are exactly why we work across electronics, embedded systems and software at once. Every stage described here, from the electrode to the decision, is a place where a better design can make the whole loop more reliable.

