BCIH: "Bidirectional Neural Interfaces" [Chp. 37]

[This post based on the Brain-Computer Interfaces Handbook, Chapter 37]

What is a bi-directional brain-computer interface?

Bi-directional brain-computer interfaces (BCIs) are BCIs that can both read and write data from the brain. For example, a bi-directional BCI might read motor data from the brain to control a prosthetic arm, and then write sensory data back to the brain about where that prosthetic is in space.

Bi-directional BCIs can be divided into their afferent (write) and efferent (read) elements, modeled after afferent and efferent neurons in the body:

1. Afferent Neurons: Assimilate information in the brain by sending info towards the Central Nervous System (CNS).
2. Efferent Neurons: Export information away from the brain/CNS and towards the rest of the body.

Thus, in the example above, the afferent element is the sensory data about where the prosthetic is in space, and the efferent element is the motor data that controls the prosthetic arm's movement. 

Challenges in Creating a Bi-directional BCI

The three main challenges in creating a bi-directional BCI are reading, writing, and decoding brain activity. 

Reading/Writing Neural Signal Challenges

1. Neuron Size: Because neurons are so small and dense, being precise about which neurons a given electrode is writing to is difficult.
1. Potential Solutions: Neuralink's polymer-based "threads," carbon nanotube electrodes, recording at multiple points on one electrode.
2. Brain Depth: The extensive folding and depth of the brain means that beneath any electrode is many layers of neurons, which makes it difficult to determine which part of a signal being read is coming from the neurons of interest, and which part is interference from neurons deeper down.
1. Potential Solutions: Grids of electrodes (e.g., an ECoG grid) combined with mathematical techniques to estimate contribution from sources at different depths.
3. Invasive vs. Non-Invasive: While the signal quality for invasive BCIs is much better than for non-invasive BCIs and write fidelity/precision is much higher, invasive BCIs are more traumatic to the brain, carry an increased risk of infection, and more difficult to maintain long-term.
1. Potential Solutions: Smaller, more flexible electrodes, a flexible neural "mesh," electrode grids/arrays instead of raw electrodes, Neuralink's "neural threads."

ECoG grids (invasive) show significant promise for chronically implanted bi-directional BCIs because the surgery is minimally invasive, high recording resolution is achieved, and the electrodes can both read and write from the area they're implanted over.

Decoding Neural Signals Challenges

1. Assuming Signals Encode Behavior of Interest: Typically, it's assumed that the signal being read directly encodes the behavior of interest; however, a precise mapping between neuron clusters in the brain and their function is not well developed, so this assumption can be of questionable validity.
1. Potential Solutions: Continued neuroscience research to determine exactly which areas of the brain and neuron groups encode what behavior. Exclusively using recent neural activity (from during behavior of interest) to predict future activity can help keep focus to only relevant neural activity. 
2. Making Sense of Neural Activity: As mentioned above, we don't have a mapping from each neuron to its meaning. Thus, we must use algorithms to try to discover the patterns linking a behavior of interest with the relevant groups of neurons. 
1. Potential Solutions: Several approaches exist to efficiently decode and make sense of neural activity. For example:
1. Population vector, based on neural tuning (correlation between a neuron's firing pattern and a certain behavior).
2. Artificial neural networks.
3. Kalman filter (creates an internal predictive model, then refines the model over time as more data is read in and error is optimized).
4. Linear discriminant analysis (LDA). 

Existing Applications of Bi-directional BCIs

Several types of bi-directional BCIs have already been created. The following table summarizes some applications and the different components of each application:

Application	Efferent Pathway	Afferent Pathway
Prosthetic	Motor cortex, to guide a robotic arm or other limb	Artificial somatosensory sensation, to provide perception of where the prosthetic is
Virtual Touch	Motor cortex, to guide a virtual hand	Artificial sensation, to provide an illusion of touch
New Sensation Source	Motor cortex, to modify aperture of a 'dataglove' to the target aperture	Arbitrary sensation provided via ECoG, which the user learns to interpret as feedback on how close their hand is to the target aperture

Future Bi-directional BCIs

The most interesting bi-directional BCIs are yet to come. Proposals for future BCIs include ones that can control cognitive and emotional states, as well as ones that would enable direct, brain-to-brain communication.

Some very basic forms of these have already been created, including one which uses brain-to-brain communication in rats for a lever task. In the study, one rat was presented with two levers and two LED lights, and it received a reward if it selected the lever corresponding to the LED that lit up on a given trial. The brain signals from that rat were then put through a sigmoid transform and sent into another rat using ICMS. Both rats then received reward if the second one pressed the same lever as the first. Results showed the second rat selected the correct lever at a level significantly above chance.

Current, mono-directional BCIs also hint at what future, bi-directional BCIs may add. For example, there are several motor BCIs that just read data and send it to muscles ('efferent' BCIs), including ones that restore paralyzed muscle control using functional electrical stimulation (FES) and ECoG-based BCIs that read brain activity to control a cursor. In the future, these might also send data back to the brain like state information for the muscles or information about what's under the cursors.

Interesting afferent-only BCIs often relate to artificial sensation, and include ones that can be used to restore hearing (cochlear implants) or simulate feeling in areas a patient has lost it. In the future, a cochlear implant might be able to be controlled and calibrated directly from the brain, and a BCI simulating feeling might also write motor movement to the areas where the patient has lost feeling.

Summary

Bi-directional BCIs (and mono-directional BCIs) are difficult to make for a variety of reasons, with the most significant being lack of knowledge about exactly what each part of the brain does, the very small size and high density of neurons, and the difficulty in separating brain signals of interest and decoding them. 

However, bi-directional BCIs for use in prosthetics and other applications already exist, and future bi-directional BCIs could have applications as currently unimaginable as direct brain-to-brain communication.

Sources

1. Brain Computer Interfaces Handbook, Chapter 37
2. Virtual Touch Bi-directional BCI: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2741294, https://www.ncbi.nlm.nih.gov/pubmed/21976021
3. New Sensation Source Bi-directional BCI: https://www.ncbi.nlm.nih.gov/pubmed/27429448
4. Brain-to-Brain Communication in Rats: https://www.nature.com/articles/srep01319