September - December 2023 Astrocyte* newsletter is here. I've bundled several months to clear my accumulated backlog. In February, I will get back to the standard monthly format.
This is a pretty long issue, so I've added a hyperlinked table of contents to make it easier to read.
Below are also a few things I'm working on. Subscribe to get these notes and a monthly newsletter into your mailbox.
Soon to be published:
- A summary/reflection note on the Royal Society Neural Interfaces Summit 2023 - November - moved to January 2024;
- A note on Synchron, an endovascular brain-computer interface - November -moved to February 2024;
- An updated version of the BCI startup landscape - December - moved to March 2024.
Please share and subscribe if you liked it.
Table of Contents
I. Reading List
- Non-Invasive/Minimally-Invasive - ultrasound rules this issue (BCI, biopsy, etc.), temporal interference, in-ear EEG.
- Invasive - speech sound encoding via Neuropixels, and a protocol to work with them in the operating room, ECoG, the transsulcal approach to microchip implantation using an existing surgery tool.
2) Data Science - real-time deep learning for BCIs, self-supervised learning on a large quantity of speech data (EEG, MEG), foundation model/LLM for EEG.
3) Misc Neurotech Reading - physics of BCIs, do we need more bandwidth, what is BCI?
1) Funding - startups working on ultrasound, EMG/SNC, fNIRS, EEG, and some invasive modalities raised funding during covered months
2) Other Announcements - examples of miniaturisation and vertical integration in the BCI domain.
Twitter handles of the authors provided after some papers are not in a particular order; based on my ability to identify authors by handles, let me know if I missed anyone.
I. Reading List 📚
Ultrasound is interesting. I first came across ultrasound for BCIs via this article about the work of @TanterM, @SumnerLN, and @mikhailshapiro (included in the April 2021 newsletter). It's fantastic to see the progress since then, which is demonstrated in Decoding motor plans using a closed-loop ultrasonic brain-machine interface. The paper presents a successful implementation of a closed-loop ultrasonic brain-machine interface.
Functional ultrasound (fUS) enables epidural BMIs that can record from large brain regions and decode spatiotemporally precise patterns of activity. fUS senses changes in cerebral blood volume (CBV) that correlate well with single-neuron activity and local field potentials. However, it requires a cranial opening or an acoustic window in large animals.
That's what has been done:
We streamed fUS data from the posterior parietal cortex of two rhesus macaque monkeys while they performed eye and hand movements. After training, the monkeys controlled up to eight movement directions using the BMI. We also developed a method for pretraining the BMI using data from previous sessions.
Interestingly, a decoder that processed the data and generated control signals to move a cursor to where the monkeys intended did not require recalibration.
In the previous newsletter, I mentioned a paper on how US helps deliver chemotherapy into the brain. The Funding section of this edition mentions Carthera, a company that builds a device for using ultrasound for drug delivery.
Another application of ultrasound is mentioned in First-in-human prospective trial of sonobiopsy in high-grade glioma patients using neuronavigation-guided focused ultrasound . The paper covers a new approach to obtaining information on a brain lesion without surgery.
Current approaches have limits: lumbar puncture for cerebral spinal fluid-based liquid biopsy is uncomfortable and carries procedural risk, and the blood-based liquid biopsy detects tumour biomarkers generally only at low abundance, primarily due to the blood-brain barrier (BBB). An alternative is the FUS-mediated BBB opening to allow brain tumour-derived biomarkers to be released into the bloodstream for diagnostic access.
A great explainer of how it works is here:
The procedure works by using focused ultrasound to target a lesion in the brain with millimeter-scale accuracy, followed by the injection of microbubbles into the bloodstream. The microbubbles travel to the targeted spot and then pop, tearing tiny holes in the blood-brain barrier that close within a few hours, leaving no lasting damage. That window of time is long enough for biomolecules from the lesion to escape into the blood, where they can be collected with an ordinary blood draw.
Transcranial focused ultrasound-mediated neurochemical and functional connectivity changes in deep cortical regions in humans demonstrates how low-intensity transcranial ultrasound stimulation (TUS) is used for neuromodulation:
In 24 healthy controls, we separately stimulated two deep cortical regions and investigated the effects of theta-burst TUS... We show that theta-burst TUS in humans selectively reduces GABA levels in the posterior cingulate, but not the dorsal anterior cingulate cortex. Functional connectivity increased following TUS in both regions.
