What they do and how machine learning fits in
‘…physics and neuroscience are in some ways the most fundamental subjects: one is concerned with the external world out there, and the other with the internal world in our minds’.
The development of technologies that study and affect the ‘internal world in our minds’ is fuelled by investment activity, among other things. In the summer of 2016, CB Insights, an investment database, published a review of 17 startups that boost the brain. In just two years, in June 2018, Neuronetics, the most well-funded startup from the list, went public. Other companies from that list raised substantial investment rounds. For example, Lumosity ($11M), Headspace ($32M), Thync ($6M) and others. Only one startup went out of business reportedly. That, what is happening at the intersection of tech and neuroscience looks exciting.
This article represents a wider and deeper overview of startups that hack the brain*. It also highlights how machine learning (ML) is/may be applied to this grand challenge. I hope this article will be useful for data scientists, who are thinking about where to apply their expertise, as well as for researchers and healthcare professionals, who want to identify technologies beneficial for their research/patients.
In this article I divide 44 startups into three groups: 1) companies that provide diagnostics capabilities, 2) those who build tech for various interventions and affect/stimulate brain, and lastly, 3) companies who contribute predominantly to brain research and development of brain interfaces. See chart 1.
Within each group, startups are aligned around the core principles that their tech is based on. For example, measuring blood flow, tracking electrical activity of the brain, or testing for certain proteins. In the Appendix, you may find information on how this data was collected.
Electroencephalography(EEG), a technology to record electrical activity of the brain, is one of the most popular diagnostics tools. Among startups that build on EEG are:
- BrainScope develops assessment of brain injury, including concussion, Ceribell designs a tool for rapid setup and triage of seizures, and ElMindAis helping to recover after brain injuries;
- NeuroSky’s EEG biosensor digitises and amplifies raw analogue brain signals and offers a platform for applications in health and wellness, education, medicine, research and other sectors.
A novel way of diagnostics is based on analysing human-smartphone interaction. Mindstrong relies on ‘…a set of digital biomarkers from human-smartphone interactions that correlate highly with select cognitive measures, mood state, and brain connectivity’.
Among various diagnostics approaches, there are approaches that are measuring fluid volumes, blood flow and oxygen levels, properties of a tissue, and even motions of the skull. Startups that apply these approaches are:
- Cerebrotech Medical Systems develops volumetric impedance phase-shift spectroscopy (VIPS). It develops a headset that uses multi-frequency electromagnetic measurements to assess fluid volume differences between the cerebral hemispheres. Its tech helps with assessment of stroke patients;
- Neural Analytics explores transcranial Doppler sonography, an ultrasound technique, for brain health assessment. ‘It combines an all-in-one neurovascular ultrasound device, designed to non-invasively measure and display brain blood flow information’. It is used for evaluation and management of patients with cerebrovascular diseases; Boston Neurosciences also uses Doppler ultrasound to non-invasively measure intracranial pressure (ICP);
- A very different approach to ultrasound is developed by Iota Biosciences. A millimetre-sized device‘…is activated by a beam of ultrasound, voltage runs between the electrodes, and this minute current is affected by the electrical activity of the tissue. These slight changes are literally reflected in how the ultrasonic pulses bounce back, and the reader can derive electrophysiological voltage from those changes’. The technology is not directly applied to the brain at the moment, but allows an ‘…interface directly with specific nerve clusters’ and ‘…opens up many doors for bioelectronic medicine and brain-machine interfaces’;
- Near-infrared spectroscopy is used to diagnose brain injuries based on oxygen concentration in blood, for example by Luciole Medical and Obelab;
- Measuring skull motion, that emerges as a result of unsymmetrical blood flow into the brain, and disruptions of such motion is the approach of Jan Medical. Their headset contains ‘…a heart rate detecting sensor, a sound pressure level sensor for detecting ambient environment noise and six accelerometers to detect the acceleration of the skull’. It helps to diagnose urgent conditions including concussion, stroke and vasospasm.
Blood tells a lot about brain health, and some startups study it to diagnose injuries and other conditions. For example:
- BioDirection develops a testing platform to detect and measure specific protein molecules released by the brain immediately after a concussion;
- Iron Horse Diagnostics develops blood or cerebrospinal fluid tests for earlier identification of brain injuries and amyotrophic lateral sclerosis.
