Slow Firing Single Units Are Essential for Optimal Decoding of Silent Speech

doi:10.21203/rs.3.rs-1196637/v1

PDF

Research Article

Slow Firing Single Units Are Essential for Optimal Decoding of Silent Speech

https://doi.org/10.21203/rs.3.rs-1196637/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted 03 Jan, 2022

You are reading this latest preprint version

Summary The motivation of someone who is locked-in, that is, paralyzed and mute, is to find relief for their loss of function. The data presented in this report is part of an attempt to restore one of those lost functions, namely, speech. An essential feature of the development of a speech prosthetic is optimal decoding of patterns of recorded neural signals during silent or covert speech, that is, speaking ‘inside the head’ with no audible output due to the paralysis of the articulators. The aim of this paper is to illustrate the importance of both fast and slow single unit firings recorded from an individual with locked-in syndrome and from an intact participant speaking silently. Long duration electrodes were implanted in the motor speech cortex for up to 13 years in the locked-in participant. The data herein provide evidence that slow firing single units are essential for optimal decoding accuracy. Additional evidence indicates that slow firing single units can be conditioned in the locked-in participant five years after implantation, further supporting their role in decoding.

Neural signals

neurotrophic electrode

single unit firings

deep learning

neural net computing

speech prosthetic

locked-in participants

A speech prosthetic is essentially a brain to computer interface between the speech areas of cortex and a computer. More specifically, in the iteration reported herein, it records signals from the motor articulatory area of the brain and transmits these neural signals to a computer that decodes the signals to produce speech in real time. There are several other approaches to developing a speech prosthesis. Most recently, Willett et al [1] have recorded from human arm motor cortex and provided a writing facility, easily convertible to speech, using Utah arrays and recording multi-units. Other researchers [2, 3, 4] have used micro-electrode arrays to control paralyzed limbs and robotic limbs, also recording multi-units. ECoG electrodes placed on the cortical surface have been used by Chang and his group [5] to develop a speech prosthesis using the frequency domain to interpret and produce the intended speech. In the present report, the Neurotrophic electrode is used to record single units [6]. A thorough review of these techniques is outside the scope of this paper and is published elsewhere [7].

The locked-in syndrome is consequent on two major conditions: 1] amyotrophic lateral sclerosis is a devastating and slow paralysis; 2] brainstem stroke it is a sudden and complete paralysis. In either case, it leads to loss of use of the articulators, and hence loss of speech, even though the neural substrate in the cortex and sub-cortical areas are intact. These cortical areas are the targets of research efforts to develop a system of stable and long lasting recordings. It follows therefore that only those electrodes that have proven stability and duration should be candidates for human implantation. The discussed systems have been demonstrated to produce functional signals (individual units, multi-units, frequency domain surface recordings) for months and years. Published data on the Utah array, however, indicates loss of 85% of single units after 3 years [8] though anecdotal evidence suggests longer duration of multi-unit signals. ECoG array duration was assessed histologically at 22 months with recordings continuing up to 18 months [9]. Histological analysis indicated giant cells and macrophages at the interface, with encapsulation of the electrodes by collagenous tissue. This histology strongly suggests attempts to reject the implant. A more recent configuration is the placement of multiple gold electrodes on the cortical surface, but no long term data has been produced to validate the longevity of this approach [10].

Another approach to avoid rejection of the electrode is to grow the neuropil inside the electrode tip. Using this approach, data indicate stability of single units up to four years in four participants and nine years in participant 5 in this study [6, 11, 12, 13]. The stability of single unit recordings after five years of implantation was such that the participant controlled movements of a cursor in a 2D formant frequency plane to activate vowel sounds [14, 15, 16], as well as decoding phones and words [6]. Post mortem histological verification 13 years after implantation indicated no gliosis, abundant myelinated axons, and normal neuropil except no neurons (the neurons remain outside the glass tip and grow neurites into the tip under the influence of neurotrophic factors) [17]. The histological analysis result is identical to prior rat and primate studies using light microscopy and electron microscopic techniques [18]. In addition, single units remained functional at year nine [13]. Thus the question of longevity and stability of recorded single units, as well as the lack of gliosis, has been established with the Neurotrophic electrode. So far, no other electrode has been shown to have these capabilities.