However, with ultrasound, '... the expansive parameter space and difficulties in controlling for peripheral auditory effects make it challenging to identify ultrasound sequences and brain targets that may provide therapeutic efficacy.'
To better understand '... how different parameters including pulsing frequency, duty cycle, stimulus duration, and acoustic intensity contribute to the elicited effects', researchers created a '... rat-wearable ultrasound device with electronic steering capabilities for high-throughput, multi-target neuro-interventional investigations. See High-throughput ultrasound neuromodulation in awake and freely behaving rats.
An overview of '...the current knowledge about the neuromodulatory effects of TUS while discussing the potential of TUS in the field of stroke recovery' is provided in Low-Intensity Focused Ultrasound Neuromodulation for Stroke Recovery: A Novel Deep Brain Stimulation Approach for Neurorehabilitation?
Non-invasive temporal interference electrical stimulation of the human hippocampus. A strategy for sculpting electrical fields is called temporal interference (TI) stimulation since the interference of multiple electric fields enables its focality. 'The strategy is about '... using multiple kHz-range electric fields with a difference frequency within the range of neural activity.'
Here we report the validation of the non-invasive DBS concept in humans. We used electric field modeling and measurements in a human cadaver to verify that the locus of the transcranial TI stimulation can be steerably focused in the hippocampus with minimal exposure to the overlying cortex. ... We then use neuroimaging and behavioral experiments in healthy humans to demonstrate focal non-invasive modulation of hippocampal memory activity and the capacity to augment memory performance.
This paper is among a few I've seen that use a new approach to work with memory, a theme overlooked, at least in the startup community.
When parameters typically used in other non-invasive tES experiments were applied, TI improved memory accuracy. 'The magnitude of the memory improvement was small...', however.
In July's issue, I've covered Apple ramping-up sensing capabilities of its earphones and also the research on in-ear bioelectronics for visual and auditory brain-computer interfaces. In-ear integrated sensor array for the continuous monitoring of brain activity and of lactate in sweat presents the following system:
... an unobtrusive and fully in-ear integrated array of multimodal electrophysiological and electrochemical sensors for simultaneous monitoring of the brain state and dynamic metabolic sweat concentration. ... The first set of features implements... electrophysiological signals, such as EEG and electrodermal activity (EDA). The second set of features implements electrochemical analysis of metabolites in the ear... lactate in sweat was selected as the analyte of choice.
Invasive 🧑⚕️ 🏥
Using Neuropixels - single-cell resolution silicon-based electrophysiology-recording probes - is getting traction. In Large-scale single-neuron speech sound encoding across the depth of human cortex high-density arrays '... were used to record [from eight participants] from 685 neurons across cortical layers at nine sites in a high-level auditory region that is critical for speech, the superior temporal gyrus'. The research found that:
Across the depth of cortex, the neuronal population is tuned to a dominant speech feature... At the same time, a relatively large proportion of neurons throughout the vertical cortical column also encode a large variety of other speech features, revealing a distinct, previously unappreciated dimension for speech encoding.
Exciting work on modifying Neuropixels 1.0 probes to better work in the human cortex and also documenting the protocol of using them in the operating room - Modified Neuropixels probes for recording human neurophysiology in the operating room. Legendary Neuropixels were modified: increased shank thickness, additional stress-compensation layers, and sharpened tips.
Our team demonstrated the first-in-human use of these probes during resection surgery for epilepsy or tumours and deep brain stimulation electrode placement in patients with Parkinson’s disease... This Protocol describes the structural reinforcement of Neuropixels 1.0-S [designed for primate models], ensuring their correct insertion into the human cortex; the extensive perioperative procedures required to maintain sterile conditions; and coordination with the surgical team.
Stable decoding from a speech BCI enables control for an individual with ALS without recalibration for 3 months. The paper focuses on how advancements in the BCI tech, ECoG specifically, translate into helping individuals directly control smart devices to perform activities (turning on lights in the home) via decoded speech and whether time spent on retraining and recalibrating the speech decoders can be reduced. The study found that:
Speech commands [via a chronic ECoG implant over the ventral sensorimotor cortex] are accurately detected and decoded (median accuracy: 90.59%) throughout a 3-month study period without model retraining or recalibration.
Reducing or eliminating the need to recalibrate decoders is an important step towards unassisted home usage of BCIs.