Machine learning for diagnostics
It appears that brain diagnostic applications rely heavily on machine learning. The core of EEG software is interpretive models that help to identify and quantify categories of mental or emotional states. These algorithms ‘…can be as simple as a mathematical formula to as complex as a machine-learning model that maps to users’ personal opinions about how they feel while engaged in a given activity’.
Machine learning also allows skull motion patterns to be recognised and associated with various brain pathologies. Algorithms ‘…differentiate between the bioimpedance profiles of various brain pathologies’.
On top of that, machine learning methods are used ‘…to show that specific digital features [of human-smartphone interaction] correlate with cognitive function, clinical symptoms, and measures of brain activity’. Therefore, a new way of diagnosis becomes available.
Startups go beyond diagnostic and build tech for affecting the brain. I divide various types of interventions into two large groups, namely:
- Technology-driven, when different types of stimulation are applied to the brain by a device. Nerves and the brain itself may be stimulated by electrical impulses, magnetic waves, low temperatures and even light. These and other stimuli affect various parameters of the brain, for example, blood flow, release of neurotransmitters, etc.;
- Technology-enabled, where a patient/user tries to manage his/her state by changing behaviour, meditating or applying other practices that do not involve stimulation by a device. These interventions may be enabled by neurofeedback, a type of biofeedback that ‘…uses real-time displays of brain activity to teach self-regulation of brain function’.
II. 1. Technology-driven interventions
Affecting blood flow, temperature
- Non-invasive stimulation of trigeminal nerve allows to decrease or increase blood flow in brain regions associated with initiation and spread of epileptic seizures and regions associated with mood, attention and executive function respectively. NeuroSigma uses this approach;
- BrainsGate’s technology is based on electrical stimulation of the sphenopalatine ganglion (SPG) that increases cerebral blood flow. In a case of an acute ischemic stroke blood circulation is compromised. SPG stimulation can increase perfusion to the areas suffering from reduced or lack of blood supply, and help to save brain tissue;
- Ebb develops frontal cerebral thermal therapy (maintained at 14–16°C) for insomnia treatment. Sleep disturbances have been shown to correlate with frontal metabolism during sleep, reduction in metabolism minimises disturbances.
Affecting release of some chemicals, e.g. hormones, neuromodulators, neurotransmitters
- Trigeminal nerve stimulation helps to meditate the brain’s major sympathetic responses to environmental stressors. The stimulation mobilises norepinephrine, a hormone that mobilises the brain and body for action, from locus coeruleus(a cluster of neurons responsible for stress and panic reaction). Thync used this approach to reduce stress;
- Invasive stimulation of vagus nerve (VNS) passes to certain brain regions and leads to a release of neuromudilators that ‘… are important for learning and memory and help increase the salience (or relevance) of the physical therapy [of those who experienced stroke]’. ‘The cholinergic and noradrenergic neuromodulatory systems are engaged by vagus nerve stimulation and represent potential pathways through which VNS may support recovery’. Microtransponder develops a VNS tech;
- Transcranial magnetic stimulation (TMS), and synchronised TMS (personalised to a patient’s individual alpha frequency, as measured by a brief EEG) are used to help patients with depression. NeoSync works on synchronised TMS; TMS provokes a release of neurotransmitters, e.g. dopamine and normalises brain chemistry. It is also considered that TMS may restore normal oscillatory patterns which enablen coordinated functions across brain regions;
- Responsive neuro stimulation (RNS) developed by NeuroPace is ‘…thought to act on an inhibitory neurotransmitter. This type of substance acts to inhibit or stop activity from brain cells that could lead to seizures’;
- Brain chemistry may also be affected by light. When ‘…photosensitive areas of the brain are stimulated by light, it affects the neural circuits in the brain via neurotransmitters (e.g. serotonin, dopamine and noradrenaline)’. Transcranial bright light exposure via the ear canals was performed by Human Charger, that develops a depression treatment. However, this approach was criticised.