Advances in decoding with neural net paradigms takes decoding to another level. These paradigms allow accurate classification of phones as described in this paper and phrases in a prior paper [6]. The classification includes results with slow firing single units as well as fast firing single units. This paper focusses on both fast (5 ips and above) and slow firing units, with particular emphasis on the importance of slow firing single units for improved accuracy of classification. There is no study known to these authors where consistently slow firing single units were specifically related to the task at hand. Some single units are known to cease firing and then burst to a high frequency, but consistently slow, non-bursting units, firing in the 0 to 5 ips range are not known to have been studied, at least with respect to neural prosthetic usage.

In this study, two human participants were implanted with the aim of understanding how to develop a speech prosthetic. The first participant, locked-in due to a brainstem stroke at the age of 16 years, was implanted in 2004. Multiple publications have described the results [6, 11, 12, 13, 14, 15, 16, 17]. Even though this locked-in participant (number 5 in FDA IDE G960032) could control his single unit firings, the technology available at that time was not amenable to the production of speech.

To better understand decoding of single units associated with the timing of audible and silent speech, one author (PK) decided to have his speech motor cortex implanted in 2014. An important result from studies in both the locked-in and speaking participants is that even though fast firing units are the most important, slow firing are important for optimal decoding. This issue is addressed in this paper. In addition, results from conditioning of slow single unit firings illustrate the importance of including slow firing in decoding, so that even if not related to the task they can be conditioned to the task ad thus improve decoding capability.

The late David Jayne, an ALS person and friend, insisted that I focus my efforts on restoration of speech, not movement, because as he said “speech will make me human again. I will talk to my children!”. So the aim of this effort is to restore at least 100 useful words to those who are mute and paralyzed. The expectation is that they speak at a conversational rate, or at least at a rate that allows the individual to hold a conversation.

Participant 5 of FDA study (G960032) gave his permission using up movements of his eyes for agreement and down movement for disagreement with the informed consent regulations of Neural Signals Inc with the regulations being read to him and his father. His parents also agreed to the regulations. Participant 6 was one of the authors (PK).

Strategy

Archived data from locked-in participant 5 and intact participant 6 (PK) are analyzed during silent speaking of phones using MATLAB’s Classification app. This app provides an accuracy rating of the ability to distinguish between individual phones. To test the contribution of fast and slow firing single units to the decoding accuracy, single units are removed from the analysis one at a time and the accuracy is recalculated. In addition, groups of fast and slow are removed and the accuracy recalculated. For conditioning of slow single units, data recorded over several days involved locked-in participant 5 listening to a tone, having a quiet period, and then ‘singing’ it in his head. When feedback was provided, conditioning improved and remained stable.

Electrode

The electrode assembly is shown in figure 1. Construction has been detailed [29]. The cone is made by pulling a heated pipette to produce tip dimensions of 1.5 mm in length, 25 microns at the deep end and a few hundred microns at the upper end to allow space for the four inserted wires. 2 mil Teflon insulated gold wires are coiled around a pipette and glued with methyl-methacrylate inside the glass cone. The other end of each coiled gold wire is soldered into a connector that plugs into the implanted electronic component. His electrode is FDA approved (IDE G960032).

Implanted Electronics

The single channel electronics are assembled in-house and FDA approved (IDE G960032). Bipolar amplifiers record pairs of wires via the low impedance (50 to 500 kohms) gold wires that are cut across the tips to provide the low impedances. These connect to an amplifier with a gain of 100x and the signals are filtered between 5 and 5,000 Hz. The signals then feed into an FM transmitter operating in the carrier range of 35 to 55 MHz. During recording sessions, a power induction coil powers the device with the induced current passing through a regulator to provide a stable +/- 3 volts. The electronics is insulated with polymer (Elvax: Ethylene Vinyl Acetate Copolymer Resin, from DuPont, Wilmington Delaware 1998) and further insulated (and protected against trauma during handling while receiving nursing care) with Silastic (Med-6607, Nusil, Silicone Technology, Carpentaria, CA). The gold pin connection to the electrodes is protected with acrylic cement (Medtronic Inc., St. Paul, MN). After implantation, the device is covered with scalp skin. The result is illustrated in figure 2: a lateral X ray of the skull (in participant 6) indicating three of the eight pairs of electrodes wires are attached to three sets of connecting pins that are, in turn, attached to three single channel electronic amplifiers and FM transmitters. There was insufficient space to place more than three electronic devices.

