Signal quality in wearable EEG: from electrode to data

A wearable EEG device is only as useful as the signal it records. Every stage between the electrode and the stored data can add noise, and little of it can be removed afterward.

Most of the decisions that set signal quality are made early, in the electrodes, the mechanics that hold them and the analog front end. The sections below follow the signal from the electrode to the stored data, and end with how to verify each stage.

What the hardware records

Electroencephalography (EEG) records the very small electrical potentials that brain activity produces at the scalp or in the ear. For the hardware, these are differential voltages between electrodes, much smaller than the offsets, interference and artifacts that arrive with them, so the design has to tolerate large unwanted signals without saturating while letting the small wanted one through cleanly.

Electrodes

The electrode is often the largest source of noise and instability.

Wet, dry and in-ear electrodes

• Wet electrodes. A conductive gel between electrode and skin gives a low, stable contact impedance, but it has to be applied and cleaned off, and it dries out over long recordings.
• Dry electrodes. Metal or conductive polymer contacts touch the skin directly, often through pins that reach through hair. They are quick to put on, but their impedance is higher and more variable, and they are more sensitive to movement.
• In-ear and around-ear electrodes. These make a discreet device, but the skin area is small, the electrodes sit close together, which tends to give smaller signals, and the fit varies from ear to ear.

Contact impedance and pressure

The junction between an electrode and the skin behaves roughly like a resistor and a capacitor in parallel, in series with a small voltage source. Its impedance depends on the electrode's material and area, the pressure on it, hair and sweat, and it changes with frequency and over time.

An impedance that changes as the wearer moves produces signals of its own. Design for consistent pressure, with springs, compliant materials or a headband of controlled tension, and keep it comfortable, since a device that presses too hard gets adjusted, which disturbs the contact.

Offsets and matched impedances

Each junction also produces a DC potential, set largely by the electrode material. The difference between two electrodes appears as an offset that can be far larger than the EEG, and it drifts as the contact changes. Use the same material on every electrode, since mixing metals makes the offsets larger.

Interference from the surroundings appears almost equally on every electrode, and a differential amplifier rejects it, but only as well as its inputs are balanced. Each electrode's impedance forms a divider with the amplifier's common-mode input impedance. When two electrodes' impedances differ, the interference is divided differently on each input, and part of it becomes a differential signal that cannot be told apart from EEG. A very high common-mode input impedance and low, closely matched electrode impedances both limit this.

The analog front end

The front end amplifies the difference between electrodes, rejects what they share and prepares the signal for the converter.

• Low-noise amplification. The first stage sets the noise floor of the system, so its input-referred noise must sit well below the smallest signal to be recorded. With high-impedance dry electrodes, current noise and input bias current matter as much as voltage noise, because they flow through the electrode impedance.
• Common-mode rejection. A high rejection ratio in the amplifier is not enough on its own. Filter and protection components in front of it must be matched across each channel's inputs, or they convert common-mode interference into a differential signal.
• Bias and reference electrodes. A bias electrode holds the body's common-mode voltage inside the amplifier's input range; driven with the inverted common-mode signal, it also reduces the interference on the body. A shared reference electrode is the point each channel is measured against, so its position affects every channel.
• Input protection and electrical safety. The electrodes make a conductive connection to the wearer, so limit the current any electrode can carry, in normal operation and if a component fails. How to limit it depends on the circuit: a resistor in series with each electrode line is common, but the right protection for each line, including the bias electrode's drive, follows from the front end's design and its manufacturer's guidance. Work out each electrode's worst-case current from the actual circuit, including with a failed component, and keep any current used to detect a loose electrode as small as the front end allows. Protect the inputs against electrostatic discharge with low-leakage parts. Series resistance adds noise, so size it within the front end's noise budget. Where the product's use calls for a formal electrical-safety assessment, the applicable standards set the limits; design to them from the start.
• Offset tolerance. The front end must accept electrode offsets without saturating. Either keep the gain low and use a converter with enough dynamic range to digitize the offset and remove it later, or couple the inputs through a high-pass filter before high gain.

