Aficionados of brain machine interfaces (BMI) have the goal of hooking up the neocortical neurons of an individual with spinal cord or brain stem damage to have him or her control a device with the neurons to restore walking or communication (Birbaumer et al. 1999; Hochberg et al. 2006; Shenoy et al. 2003; Taylor et al. 2002; Wessberg et al. 2000). Since single-cell organisms can be conditioned (Saigusa et al. 2008), it should not be surprising that a single cell of the neocortex can also be conditioned for BMI development. In the study of Prsa et al. (2017), a single neuron of a head-fixed mouse was conditioned in the motor cortex (as measured with two-photon imaging), and feedback of successful conditioning was achieved by optogenetic activation of cells in the somatosensory cortex. The mouse was rewarded with a drop of water following the volitional discharge of a motor cell using the method of Fetz (1969). The conditioning was achieved after 5 minutes of practice, which highlights that the neocortex has a tremendous capacity for making associations (as already discussed), and this is why the neocortex has been the focus of BMI development (Tehovnik et al. 2013).
For an amoeba to learn it must be able to transmit information through its cell membrane so that the internal state of the cell can be modified and have the information stored for long-term use (Nakagaki et al. 2000; Saigusa et al. 2008). As mentioned, like single-cell organisms, multicellular organisms must also internalize and store changes to the environment for learning. The success of BMI, therefore, depends on the extent of feedback during learning (Birbaumer 2006). In the study of Prsa et al. (2017) the feedback came from two sources: from the activation of a population of neurons in the somatosensory cortex and from the delivery of reward which would have engaged the reward circuits of the brain (Olds and Milner 1954; Olds 1958; Pallikaras and Shizgal 2022; Yeomans et al. 1988). In the study of Fetz (1969), monkeys were conditioned by associating neural responses in the motor cortex with the delivery of a reward. It was found that this association could be abolished by cutting the proprioceptive input (Wyler and Burchiel 1978; Wyler et al. 1979), given that when learning the association monkeys often moved their limbs to drive the neurons in the motor cortex to facilitate reward delivery (Fetz 1969; Fetz and Baker 1973; Fetz and Finocchio 1971, 1972). It is noteworthy that following transection of the spinal cord to abolish proprioceptive input, a monkey was observed moving its flaccid arm by the functional arm to try to drive the cells in the motor cortex to obtain a reward (Wyler et al. 1979).
As the sensory feedback of a monkey is reduced on a BMI task, the extent of modulation of the neocortical neurons during task performance is reduced (Tehovnik et al 2013). A monkey was trained to use a manipulandum to move a cursor from the center of a computer monitor to acquire a peripherally located visual target in exchange for a reward (O’Doherty et al. 2011). Three conditions were considered as neurons in the neocortex were activated to move the cursor: (1) moving the hand-held manipulandum to acquire the target, (2) having the hand-held manipulandum fixed in place as the target was being acquired, and (3) having no manipulandum and allowing the animal to free-view the monitor to acquire the target. Going from condition 1 to condition 3, it was found that the modulation of the neocortical neurons dropped by 80%. Thus, as one reduces the number of feedback channels for BMI, expect the firing of the neocortical neurons to decline. This has direct implications for patients who are paralyzed and must therefore rely on the non-tactile and non-proprioceptive senses to engage the neurons of the neocortex.
Another factor that affects the BMI signal is that as the number of recorded neurons in the neocortex surpasses 40 using an electrode array, the information transmitted to drive an external device begins to saturate (Figure 32, Tehovnik and Chen 2015). The best area of the neocortex for getting an optimal BMI signal is in the motor cortex when trying to move an external device that is based on the movements of the forelimbs that engages the visual system, for example (Tehovnik et al. 2013). Also, primary cortical areas are superior to association areas for electrode implantation, since the best signals for BMI are found in areas M1, S1, and A1 (Lorach et al. 2023; Martin et al. 2014; Metzger et al. 2023; Tehovnik and Chen 2015; Tehovnik et al. 2013; Willett et al. 2021, 2023).