Implantation Target Site

Because the speech prosthetic is based on movement of the articulators, the speech motor cortex is targeted for implantation. Functional MRI studies during audible and silent speech confirmed the target location as previously described in detail [13]. This area extends from the Sylvian fissure medially for 30 mm and about 20 mm in the rostro-caudal dimension.

Surgery

Briefly, under fully sterile conditions, a craniotomy is performed over the left sided speech motor cortex and electrodes are implanted as previously described [13]. The electronics are attached in participant 5 and at a later surgery in the speaking human. Recordings began at month 4 in both participants.

Recording

The recording systems are detailed elsewhere [13]. Briefly, the power induction coil is placed over the scalp and the underlying receiving coil, and powers the implanted electronics. Data from the FM transmitters are received via a coil placed on the scalp over the transmitting coil. These data are sent to a receiver that sends it on to the Neuralynx (Bozeman, MT) computer that contains the Cheetah cluster cutting system that separates the single units from the continuous data stream. These are then transferred to a Laptop computer for off-line analysis employing the Classification App from MATLAB (Natwick, MA).

Paradigms

The paradigm used for locked-in participant 5 is shown in figure 3a. It consists of the computer first outputting a phrase “listen to the sound” followed by the phone or word repeated three times. Then the computer emits a phrase “say the sound”. We then assume the participant says the sound in his head. This is repeated at least ten times. We ask him later to confirm that he did say the phone or word by rolling his eyes up, or deny saying it by rolling his eyes down.

The paradigm used in the speaking human (PK) consists of a control period, followed by a speaking period and then silently (in the participant’s head) as shown in figure 3b. Note the timing marker that is activated with the left hand. The left hand is preferred to the right so as to avoid possible contamination of the data if the right hand is used (because it is contralateral to the left speech motor cortex). The event markers indicate the control period followed by the audible speech and then the silent speech.

The conditioning study: Participant 5 first listened to a tone for 10 seconds, had a rest period for 10 seconds and then attempted to sing the tone in his head for 10 seconds. Tones are emitted from a sine wave generator as major keys C D G and A in different octaves. All single units are involved in this conditioning study, whereby the participant is asked to sing in his head as accurately as he could. In the first session no feedback is provided. During the second session feedback of single unit firing is provided as a single brief tone each time the single unit fires. During the final session, feedback of single unit firing is provided and the volume of the feedback is directly related to the firing rate.

Spike Sorting

An example of a continuous stream of multi-single unit neural activity is shown in figure 4a over a 40 ms time base. Cheetah Spike Sorting software (Neuralynx, Boseman, MT) is employed to sort the continuous data streams into identifiable single units. The preferred algorithm is the convex hull technique which uses a combination of parameters such as peak, valley, height and area under the curve of the presumptive spikes that are shown in Figure 4b. The program first separates presumed single units into upward or downward action potentials thus creating two channels of data. It then applies the parameters for single unit separation to each channel (panel 1). The clusters are selected by circling them with the cursor to produce multiple waveforms (panel 2). These are then cut (panel 3) by placing a marker above the presumed waveforms which deletes the outlying waveforms. This technique is further used to remove extraneous signals from the waveform. Finally, examples of various resultant waveforms are shown in panel 4. Time base is 1 ms in panels 2, 3 and 4. These waveforms are then subjected to auto-correlograms to provide further assurance that they can be designated as single units as evidenced by the single peak (one example in panel 5). Inter spike interval histograms are used to verify fast firing units as single units as evidenced by the 0.5 ms gap at origin (figure 4c). Slow units will have a false gap and are not shown. Further validation that units are single is dependent on functional studies as described below.

Decoding of single unit patterns

Classification Learner (Matlab): The classification learner is a probabilistic neural network. When the input vector arrives at the first layer of the network, the distance from the input vector to the training input vectors is computed. This produces a vector whose elements indicate how close the input is to the training input. All classes of inputs are summed to produce as its net output a vector of probabilities. The final step is for the output of the second layer to use a transfer function to produce a 1 for maximum probability and a 0 for the other classes of inputs which thus generates the classification.