Filtering before conversion

A high-pass filter removes offset and slow drift. A lower corner keeps more of the slow content but settles more slowly after a disturbance, such as an electrode being reseated, so many designs keep the analog high-pass very low and do the final filtering digitally, where it can be changed.

An anti-aliasing filter attenuates content above half the output sampling rate, which would otherwise fold back into the recorded band. It reduces aliasing rather than removing it, and the attenuation needed depends on how strong the out-of-band content is, such as muscle activity or radio interference, compared with the smallest signal to be recorded. Oversampling (delta-sigma) converters do most of this filtering digitally, which relaxes the analog filter in front of them. Their digital filter cannot reject content near the converter's internal sampling frequency and its multiples, so an analog filter is still needed. How much it must attenuate depends on the converter, through its internal sampling frequency and digital filter, and on the interference where the device is used, such as nearby radio transmitters and switching circuits; the converter manufacturer's guidance is the starting point. Match its parts across each channel's inputs, as for the protection parts. Record the filter settings with the data, so that anyone who analyzes it later knows what the hardware removed.

Conversion

• Converter noise and range. The converter's own noise, referred to the input at the chosen gain and data rate, should be well below the front end's, so that the amplifier sets the noise floor. Its input range at that gain must take electrode offsets and artifacts without clipping. Compare converters by their noise or effective resolution at the data rate and bandwidth to be used, not by their nominal bit count; in an oversampling converter, noise falls as the data rate is lowered.
• Sampling rate. Sample comfortably above twice the highest frequency to be kept. A higher rate eases the anti-aliasing filter but raises the data rate and the power the radio draws.
• When simultaneous sampling matters. A multiplexed converter samples channels one after another, so each channel's samples are shifted in time by a fixed amount. The shift acts as a phase error that grows with frequency. It matters when channels are combined or compared: subtracted in software, for example when the reference is changed, where shared interference no longer cancels exactly, or analyzed for timing relationships between channels. Where it matters, use a converter that samples all channels at the same instant, or correct the known shift in processing. In either case, run every channel from one stable clock.

Interference

Power wiring in walls and nearby equipment is the most common source of interference. The body couples capacitively to it and picks up a voltage at the power-line frequency, mostly as a common-mode signal. The rest enters through the leads, by electric and magnetic coupling.

• Short leads, routed together. Keep electrode wires short and route each input close to its reference, so that coupling affects both equally and the loop between them stays small.
• Buffering at the electrode. Active electrodes put a buffer amplifier at each electrode, so the lead is driven from a low impedance and becomes far less sensitive to coupling and cable movement. The cost is power, parts and a buffer's noise and offset at every electrode.
• Battery power. Running from a battery, with no wired connection, reduces power-line interference but does not eliminate it. The body and the device still couple capacitively to nearby power wiring, so a common-mode signal at the power-line frequency remains, and imbalance between electrodes turns part of it into the recording; good common-mode rejection and the bias electrode are still needed. A charger, or a cable to a computer powered from the wall, adds a much stronger interference path, and if that equipment develops a fault, it can expose the wearer to a hazardous voltage through the electrodes. The safest design does not record while charging, and keeps the charging contacts out of reach while the device is worn. If recording while connected is required, isolate the connection and test that case separately.
• The device itself. Switching regulators produce ripple, and the radio draws current in bursts and transmits next to the inputs, where its energy can be rectified into a signal at the packet rate. Give the front end a quiet supply, and keep the radio and switching circuits away from the inputs.

Motion and other artifacts

An artifact is anything in the recording that does not come from the activity the device is meant to record. Some can be reduced by design; the rest should at least be easy to recognize.

• Cable movement. A moving cable changes its capacitance and can generate charge by friction, and on high-impedance inputs both appear as signals. Keep cables short and fixed, or place the electronics next to the electrodes.
• Electrode movement and drift. When an electrode shifts or its pressure changes, its impedance and DC potential change too, producing large, slow deflections. Sweat and drying gel cause slower drift. Steady mechanical contact is the main defense.
• Muscle activity. Jaw, face, scalp and neck muscles produce broadband electrical activity, often much larger than the EEG. Electrodes near the temples, the jaw and the ears pick up more of it.
• Blinks and eye movements. These produce large, slow potentials, strongest at electrodes near the front of the head, and appear in almost every recording.