Other types of stimulation
- Deep brain stimulation (DBS) does not act directly on dopamine producing cells and does not affect brain dopamine levels. Instead, it compensates for one of the major secondary effects of dopamine loss, it corrects excessive and abnormal activity provoked by the loss of dopamine-producing cells. Aleva Neurotherapeutics applies DBS to subthalamic nucleus as well as in the nucleus ventralis intermedius of the thalamus. Functional Neuromodulation is researching the use of a surgically implanted device for stimulation of fornix, a place in the brain that plays a central role in memory;
- Nativis Voyager, a device that uses ultra-low radio frequency energy, ‘…records the electronic frequency profile of a molecule, in this case a chemotherapy called Taxol. Then … ‘plays back’ the signal at a patients’ tumor. In theory, the signal impacts the cancer cells the same way molecules of Taxol would’. Taxol works ‘… by damaging the RNA or DNA that tells the cell how to copy itself in division. If the cells are unable to divide, they die’. You could read critics of Nativis’ approach here;
- Electrical stimulation ‘…amplifies the residual commands from the brain, enabling voluntary control over specific leg muscles in people with severe spinal cord injury’. What happensis ‘…a massive reorganization of both the relevant neural circuits in the spinal cord and residual pathways towards/from the brain’. Experiments showed that neuromodulation leads to rerouting of cortical information through new pathways in a case of spinal cord contusions. GTX Medical works on lumbosacral spinal cord stimulation for rehabilitation;
- Halo Neuroscience uses transcranial direct current stimulation (tDCS), i.e. applies very low levels of constant electrical current to the brain area of interest via electrodes on the scalp. That increases ‘…brain’s natural ability to generate and optimise new neural pathways in the motor cortex’ and improve ‘muscle memory’.
II.2. Technology-enabled interventions: biofeedback and other techniques
- EEG-powered: Muse‘…translates brainwaves into the guiding sounds of weather’ and helps to meditate. Dreem uses neurofeedback as a relaxation technique. The level of relaxation, measured by the shape of brain waves, is also translated into a sound signal. User tries to affect the sound and eventually relaxes. Emotiv helps to learn what time of a day users are most focused or what activities make them be the most relaxed. It also recommends strategies for managing stress or improving focus. BrainCo ‘… helps students to enhance their abilities to focus and study efficiently. Users can track their attention scores and play cognitive training games…’. Atentiv also provides EEG-based cognitive training and therapy in a form of games;
- Augmented/virtual reality and wearable devices/sensors also are applied to tinkering with brain function. For instance, these technologies allow movements of a patient in the real world being reproduced as movements of an anthropomorphic avatar in the virtual environment. Embodied sensorimotor feedback allows ‘…the integration of motor priming techniques and cognitive principles related to body perception and action’, ‘…which have been shown to improve functional recovery’. MindMaze applies VR for rehabilitation of stroke patients, for example;
- Mobile phone interaction: BrainHQ shares feedback on how users are playing games. Lumocity shares insights on training that users undergo playing cognitive mobile games. Happify provides general feedback in a form of a ‘happiness score’. It is not yet clear how significant is the impact of these games on brain training, however there are evidences that ‘…brain’s neural timing improves with training, with neurons responding more quickly to speech’, for example.
Training self-awareness is another way how technologies can help to manage brain function
- Startups such as Calm and Headspace do not provide neurofeedback, however they guide a user through meditation practice and help to build an ability to understand own body and mind. The effects of meditation on the brain are not yet fully researched. More or less visible effects are evidenced on an example of experienced meditators, who practiced months and years.
Machine learning for interventions
There are at least three ways how machine learning may be applied to hacking the brain, namely:
- To identify the right time for an intervention/simulation;
- To customise a stimulation itself for needs of an individual patient;
- To translate raw data into actionable insights that are required for neurofeedback.
Selecting the right time for stimulation is an important task. Sensor fusion algorithms may be applied to recognising indented movement of a patient and selecting the right moment for electrical stimulation. GTX Medical and NeuroSigma develop devices that send stimulating signals based on real-time feedback from body worn sensors and implanted sensors respectively. Seizure detection is an element of a pulse generator by AspireSR.
Parameters of an intervention themselves are not easy to define and programme.
- Currently, a doctor or an experienced nurse is required to program a stimulator of the vagus nerve, also a special software is required. Moreover, in case of a deep brain stimulation, reprogramming of a device happens due to fluctuations in the impedance at the level of electrode-tissue interface;
- The problem of choosing parameters of an intervention is becoming even more challenging with more complex interventions. For example, postoperative management is required to balance deep brain stimulation and medication treatments. Therefore, clinical decision support systems based on machine learning algorithms may be useful for managing deep brain stimulation.