Application: The .txt file containing the single unit firings in 5, 10, 25 or 50 ms time bins is imported into Matlab workspace as a Table. 25 ms bins provide the optimal decoding accuracy. The Classification Learner app is opened by clicking on the app and then clicking on New Session and selecting the imported .txt file. All classifier types are selected and the Train button is activated. Of the many classifiers, the Support Vector Machine (SVM, fine or coarse) and Known Nearest Neighbor classifiers proved to be optimal for decoding these data. Each .txt file (corresponding to production of one phone or word) typically contains 500 ms of single unit firings as determined by the acoustic channel that represents audible speech or by the event marker during the control or silent speech periods. 100 ms of data is selected before the acoustic record or event marker because the neural activity has to begin prior to the speech and most likely falls within 100 ms prior to speech onset.

Approval for the implantation in Belize is obtained from the Belize Committee on Human Experimentation, Belize City, Belize. The participant 5 part of the study is undertaken under IDE G9600032. The study adhered to The Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans http://www.wma.net/en/30publications/10policies/b3/index.html; EC Directive 86/609/EEC for animal experiments http://ec.europa.eu/environment/chemicals/lab_animals/legislation_en.htm; Uniform Requirements for manuscripts submitted to Biomedical journals http://www.icmje.org. The Human Investigation Committee of Neural Signals Inc also approved the study.

Phone classification in the intact participant 6 speaking silently

Classification is performed as described above using Matlab Inc.’s app Classification Learner. All paradigms (including tree, linear discriminant, quadratic discriminant, support vector machine (SVM), known nearest neighbor (KNN) and ensemble), are trained by the system. SVM (fine or coarse gaussian) almost invariably is the most accurate with KNN the next most accurate.

The accuracy of phone classification is illustrated in figure 5a for participant 6 speaking silently. When all single units are included the accuracy is 96.4% (histogram bar ALL on left). The accuracy varies with removal of each single unit in sequence. When the non-firing single units (non-firing in this task) are removed as a group (#1, 5, 11, 12, 13, 14, 15, 18,19), the accuracy remains at 96.4% (third histogram bar from right labelled NO). When all the fastest firing are removed (#3, 4, 8, 22 and 23) the accuracy drops dramatically to 64.3% (second bar from right labelled FAST). Most importantly, when slow firing are removed as a group (#2, 6, 7, 9, 10, 16, 17, 20 and 21 labelled SLOW) the accuracy drops to 80% (final bar on right). Data indicating the total firings for each of the 23 single units over 4.5 seconds are illustrated in figure 5b.

Phone classification in locked-in participant 5 speaking silently

The accuracy of phone classification is illustrated in figure 6a for participant 5 during silent speech. When all single units are included the accuracy is 50.7% (ALL). The accuracy varies with removal of each single unit in sequence. When all the non-firing (in this task) single units (1, 6, 7, 8, 12, 13, 14, 15, 16, 19, 24) are removed the accuracy is virtually unchanged at 49.9% (NO). When all the fastest firing (9, 10, 17, 20, 21, 22 and 23) are removed the accuracy drops dramatically to 22.2% (FAST). Most importantly, when slow firing are removed as a group (2, 3, 4, 5, 11, 18, 25, 26) the accuracy drops to 37.9% (SLOW). Data indicating the total firings for each of the 26 single units over 4.5 seconds are illustrated in figure 5b.

Conditioning of slow firing single units

In light of the above results, it should follow that the slow firing single units have some functional value. If they have value then it should be possible to condition them. In a different set of studies in the locked-in participant 5 over year four after implantation, conditioning of slow firing single units is achieved [12]. This study involves silent singing a tone as described in Methods. The data in figures 7a,b,c are extracted from the data in the 2011 study. We focus here on the slow firing single units only. Illustrated is the result of presenting a 523 Hz frequency tone (high C).

There are 10 separate trials in figure 7a. The firing rate in spikes per second are shown for the listen period, the rest period and then the sing period (10 secs each), followed by a 30 second pause until the next presentation. Clearly, the firing rates (without feedback) of these two slow firing units are random on day 1549 after implantation. The next session, on day 1553, audible feedback of the firing rate is presented to the participant. As figure 7b illustrates, progressing through the session and compared to the previous session on day 1549, the firing rates became much less random. On day 1556, the audible feedback of the firing rate increased in volume as a function of increasing firing rate. Clearly each trial produces a symmetry of the firing rates (figure 7c). Note that the firing rate do not exceed 4 ips.