The signals trial is ongoing and is now recruiting two more participants at Johns Hopkins and two at the partnering Utrecht University in the Netherlands. Nathan Crone, who is conducting the research as part of the CortiCom clinical trial, co-authored a review of the best-performing BCIs that I've included in the previous newsletter. Crone will act as principal investigator in another NIH-funded trial on a new type of MEA device, so follow the space.
High-resolution neural recordings improve the accuracy of speech decoding. Motivated by previous studies that confirmed that '... Neural speech decoding using 4-mm-spaced arrays showed up to a 5× increase in phoneme prediction compared to 10-mm-spaced arrays...', the researchers fabricated µECoG arrays with even higher resolution.
... In the present work, we demonstrate the use of high-density µECoG for speech decoding in humans [3 movement disorder patients and 1 tumor patient]. We recorded intra-operatively from speech-abled patients using liquid crystal polymer thin-film (LCP-TF) µECoG arrays (1.33 – 1.72 mm inter-electrode distance, 200 µm exposed diameter electrodes) placed over SMC during a speech production task.
Subcortical implantation of a passive microchip in rodents -an observational proof-of-concept study. Microchips are the bedrock of modern tech. This technology will affect BCIs in the same, or even more profound way, as it affects personal computing. It explores a novel way of implanting microchips.
... we conducted an observational, proof-of-concept study to evaluate the use of a modified BrainPath device to implant subcortical microchips in a rodent model. All animals survived the procedure and lived with a microchip in situ for one or three months with no apparent neurological effects... The BrainPath is a tubular retraction system that uses a transsulcal approach to exploit naturally existing anatomical corridors using intraoperative navigation. Using a parafascicular, fiber-sparing approach, the device [was] inserted in parallel to white matter tracts, displacing rather than dissecting or transecting tissue or fiber tracts, thereby minimizing tissue disruption...
Among other things, two items excite me about this study - 1) microchips and 2) using an existing tool.
Microchip manufacturing techniques look promising. If the implantation challenge is solved, then adopting microchip approaches to BCIs will be facilitated. The study used the Rogers Research Group chips at Northwestern University, Evanston Illinois.
Earlier, researchers from this group redesigned the circuitry of an electrocorticography device to make microchip manufacturing techniques applicable to it. The aim was to use more electrodes to boost the chances of detecting smaller signals and encase the system to lengthen the life of a device.
This neurosurgical tool used to implant the chip is widely adopted to safely access and remove subcortical lesions such as tumours and hematomas. As Syncrhon demonstrated with stents, repurposing existing tools for BCI implantation is a great idea that may positively affect their adoption (less additional training for surgeons, fewer costs for tools purchasing, better understanding of risks, etc.).
The mismatch between soft living tissues and stiff implants challenges invasive modalities. Biohybrid neural interfaces: improving the biological integration of neural implants describes various types of tissue response to neural implants and outlines '... experimental strategies in bioelectronics research aimed at engineering the device–tissue interface through the integration of biomaterial- and cell-based strategies'.
Device-tissue interface strategies: surface functionalisation with bioactive molecules, bioactive coatings, controlled release of drugs and growth factors, implant architecture, surface topography, device form factor, and strategies of functionalisation of interfaces with live cells.
Data Science 💻🧮
Applying deep learning to data from BCIs is hard; among other things, one must enable real-time and efficient processing and handling of randomly missing brain signals (which happens in wireless BCIs). To enable 'flexible inference causally, non-causally and in the presence of missing neural observations', Dynamical flexible inference of nonlinear latent factors and structures in neural population activity presents:
a neural network that separates the model into jointly trained manifold and dynamic latent factors such that nonlinearity is captured through the manifold factors and the dynamics can be modelled in tractable linear form on this nonlinear manifold.
A large EEG database with users’ profile information for motor imagery brain-computer interface research. One of the approaches to the BCI is through reading how a user is performing motor imagery tasks, e.g., imagining hand movements. However, 10–30% of naive users cannot control the BCI via this language. There are also problems with recognition accuracy. To '... understand, model and optimize user’s training..' for using motor imagery-based BCIs, a dataset was collected. 87 participants, more than 20,800 trials, or approximately 70 hours of recording time. The database includes detailed information about the demographics, personality profile, cognitive traits, experimental instructions, and codes.