Machine learning is vital for translating raw signals that are sent by the brain into meaningful insights that allow neurofeedback.
- For example, every 30 seconds, Dreem’s machine learning algorithms determine what sleep stage users are in. Hypnogram, the sum of these stages of sleep over a full night, is shared with a user;
- Researchers at MIT and Harvard used machine learning to detect and distinguish signals associated with pain when participants wore Muse’s meditation headband;
- Computer vision helps MindMaze to decode real-life facial expressions of a user in tens of milliseconds and instantly replicate that expression on a virtual reality avatar that assists stroke patients in their recovery.
There are at least two wide research themes approached by neurotech startups. Some startups work on making brain more visible, while others try to connect it with the outside world.
- Higher resolution/better visibility. Inscopix and its microscope system allows studying the brain of a mouse at the cell level, constantly while an animal lives its normal life. 3Scan uses ‘…computer vision to extract spatial data from tissue samples. The results are detailed 3D representations of anatomical structures’;
- Brain interfaces. Stentrode by Synchron is implanted inside the brain in the motor cortex. It captures and sends signal to a wireless chest-implanted antenna that forwards it to an external receiver. The concept of ‘neural dust’ was proposed by Neuralink. The interface ‘… would consist of thousands of microscopic independent sensor nodes, and one ‘sub-cranial interrogator’ that connects and powers them’. Kernel explores ways of implanting microchips into human heads. Ctrl-labs develops a wearable device thatworks based on differential electromyography, i.e. measuring changes in electrical potential caused by impulses traveling from the brain to muscles.
After having reviewed these startups, I got amazed by several things:
- We have an astonishing chemical factory within our body, and being able to control it may help to fight various diseases with lesser reliance on drugs. Exploring different types of brain stimulation and brain research techniques brings us closer to the control deck of this factory;
- Growing interest to mindfulness and meditation (see chart 2) may increase the adoption of wearable EEG sensors and other devices for neurofeedback. Headspace, one of the most popular meditation apps, has more than 16M downloads, and even a share of its users may help to grow a small portable EEG market ($50M in the US in 2017, chart 2);
- Data generated from growing EEG adoption, coupled with data from human-mobile phone interaction, opens up new opportunities for diagnosis based on digital biomarkers;
- There is a risk that despite increasing interest to neurofeedback and mindfulness, some technologies will be pushing us to a rather passive role in the process of managing our mind, e.g. stimulation instead of meditation;
- I would expect that neurotech will have a massive effect on economy/business and business models. Imagine the demand for neurologists if brain implants become mainstream and surgeries to implant them could become routine? In 2004, European countries had 4.84 neurologists per 100K population. To compare — in 2007, OECD countries had 61 dentists per 100K on average. These figures are a bit outdated, but I believe they demonstrate the challenge I am talking about. Then imagine a situation when a doctor would ask to remotely tap into patient’s memories for a certain day and explore 24 hours’ worth of video, data on brain activity, etc. Would the infrastructure of mobile networks withstand that? These are just some obvious operational challenges, that do not even take into account some other effects such as influence on population healthcare, countries’ competitiveness, philosophical/ethical issues, etc.
- Currently we are at the very early stages of brain tech, just unlocking its value to healthcare, wellness and sports. However, horizontal platforms like NeuroSky and Ctrl-labs may help to bring brain tech to other markets, e.g. education, gaming/entertainment.
If you find this article encouraging, and, if you want to launch/join a neuro tech startup I would strongly advise you to team up/consult with someone with neuroscience or an adjacent background. What I see from my brief dive into the field is that computer science/data science/engineering expertise alone rarely lead to success.
Data collection and sources
* This review covers privately owned companies that raised $10M or more, are engaged in neuroscience research and develop software/hardware products for healthcare professionals, researchers and consumers. Startups for CT/MRI are not included, as it is a theme on its own. Also, the review does not include analysis of pharma companies.
Data is taken from Pitchbook and CrunchBase. If these data bases provide contradictory data, Pitchbook’s data is used.