While RNS and DBS are brain implants on the market with on and off label usages, there is also a class of brain implant devices which are purely in the clinical research realm. These brain machine interfaces use microelectrodes which record cellular level data and allow machine learning algorithms to control computer cursors and robotic arms. The first demonstration of this type of device’s efficacy was in non-human primates by the seminal work of Drs. Dawn Taylor, Andrew Scwartz, and colleagues.⁴¹ The microelectrode array, the ‘Utah array,’ was created in Salt Lake City, Utah, by the pioneering implant company, now called Blackrock Neurotech (Salt Lake City, Utah). This 4 × 4 mm array resembles a pin cushion that gets impacted into the cortical tissue with a precise pressurized insertion device (Figure 3). The adaptive-learning algorithm was engineered to sense neuronal firing patterns from the brain tissue and then uses those signals to control a device such as a computer cursor or robotic arm based on these patterns. The concept of ‘decoding neural data’ using machine learning is the foundation of BMIs and came from work by Dr. Schwartz and his mentor Dr. Apostolos Georgopoulos. Amazingly, animals and patients can adapt their own neural activity in motor cortex or parietal cortex through training an adaptive computer algorithm to learn the patient’s brain signals related to the intention to move, and then moving a robotic arm with varying degrees of freedom accordingly. Here AI is the computer model that trains on neural activity related to the desired output such as a robotic arm movement. This model learns a ‘transform function’ which it uses to predict when and how the patient wants to move the robotic arm in a future planned movement. Once trained, the patient can control a machine using the brain implant with their mind. The machine is effectively “mind-reading” via the learned transfer function. This concept opens the door to treating patients who are tetraplegic or otherwise locked-in and unable to communicate or interact with the world. It also leads to some interesting privacy issues such as, should and could there be controls in place for the computer not to read certain types of neural signals?

The first use of brain implants to treat such patients was led by Drs. John Donoghue, Leigh Hochberg, and their team at Brown University and Massachusetts General Hospital, via the BrainGate clinical trials.^42,43 The BrainGate2 clinical trial (NCT00912041) is currently active and recruiting patients with tetraplegia from amyotrophic lateral sclerosis or spinal cord injury. These patients have a Blackrock NeuroPort electrode-based BCI device implanted into the motor cortex or other cortical areas. Patients use their brain activity to train a machine learning algorithm to then control an assistive device. While these clinical trials are certainly tailored to the individual patient, these trials help researchers develop better control algorithms for other BCI applications and helps researchers gain insights into how the human brain works, which they otherwise would not be able to learn. For example, in a study with stroke patients at Washington University in St. Louis, it was noted that patients could control the limb ipsilateral to a control device in motor cortex, when generally we do not think about possible ipsilateral limb control capabilities of motor cortex.⁴⁴ Note that the Blackrock NeuroPort electrode (which is the human version of the Utah array) is not fully implanted. It requires a head-mounted pedestal to transfer data and that piece is exposed outside the skin which may carry a higher risk of infection than a fully implanted device.⁴⁵ Neuralink’s (Fremont, California) N1 Chip mentioned above, is fully implantable and has 1,024 electrodes. Several patients with tetraplegia or tetraparesis have been implanted with this research device in the ongoing PRIME clinical trial (NCT06429735). Paradromics (Austin, Texas) has the Connexus BCI interface that is also fully implantable and supports 1,600+ channels of data, again supporting AI models that require large amounts of data and has also been implanted in humans. Precision (New York City, New York) has a thin seven-layer film designed to capture data at the level of LFPs (NCT05182437) and is designed to treat epilepsy. It is also fully implantable with a battery in the chest and can capture wave phenomena on the brain and has been implanted in several patients. Finally, Synchron (Brooklyn, New York) has created the Stentrode, which is a device with electrodes mounted on a stent that is then implanted in a cerebral vessel near motor cortex. The device records cortical neural activity that is rich enough to run an AI algorithm to control a touchscreen device. The potential advantage here is perhaps a lower rate of infection by being intravascular, as opposed to the immune sheltered environment of the brain. The SWITCH trial (NCT 03834587) enrolled five patients with results pending.

Aside from motor control, speech prostheses designed for communication have also emerged. Here the concept is to decode speech directly from speech-related motor areas including ventral sensorimotor cortex and midprecentral gyrus using a brain implant.⁴⁶ Patients most appropriate have motor paralysis causing dysarthria or anarthria, which is the total inability to produce speech. This could be a result of stroke or amyolateral sclerosis. First demonstrations of speech decoding came from the lab of Edward Chang, MD, followed by others.⁴⁶ This does require that the patient’s ability to understand speech is intact. The control signal is generated usually by imagining the speech. Most recent iterations involve a patient having an avatar perform realistic facial movements as well as generate something similar to the patient’s voice.⁴⁷ Here you can imagine that if the decoding is accurate, any words the patient imagines would be projected, which may compromise patient privacy to some degree.