The data presented above indicate that decoding accuracy is optimal when all single units, slow as well as fast, are included in the analysis. Dropping out individual single units, or groups of single units, reduced the accuracy as shown in figures 5a and 6a. This was more marked for fast firing single units, but also very noticeable for slow firing single units. The importance of this finding is highlighted when considering the clear drops in accurate decoding.

Conditioning of single units implies that single units can be used for functions that are not their original function. Conditioning of fast and slow single units has been achieved [12]. To buttress the present result, an example is extracted from the 2011 paper to illustrate this point and in illustrated in figures 7a,b,c for slow firing single units only. These data illustrate that using auditory feedback, the locked-in participant is able to condition the single unit firings as illustrated in figure 7b and more impressively in figure 7c. It is worth noting that the feedback involved only the audible firing rate and the participant had no visual representation of the result. In fact, neither the researchers nor the participant had any idea that the participant was conditioning the single units while these data were being acquired. We only asked him to sing the tone in his head on a ‘look see’ basis. It was not until the analysis was completed offline that we understood that the single units were being conditioned. The importance of conditioning of slow or fast firing single units is that they can be conditioned to a new task such as vocalization, movement or other tasks. It follows that just a few (5 to 10) single units can be used to decode the phone, word or phrase, as demonstrated by the data above. This implies that even though thousands of units can be used for decoding [1, 2, 3, 4], a large number of units are NOT needed for adequate decoding, though the accuracy improves with involvement of a larger number of units. Clearly, more units improves the resolution, but are not essential for adequate decoding as described here and elsewhere [16, 17].

Slow firing single units are infrequently analyzed in brain computer interfacing research. One reason is that when trying to develop a motor prosthetic, there needs to a driving force to drive velocity, for example, emanating from the neuronal activity and only fast firing units can provide a high velocity related driving force. Slow firing units could only supply a low velocity related driving force that would be of limited usage. For that reason slow firing single units are rarely described. However, they have been described in the hippocampus in relationship to theta rhythmic clock-like activity [19]. The present data strongly suggests they have a role in the accurate decoding of patterns of neural activity in a speech prosthetic. Theoretically, they may also be useful in precise digit movement generation.

Alternative approaches

Undoubtedly, the Blackrock (Utah) microarray and other tine type electrodes have produced remarkably useful single unit firings over months and years [2]. However, 85% of the single units are lost after 3 years [8]. Thus, the Utah array and similar electrodes are not presently suitable as long term recording electrodes, with long term defined as a minimum 50 years. Thus they would not be suitable for restoration of speech over the long term.

Many other workers in the field take a different approach to decoding neural activity. The ECoG approach of various workers [5, 20, 21, 22] involves decoding spoken words and phrases using cortical surface electrical activity, clearly a data set with less resolution compared to single units. Their data indicate the most effective recording resulted from electrodes over the ventral motor speech cortex [22], which is the cortical area targeted for implantation in both the present participants. Conant et al [22] did not show that silent speech, (defined again as speaking “in the head” but not using lips, tongue and jaw, and it is not imagined speech) could be detected because they used movements of the articulators without producing sound (they called it miming) which is obviously not realistic for paralyzed and mute participants. However, a more recent study demonstrates that phrases could be decoded using ECoG electrodes over the primary motor and sensory cortex in a locked-in participant [5]. So ECog recording can be used to decode speech in locked-in participants. The main problem with ECoG, however, is that the ECoG electrodes are not compatible with the cortical surface that always rejects foreign bodies. As described in the introduction, signal is lost in less than two years [9].