Here we introduce a model trained with contrastive learning to decode self-supervised representations of perceived speech from the non-invasive recordings of a large cohort of healthy individuals. 175 volunteers recorded with magneto-encephalography or electro-encephalography while they listened to short stories and isolated sentences. The results show that our model can identify, from 3 seconds of magneto-encephalography signals, the corresponding speech segment with up to 41% accuracy out of more than 1,000 distinct possibilities on average across participants...
Interesting steps towards non-invasive speech decoding, an attempt to train a model on a large cohort, not an individual patient, using self-supervised learning on a large quantity of speech data (not limited set of interpretable features like Mel spectrogram, letters, phonemes, a small set of words). To achieve mass adoption, neurotech companies must find a way to work with data from non(less)-invasive modalities and to minimise training on specific patient/user data.
Neuro-GPT: developing a foundation model for EEG. This attempts to develop a foundation model on a large dataset (almost 11K subjects) and then fine-tune it on a small dataset for a downstream task (9 subjects).
Neuro-GPT is a foundation model consisting of an EEG encoder to extract spatio-temporal features from EEG data, and a GPT model that uses self-supervision to predict the masked chunks. The foundation model is pre-trained on the . We then fine-tune the model on a Motor Imagery Classification task where only 9 subjects are available.
DeWave: discrete EEG waves encoding for brain dynamics to text translation. An interesting approach to '...bridge the gap between the brain and languages' via the power of LLMs and avoid using eye-tracking fixations or event markers to segment brain dynamics into word-level features.
... we introduce a novel framework, DeWave, that integrates discrete encoding sequences into open-vocabulary EEG-to-text translation tasks. DeWave uses a quantized variational encoder to derive discrete codex encoding and align it with pre-trained language models.
Some technical concerns about this paper by @JeanRemiKing:
Misc Neurotech Reading 👨💻
Plug me in: the physics of brain-computer interfaces - a great overview of how invasive and non-invasive modalities work from the physics perspective by @physp. It covers near-infrared light, magnetic fields, high-frequency sound waves, and other phenomena neurotech device developers use.
Elon Musk wants more bandwidth between people and machines. Do we need it? By @antonioregalado. A brief answer is 'Yes' and 'No'. More bandwidth is less likely required to enhance communication between nondisabled people unless brain processing speed increases proportionally. It will likely be required for restoring speech, as more bandwidth will allow reading from more neurones and restore with higher accuracy. There are also use cases where bandwidth may allow communication in alternative modalities, e.g., via images or exchanging emotions.
Old but gold. A summary of Facebook Realty Labs component of the Facebook Connect Keynote 2020 that covers CTRL Labs (an EMG interface developer acquired by FB in 2019). There is also an interesting discussion in commentaries about defining a BCI.
Startup/Corporate News 📰💰
Funding announcements from September to December 2023 include companies operating across various non-invasive and invasive modalities, including ultrasound.
🇺🇸 Prophetic raised a $1.1 million funding round - it develops a wearable device for stabilizing lucid dreams via focused ultrasound signals.
🇺🇸 🇫🇷 Carthera, an innovative ultrasound-based medical device to treat a wide range of brain disorders, announced today an additional €4.5M in funding to complement its Series B financing round. The device emits ultrasound to temporarily increase the permeability of the blood vessels in the brain to improve the delivery of therapeutic molecules.
🇺🇸 Sonera secures $11M in seed funding. The first application of Sonera's technology is the S1 chip for muscle sensing, which measures magnetic fields produced by electrical currents from neural activity.
I'm curious how Sonera will stack against incumbents. For example, Mudra uses surface nerve conductance (SNC) to control the Apple Watch and offers it as an SDK. The company behind Mudra was founded in 2014, and recently Wearable Devices announced closing of $2.0 Million underwritten public offering 🇺🇸 🇮🇱.
🇺🇸 🇮🇱 Arctop, a software platform for brain decoding, announced today $10 million in Series A funding. 'Arctop decodes brain signals measured from wearables like headbands, earbuds, and virtual and augmented reality headsets equipped with basic electric skin sensors. ... software is licensed to enterprise developers who are creating unprecedented user-centric applications, such as personalized skill training, assistive communication, and emotion-adaptive experiences.' Multimodal, incl. EEG.
🇺🇸 Cognixion raised $1.9M via debt financing. It aims to build an EEG-powered device that allows fully paralyzed or locked-in patients to communicate.
