Brain-Computer Interfaces: History and Overview

Brain-Computer Interfaces (BCI) Overview

• BCIs determine a user's intent by analyzing and classifying brain activity patterns.
• BCIs analyze typical patterns, including electric activity of the brain, and oxygen change.
• BCIs directly provide feedback to the user through visual, auditory, or moving medium.

History of BCIs

• 1929: Hans Berger discovered EEG
• 1950s-1960s: First neurofeedback studies conducted on cats and humans.
• 1970s: Conducted first controlled neurofeedback study in children with ADHD.
• 1973: Jacques Vidal at UCLA first used the term "Brain Computer Interface"
• 1990s: First non-invasive BCI lab by Pfurtscheller, & foundational work by Birbaumer and Wolpaw.
• 2000s: Main interest for BCIs were for communication for adults with disabilities
• 2010s-Present: The field experienced rapid expansion and growing interest in applications.
• Neurofeedback is better marketing for BCIs

Parts of a BCI

• BCIs integrate multiple intersecting research fields.

Key Components of a BCI

• Specific brain activity or pattern is the mental activity to pick up, including responses to external stimuli (exogenous) and internal activity (endogenous)
• Signal acquisition records brain activity using electrodes on the scalp or brain surface. Functional MRI or ultrasound can also be used.
• Signal processing and classification extracts relevant features from the brain pattern, that removes artifacts, identifies features, filters, and uses machine learning to classify features.
• Control signal and feedback provides feedback to the user through a control signal and indicates if the BCI is accurately interpreting the user's intent.

BCI Operation

• BCIs are controlled when a user performs mental activity, that produces a repeatable, identifiable neurophysiological pattern.
• Brain activity is recorded using EEG or ECOG.
• Samples are recorded at rates ranging from 128 Hz to 30k Hz.
• The recorded brain activity is processed to remove noise and artifacts that could contaminate the pattern of interest.
• Incoming signals are processed to remove artifacts, identify target patterns, and extract features of interest.
• Machine learning algorithms identify neurophysiological patterns or features present in the recorded signal.
• Classification steps require clear, labeled examples of neurophysiological features of interest, known as supervised learning. Then a control signal to execute a certain command or drive a specific application to provide feedback is sent.

Electric Activity in the Brain

• Excitatory neurons synapsing (EPSP) leads to generating action potential.
• Inhibitory neurons synapsing (IPSP) stops action potentials.
• Measuring the change in electric activity near neurons, and in the extracellular space near them uses dipoles.
• Local Field Potentials (LFPs) measure pooled synaptic potentials from groups of neurons near a recording electrode.
• Local Field Potentials provide insights into correlated neural activity, and are known as "micro" or "intracranial" EEG.
• The cortex gets a dipole with the pyramidal neurons orientation for easy measurement with the difference in electric activity across the neuron from positive to negative.
• Greater dipole strength can be measured through tissues.

Electroencephalography (EEG)

• An electroencephalography (EEG) records the LFP integrated over a relatively large spatial area (10 cm^2 or more).
• In electroencephalography, sensors are placed on the scalp.
• EEGs have good temporal resolution, but poor spatial resolution.
• Hard to tie contributions of individual neurons because of distortions between the current source and the electrodes in the scalp.
• Advanced techniques can get better spatial resolution.
• The signal is distorted due to the CSF, skull, and other tissues.
• Volume conduction effect is caused because the sensor is far from the source of electric activity.
• EEGs are the most used recording method for non-implanted BCIs.
• "Standard" 10/20system for measuring EEG, with Nasion -> Inion in in 10% and 20% intervals.
• Channel names are roughly representative of their primary lobe.
  • Frontal -> F (prefrontal sometimes Fp)
  • Central -> C (roughly near central sulcus and motor cortex)
  • Parietal -> P
  • Occipital -> O
  • Temporal -> T
• Odd numbers indicate left hemisphere and even numbers indicate right hemisphere, and F3 is left-hand of frontal region
• “z” indicates midline, with Pz being the parietal lobe and midline
• Numbers increase laterally from midline
• 04 would be more lateral than O2 on the right side of the midline

Recording at Depth

• There are many different techniques to record electric activity of the brain, each has advantage/disadvantage
• Deeper electrodes have much higher spatial resolution and less noise

Electrocorticography (ECoG)

• ECoG records electric activity with sensors on the exposed surface of the cortex.
• ECoG still records LFPs, but has higher resolution, and less signal atenuation, and uses biocompatible materials, to get resolution < 5 mm^2.
• “Burr hole" or craniotomy is required for ECoG placement.
• Craniotomy removes whole sections of the skull to place ECoG or other implants, while Burr holes use a smaller precise drill to create a smaller well.

Controlling a BCI

• BCIS use mind control

Exogenous vs. Endogenous Paradigms

• Exogenous BCIs use brain activity elicited by an external stimulus.
  • Control signals have a clear pattern, less training required.
  • Signals can often be found with fewer channels.
  • Some users may find that it is difficult or tiring to attend to an ongoing stimulus.
  • Event Related Potentials (ERPs)
    • P300 (visual, auditory, somatosensory)
  • Visual evoked potentials (VEPs) -Steady-state (SSVEP)
• Endogenous BCIs use self-paced control of brain activity without any external stimuli.
  • Have lower information transfer rates, and are operated at own will
  • Better those with sensory impairments.
  • Time-consuming training
  • Some users may be unable to master control.
• Involve motor imagery, slow cortical potentials, and mental tasks like mental arithmetic, number or word generation.
• Both exogenous and endogenous Paradigms sample activity from well defined areas of the brain.

Exogenous Paradigm

• Generate a P300 response, and use an oddball paradigm, that presents the user with repetitive ‘non-target’ stimulus, and occasionally interrupt with presentation of the 'target' stimulus.
• Presenting a flashing light source at a specific frequency generates an evoked response/pattern in the EEG of the visual cortex that will oscillate at the same frequency, as shown with Steady-State Visual Evoked Potentials.
• The EEG signal is analyzed in the frequency domain, where a peak indicates the oscillating frequency.

Endogenous Paradigm

• Imagined motor movement causes localized, synchronous activity over the motor cortex.
• Actual motor movement is not needed
• Motor imagery on the left side of the body produces different patterns than a limb on the right side as well as hands vs feet and that allows for multiple control options.
• The user imagines a movement evaluated by trained ML decoders and recognizes the desired pattern.

Types of BCIs

• BCI design involves balancing invasiveness, signal quality, cost, safety, comfort, spatial/temporal resolution, and ease of use.
• Implantable BCIs have a higher spatial and temporal resolution, which make them very powerful, at higher risks.