The digital side

Once the signal is digitized, the task is to deliver every sample with its correct time, and to make every missing sample visible.

• Timestamps at the source. Count samples on the device, from the converter's clock. Arrival times at the receiver are unreliable, because a wireless link delivers data in bursts with variable delay.
• Synchronization. Events, such as a button press or a marker sent from the receiving device, need a timestamp on the same timebase or a measured offset to it. The device and receiver clocks drift apart, so measure that offset repeatedly, not only at the start.
• Buffering through dropouts. Wireless links stall and lose packets, especially when the body blocks the path. Buffer enough data on the device to ride through the expected gaps and retransmit, or store the recording locally and transfer it later.
• Data rate against radio bandwidth. The channel count times the sampling rate times the bits per sample, plus headers, timestamps and retransmissions, must fit within the throughput the link achieves while worn, with a margin. Lossless compression can help.
• Lost samples marked, not hidden. Put a sequence number in every packet so that the receiver can detect gaps, and record each gap in the data as an explicit marker, rather than interpolating across it or joining the samples on either side.

Power

Two trade-offs dominate the power budget.

• Front-end noise against current. Lower amplifier noise generally costs more current, for every channel and, with active electrodes, at every electrode. Choose the front end that meets the noise target at the lowest current, not the quietest one available, and switch off unused channels.
• Radio duty cycle. The radio is often the largest consumer. Collecting samples into larger packets lets it stay off for longer than sending each sample as it arrives, at the cost of latency and a larger buffer. Plan that buffer together with the one that covers dropouts.

Verification

Verify signal quality on the bench with known inputs first, then while the device is worn.

On the bench

• Shorted-input noise. Short the inputs at the electrode connector following the front end manufacturer's recommended procedure, which sets how the inputs, the reference and the bias connection are joined, so that the result can be compared with the specified noise. Record with the device on its battery, away from power cables and other sources of interference, using the gain, data rate and digital filtering of normal recording, and long enough for a stable result. Compare the noise with the front end's specified input-referred noise for the same bandwidth and conditions. Many front ends also offer an internal input-short setting, which measures the amplifier and converter without the connector and wiring; the difference between the two shows what the wiring and layout add. Then repeat with the radio transmitting, to expose interference from the device itself.
• Known test signals through a divider. A resistive divider brings a signal generator's output down to the small amplitudes of EEG. Use it to check gain, frequency response, filter corners, channel matching and timing. Apply it to all inputs at once to measure common-mode rejection, and add a resistor in series with one input to see the effect of imbalance.
• Impedance checks. Many front ends estimate contact impedance by passing a small known current through the electrodes. Verify this against known resistor and capacitor networks, then use it to check the contact before and during recordings.

While worn

Begin with the wearer seated and still, then record while walking, talking, chewing and turning the head, and near laptops, chargers and power cables. Label each segment as it is recorded, so that artifacts can be matched to their causes. Repeat with several wearers, since head shape, hair and ear shape all change how the electrodes sit.

A checklist for wearable EEG signal quality

• Every electrode uses the same material and is held with steady, comfortable pressure.
• The front end's input impedance is high compared with the electrodes', and its filter and protection parts are matched across inputs.
• The front end accepts electrode offsets without saturating, and the filter settings are recorded with the data.
• All channels run from one stable clock, and are sampled at the same instant wherever channels are combined or compared.
• Each electrode's worst-case current is limited by protection suited to its circuit, and the device does not record while connected to wall-powered equipment unless that connection is isolated.
• Input leads are short and routed together, or the electrodes are buffered.
• The radio and switching regulators are kept away from the inputs, and their effect is checked with shorted inputs.
• Samples are counted on the device, and lost samples are marked in the data.
• The data rate fits the wireless link as worn, with a margin.
• Bench tests pass before recordings while worn begin, and every worn recording is labeled by activity.