Other researchers such as [23, 24] are using EEG recordings allied with deep learning paradigms. It is unlikely that artificial intelligence (AI) will adequately interpret the EEG signals to produce free flowing speech because the resolution with external EEG recordings is too low. Whether or not AI can compensate for this low resolution remains to be seen, but it is unlikely to fully compensate, as recently reviewed [7]. Other techniques include near field infra-red (NFIR) that penetrates through the skull and measures the blood flow within the cortex as used by Chaudhary et al [25]. Again, low resolution is the issue for this technique though useful information is unlikely to be decoded for speech. Other recent developments include tightly spaced multi-electrodes from Neuralink [10]. These are likely to suffer gliosis similar to all other penetrating electrodes with subsequent loss of useful signal [8]. The same gliosis issue will likely be true for the stereotaxic-EEG electrodes [21]. Other approaches include ‘neural lace’ whereby micro devices powered by ultrasound record directly from neural tissue [26]. Longevity has not been published but gliosis may intercede between the device and the neural surface. Efforts to minimize the mechanical rigidity of the electrode and thus minimize rejection are underway [27]. In summary, these data indicate that for human prosthetic lifetime usage there is presently no alternative to the Neurotrophic electrode.

Slow firing single units are just as important as fast firing for optimal decoding accuracy. The ability to condition single units implies that unrelated single units could be made to relate to the decoding task.

Conflict of Interest: PRK has 98% ownership of Neural Signals Inc.

Funded by Neural Signals Inc.

Approval for the implantation in Belize is obtained from the Belize Committee on Human Experimentation, Belize City, Belize. The participant 5 part of the study is undertaken under IDE G9600032. The study adhered to The Code of Ethics of the World Medical Association (Declaration of Helsinki) for experiments involving humans http://www.wma.net/en/30publications/10policies/b3/index.html; EC Directive 86/609/EEC for animal experiments http://ec.europa.eu/environment/chemicals/lab_animals/legislation_en.htm; Uniform Requirements for manuscripts submitted to Biomedical journals http://www.icmje.org.

The Human Investigation Committee of Neural Signals Inc also approved the study. Consent was obtained from participant 5 in the presence of his father and mother by communicating with eye blinks.