🇺🇸 CenSyn secures NSF SBIR grant and Arizona State funding to accelerate commercialization of innovative neuro health medical platform. The platform includes proprietary cloud software and a two-channel EEG sensor that enables 30-second EEG recordings.
🇺🇸 BrainScope receives investment from the Alzheimer's Drug Discovery Foundation. BrainScope is a medical neuro-technology firm utilizing EEG technology to assess the full spectrum of mild traumatic brain injury.
🇬🇧 FC Labs secures £150k investment for its AI-powered solution to reducing workplace accidents. It aims to reduce the risk of workplace accidents and fatalities caused by human error by spotting the tell-tale signs that an employee may be at risk through monitoring trends in their brain activity. It develops a wearable device that uses near-infrared light to measure blood flow and oxygen levels (fNIRS?) in the prefrontal cortex area of the brain, which helps regulate focus, actions and emotions.
🇺🇸 Kernel raised $5.2M - likely a bridge funding round, TD-fNIRS headset-based neuroimaging platform to accelerate treatment development, improve patient outcomes, and reduce healthcare costs.
🇨🇦 Ontario Brain Institute announces $100K grant to Axem Neurotechnology. The device uses functional near-infrared spectroscopy (fNIRS) to measure changes in brain activity during movement to allow healthcare professionals to quickly obtain brain activity measurements that can be used to track motor recovery and guide treatment in stroke patients.
🇫🇷 Injectsense and Injectpower raise $9.4m for miniature implantable devices - an enabling technology - ultra-miniature batteries. that can be used in neurotech. One of the applications is micro-batteries for a device embedded in a shunt that will measure intracranial pressure (ICP).
🇪🇸 Inbrain Neuroelectronics obtains a loan of 20 million euros from the European Investment Bank (EIB). Inbrain '...is developing a neural graphene interface [iEEG] that can detect more biomarkers and do so more invasively than the current metal devices on the market for the treatment of central nervous system (CNS) diseases.' It was a bridge round '... for the round it is preparing for 2024.'
There are a couple of interesting articles about scaling neurotech, covering the miniaturisation of the devices required for mass adoption and increasing the capacity of production facilities.
🇺🇸 Niura’s EEG-implemented earbuds scan your brain health and recommend music to your mood. By @kateparknews: Interesting benchmark on shrinking EEG to fit earbuds:
the company has progressively shrunk it with NeuralONE (30 x 30 mm), “a chip encompassing the EEG reader, audio data and data processing parts,” then NeuralTWO (22 x 22 mm) and NeuralTHREE (20 x 12 mm), eventually compressing the PCB to fit into regular earbuds.
Our business model revolves around licensing our product, ensuring partnerships maintain the highest standards of security. This strategic collaboration with leading companies also allows us to advance our research collectively.
🇺🇸 Beacon Biosignals receives FDA 510(k) Clearance for the Dreem 3S, a wearable EEG headband with integrated ML algorithms to capture data from the brain to monitor sleep architecture and aid in the diagnosis of disturbed sleep.
🇨🇭 Neurosoft Bioelectronics’ subdural brain interface was used for recording the human brain in the context of an epilepsy resective surgery. Two epileptic patients and one brain tumour patient have undergone a successful surgery, while detailed recordings were obtained in parallel from the surface of the brain (ECoG).
🇺🇸 Precision Neuroscience buys a factory to build its brain implants - to me, that signals the soon-to-be-expected wider adoption of BCIs. The capacity uplift is significant - previously '... Precision worked with a facility that took over a year to manufacture six arrays, and now, the company can manufacture more than 100 arrays in a single week'.
Another rationale for owning the factory is that it allows it to move faster, given that the competition with other BCI developers heats up. Owning a factory '... allows us to iterate really quickly, improve performance, longevity, different form factors of the device...'.
Precision is not the first. Science, another BCI heavyweight, acquired a MEMS foundry facility in 2022. The reasons behind vertical integration: 'The sophistication of the devices we can make is directly limited by the tools we can access, and there’s nothing like walking down the hall for iteration speed.' Science also promised to externalize the in-house MEMS capabilities '... as a commercial platform for a limited group of like-minded customers.'.
📬📬📬 Subscribe for a monthly update on neurotech and neurotech investment/commercialisation activity.