Signal Acquisition

Here we will outline some common examples:

EEG

When brain areas are activated, a large population of neurons fire together in synchrony and generate a quantifiable electric field. This electric field can be read by an electroencephalogram (EEG) sensor, a standard hardware component in many brain-computer interface devices.

fMRI/fNIRS

Functional magnetic resonance imaging (fMRI) and functional near infrared spectroscopy (fNIRS) measure hemodynamic (blood flow) changes which reflect neuronal activity. How are these linked? Neuron firing is extremely energy intensive. Studies estimate that the brain uses ~20% of glucose derived energy in the body. Replenishing glucose, oxygen and other nutrients is essential, and requires blood flow into brain tissues.

How do the devices capture signals?

This depends on the the type of machine/device.

fMRI tracks a blood-oxygen-level-dependent (BOLD) signal, which changes based on the amount of blood flow and brain oxygenation to areas of the brain. These changes infer a change in neuronal activity at localized sites. Activity levels in different areas might indicate higher levels of specific functions, e.g mental planning and executive decision at the frontal lobe, or motor commands for movement at the primary motor cortex. This type of information is suited for longer patterned activity on the order of seconds.

fNIRS similarly estimates cortical hemodynamic activity but by measuring absorption of near-infrared light by the same chromophores (oxyhemoglobin and deoxyhemoglobin) used to measure BOLD signals in fMRI. This type of information has better temporal resolution than fMRI and can sample signals at around 1-10Hz, and typically up to 100Hz.

For EEG signals, raw voltage data and variations in amplitude can infer changes in neuronal activity. Mathematical transformations allow for estimates of signal sources within the brain, known as source localization. The transformations are done to mitigate a phenomenon called volume conduction, which describes how electric fields spread out in the cerebral spinal fluid (CSF) that surrounds the brain. In other words, the CSF causes sensors to pick up signals that originate from areas distant to the site of the sensor itself. This is less of a problem for capturing mental states and functions for an AI model if the focus is on patterns of activity, rather than locations of activity. This type of information has the best temporal resolution with a sampling rate typically in the range of 500Hz-1000Hz.

Device Availability and Considerations

fMRI machines are large and not easily moved. Access is limited to hospitals, clinics and research centers. EEG and fNIRS devices are wearable and are available commercially. In terms of temporal resolution, the rank order is fMRI<fNIRS<EEG. In terms of spatial resolution the opposite is true. Another method not discussed here is Magnetoencephalography (MEG). This method is quickly gaining popularity due to its easier machine accessibility and relatively easier portability compared to fMRI. It blends the advantages of EEG and fMRI, providing high spatial and temporal information.