Willett FR, Avansino DT, Hochberg LR, Henderson JM, Shenoy KV. High-performance brain-to-text communication via handwriting. Nature. 2021 593(7858):249–254.
Hochberg LR, Bacher D, Jarosiewicz, B, et al. 2012. Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature. 485 (7398), 372-5.
Collinger JL, Wodlinger B, Downey JE et al. High-performance neuroprosthetic control by an individual with tetraplegia. Lancet. 2013 Feb 16;381(9866):557–64.
Jafari M, Aflalo T, Chivukula S, et al. The human primary somatosensory cortex encodes imagined movement in the absence of sensory information. Commun Biol. 2020 Dec 11;3(1):757.
Moses DA, Metzger SL, Jessie RL, et al, Neuroprosthesis for decoding speech in a paralyzed person with anarthria. N Engl J Med, 2021: 385:3.
Kennedy PR Gambrell C, Ehirim P, and Cervantes A. Advances in the development of a speech prosthesis. Book chapter in Brain-Machine Interfaces: Uses and Developments. 2017
Kennedy PR. To invade or not to invade, that is the question for brain computer interfacing. Current Trends in Neurology, 2018 12:31–38.
Downey JE, Schwed N, Chase SM et al. Intracortical recording stability in human brain-computer interface users J Neural Eng., 2018 15 (4): 046016.
Degenhart A D, Eles J, Dum R, et al. Histological evaluation of a chronically-implanted electrocorticographic electrode grid in a non-human primate. J. Neural Eng. 2016, 13:046019.
Musk, E. An integrated brain-machine interface platform with thousands of channels. BioRxiv, 2019, 18:1–12.
Kennedy PR, Andreasen DS, Bartels J, et al. Making the lifetime connection between brain and machine for restoring and enhancing function. Proceedings in Brain Research, 2011, CH.1.
Kennedy PR. Changes in emotional state modulate neuronal firing rates of human speech motor cortex: A case study in long-term recording. Neurocase, 2011 17(5), 381–393. **
Kennedy P.R., Andreasen D.S., Bartels J., Ehirim P., E. Joe Wright, E.J., Seibert, S., Cervantes, A.J. 2018. Validation of Neurotrophic Electrode long term recordings in Human Cortex. Handbook of Biomedical Engineering.
Guenther FH, Brumberg JS, Wright EJ, et al. A wireless brain-machine interface for real-time speech synthesis. PLoS One. 2009 4(12):e8218.
Brumberg J., Wright EJ, Andreasen D, et al. Classification of intended phoneme production from chronic intracortical microelectrode recordings in speech motor cortex. Frontiers in Neuroscience 2011 5(65)1–14.
Brumberg JS, Nieto-Castanon A, Kennedy PR, et al. Brain-Computer Interfaces for Speech Cmmunication. Speech Commun. 2010 Apr 1;52(4):367-379.
Gearin M and Kennedy PR. Histological confirmation of myelinated neural filaments within the tip of the Neurotrophic Electrode after a decade of neural recordings. Frontiers in Human Neuroscience 21 April 2020 <background-color:#CFBFB1;uvertical-align:super;>https://doi.org/10.3389/fnhum.2020.00111</background-color:#CFBFB1;uvertical-align:super;>
Kennedy PR, Mirra S, and Bakay RAE. The Cone Electrode: Ultrastructural Studies Following Long-Term Recording. Neuroscience Letters, 1992 142: 89–94.
Zhang Lu, Ma Xiaoyu, Chen Guifen, Barkai Edi and Lin Longnian. Theta rhythmic clock-like activity of participant in the mouse hippocampus. J Neurosci, 2016, 20; 36(16): 4415–4420.
Makin JG, Moses DA, Chang EF. Machine translation of cortical activity to text with an encoder-decoder framework. Nat Neurosci. 2020 23(4):575–582.
Herff C, Krusienski DJ and Kubben P. The potential of stereotaxic-EEG for brain-computer interfaces: Current progress and future directions. Front. Neurosci. Feb 27, 2020.
Conant., D, Bouchard, K.E. and Chang E.F. Speech map in the human ventral sensory-motor cortex. Current Opinion Neurobiology. 2014, 24 (1), 63–7.
Lawhern V, Solon A, Waytowich N, Gordon SM, Hung C, Lance BJ. EEGNet: a compact convolutional neural network for EEG-based brain--computer interfaces. J Neural Eng. 2018, 22, 1741–2552.
Balaji A, Haldar A, Patil K, Ruthvik TS, Ca V, Jartarkar M, Baths V. EEG-based classification of bilingual unspoken speech using ANN. Conf Proc IEEE Eng Med Biol Soc. 2017, 7:1022–1025.
Chaudhary U, Xia B, Silvoni S, Cohen LG, Birbaumer N. Brain-Computer Interface-Based Communication in the Completely Locked-In State. PLoS Biol. 2017, 31;15(1).
Neely RM, Piech DK, Santacruz SR, Maharbiz MM, Carmena JM. <background-color:#CCCCFF;uvertical-align:super;>Recent advances in neural dust: towards a neural interface platform</background-color:#CCCCFF;uvertical-align:super;><uvertical-align:super;>.</uvertical-align:super;> Curr Opin Neurobiol. 2018 Jun;50:64–71. doi: 10.1016/j.conb.2017.12.010. Epub 2018 Jan 11.
Zhengtuo Zhao, Xue Li, Fei He, Xiaoling Wei, Shengqing Lin, Chong Xie. Parallel, minimally-invasive implantation of ultra-flexible neural electrode arrays. J Neural Eng. 2019 Jun;16(3):035001. doi: 10.1088/1741-2552/ab05b6. Epub 2019 Feb 8.
Bartels J, Andreasen D, Ehirim P, Mao H, Seibert S, Wright EJ, Kennedy PR. Neurotrophic electrode: method of assembly and implantation into human motor speech cortex. J Neurosci Methods. 2008 Sep 30;174(2):168–76. Epub 2008 Jul 10.

Competing interest reported. Dr. Kennedy's work has been funded by Neural Signals Inc. He owns 98% of the stock in Neural Signals Inc. Ms. Ganesh and Dr. Cervantes declare no potential conflict of interest.

APPENDIXSlowfiringunits.docx

PDF

Version 1

posted 03 Jan, 2022

You are reading this latest preprint version

Slow Firing Single Units Are Essential for Optimal Decoding of Silent Speech

Status:

Version 1

Abstract

Figures

Introduction

Strategy

Methods

Electrode

Implanted Electronics

Implantation Target Site

Surgery

Recording

Paradigms

Spike Sorting

Decoding of single unit patterns

Results

Phone classification in the intact participant 6 speaking silently

Phone classification in locked-in participant 5 speaking silently

Conditioning of slow firing single units

Discussion

Alternative approaches

Conclusion

Declarations

References

Competing Interests

Supplementary Files

Status:

Version 1