Furthermore, it has been known for some time that for a human subject to operate a BMI device using the neocortex, tremendous concentration is necessary and its seems that even with practice the amount of concentration is never reduced (Bublitz et al. 2018). As discussed, a central feature of learning via the neocortex is that with the learning of a task the behavior becomes automated, thereby bringing about the reduction of needed CNS (central nervous system) neurons to perform a task. There is no evidence that neocortically-based BMIs can be automated, since devices need to be calibrated daily (Ganguly and Carmena 2009). Finally, it is known that the neurons in the neocortex, e.g., in area M1, do not follow every behavior faithfully given that the signals are highly variable and prone to wandering when examined across days and months (Gallego et al. 2020; Rokni et al. 2007; Schaeffer and Aksenova 2018). If the neocortex is indeed the center of consciousness (as is presumed here) then one should expect that the neurons in this part of the brain do not discharge lawfully to every motor response as do the motor neurons in the brain stem and spinal cord (Schiller and Tehovnik 2015; Sherrington 1906; Vanderwolf 2007). Whether implanting neurons in the cerebellum might overcome some of the shortcomings found for the neocortex should be considered.
So, how much information is transmitted by a BMI device in bits per second when recording from the neocortex? In 2013, the amount of information transmitted averaged 0.2 bits per second, which was based on work done on behaving primates as well as human subjects (Tehovnik et al. 2013). This value is comparable to the amount of information transmitted by Stephen Hawking (who suffered from amyotrophic lateral sclerosis, ALS) using his cheek muscle at 0.1 bits per second (corrected for information redundancy and based on data from De Lange 2011).[1] This means that at this time there would not have been any advantage for Hawking to use a BMI.
Several studies have been done in recent years that have increased the information transfer rate of BMIs above 1 bit per second. Metzger et al. (2023) developed a BMI to recover language in a patient that had experienced a brain stem stroke that abolished speaking and eliminated the ability to type. A 253-channel electrocortical array was placed on the surface of the sensorimotor cortex over areas that mediate facial movements. It was found that as the subject engaged in the silent reading of sentences, signals could be extracted from the neocortex (with the assistance of artificial intelligence) to generate text at a rate of 78 words per minute at a percent correctness of 75%. This translates into 2.5 bits of information per second, or 5.7 possibilities per second [to derive the bit-rate corrections were made for information redundancy, Reed and Durlach (1998); see Tehovnik et al. (2013) for other details]. This value is consistent with what has been reported by others using depth electrodes implanted in the motor cortex of the face and hand area for silent reading and imagined writing (i.e., ranging from 1.2 to 2.1 bits per second while using the assistance of artificial intelligence, Willett et al. 2021, 2023; also see Metzger et al. 2022).[2] Overall, 1.2 to 2.5 bits per second is to predict 2 to 6 possibilities per second, which is far short of the performance of a cochlear implant (which can predict over 1,000 possibilities per second, Baranauskas 2014), and way short of normal language (which can predict over a trillion possibilities per second, Reed and Durlach 1998).
As for restoring locomotion to spinal cord patients, a major effort was made by Miguel Nicolelis to fit a paralyzed patient in an exoskeleton such that signals were collected from the subject’s neocortex to have him kick a soccer ball with the exoskeleton, which was used to open the 2014 FIFA World Cup (Nicolelis 2019). Realizing that the demonstration was not working, FIFA and the media networks cancelled the broadcast before having the failure transmitted throughout the world (Tehovnik 2017b). Nevertheless, since this time investigators have not given up on the idea of restoring locomotor functions to patients with spinal cord damage. Similar to the study of Ethier et al. (2012), who found that activity from the neocortex could be used to contract the skeletal muscles by discharging the cells in the spinal cord of a monkey, Lorach et al. (2023) found that signals from the sensorimotor cortex of a paralyzed patient could drive the muscles in the legs by having the cortical signals transmitted to the lumbar spinal cord. Recordings were made from each hemisphere using an array of 64 epidural electrodes positioned over each somatosensory cortex. When the patient thought about moving his legs the signal generated in the neocortex stimulated an array of 16 electrodes positioned over the dorsal lumbar spinal cord, such that some combination of 8 electrodes was activated over the dorsal roots of the left spinal cord and some combination of the remaining electrodes was activated over the dorsal roots of the right spinal cord. Consistent with the anatomy, the right neocortex (upon thinking to move) engaged the left spinal cord and the left neocortex engaged the right spinal cord, which elicited a stepping response at a latency of ~100 ms following the discharge of the neocortical neurons, which matches the normal latency.
The walking induced by the implants was slower than that found for an intact system and the patient typically had to walk with the assistance of crutches since postural support was impaired. The minimal number of muscles utilized to walk is eight per leg (including gluteus maximus, gluteus medius, vasti, rectus femoris, hamstrings, gastrocnemius, soleus, and dorsiflexors) for a total of sixteen muscles required (Lorach et al. 2023; Liu et al. 2008). Accordingly, if a ‘0’ and ‘1’ are assigned to the absence and presence of a muscle contraction, then a minimum of 16 bits of information is needed to perform a stepping response. It took the paralyzed patient 4.6 seconds to complete a step (derived from Figure 4f of Lorach et al. 2023), whereas a normal subject takes a 10th of this time to complete a step (based on the step duration of one of the authors). Therefore, the information transferred by the patient was 3.5 bits per second (16 bits/4.6 sec) and that transferred by a normal subject would be 35 bits per second (16 bits/0.46 sec), namely, one order of magnitude less for the patient. Finally, it was found that in the absence of the neocortical implant but having stimulation delivered to the spinal cord implant, the patient could still walk but at a bit-rate of 3 bits per second (derived from Figure 4f or Lorach et al. 2023). Thus, the neocortex added 0.5 bits per second to the information throughput.
Following from the foregoing ‘minimal’ analysis, if each skeletal muscle in the body represents 1 bit of information, then the entire collection of muscles in the body (totaling 700, Tortora and Grabowski 1996) represents 700 bits. We know that language generation requires a minimum of 20 or so skeletal muscles (Simonyan and Horwitz 2011) or 20 bits of information.[3] Generating a muscle contraction every 500 ms would put the bit-rate for language up to 40 bits per second. Accordingly, the skeleto-motor throughput by itself falls well-short of the trillion bits per second estimated for the neocortex or the cerebellum, suggesting further that the information transfer capacity of these structures is mainly for internal use.
Summary
1. A neocortically-based BMI—like a functional brain—is dependent on feedback from the senses to remain operative. The more feedback channels available, the better the signal.
2. An information ceiling occurs when recording from more than 40 neurons in the neocortex using implanted electrode arrays.
3. Neural signals derived from the neocortex are not good for long-term use, and therefore a device would need to be recalibrated daily. Furthermore, to operate such a device requires much concentration on the part of a patient, since the signal does not seem amenable to automation for long-term functionality.
4. In 2013, the amount of information transmitted by a BMI averaged 0.2 bits per second. At this time, it would not have made sense for Stephen Hawking to use such a device to overcome ALS.
5. When electrodes are centered on the writing or speech areas of the motor cortex the amount of information transmitted by the neurons ranges from 1.2 to 2.5 bits per second. This translates into accurately predicting 2 to 6 possibilities per second, which is far short of the performance of a cochlear implant—which can predict over 1,000 possibilities per second—and way short of normal language—which can predict over a trillion possibilities per second.
6. To elicit a stepping response with a BMI, a throughput of 3.5 bits per second has been achieved. This rate is one order of magnitude below the required rate of 35 bits per second to produce a stepping response.
Footnotes:
[1] To determine the information transfer rate from behavioral performance data, see Tehovnik and Chen 2015; Tehovnik et al. 2013).
[2] For the imagined writing, the hand contralateral to the implant was somewhat functional through movement, which could have contributed to the imagined writing (Willett et al. 2021).
[3] A total of 100 muscles are used for speech which control the voice, swallowing, and breathing (Simonyan and Horwitz 2011).
Figure 32. Normalized BMI signal is plotted as a function of neurons. See Tehovnik and Chen (2015) for details. (file: auto_003.gif)