Meet Hemispheric: The Israeli AI startup aiming to revolutionize brain diagnostics

The sleek, three-story headquarters of Hemispheric in central Tel Aviv has two entrances. The smaller one is for employees. The larger, more spacious entrance is for test subjects. That is no coincidence: The people who come here generate the lifeblood of any artificial intelligence company — its database, its data. How does it work? Young and old, men and women, are fitted with a sensor-equipped cap worn on the head. They are then led into a quiet room, where for about an hour they are asked to perform certain tasks in front of a computer screen, with the results recorded immediately on the company’s servers. “In recent years,” says Gidi Littwin, 44, Hemispheric’s chief technology officer and co-founder, “one of our main focuses has been building one of the world’s largest brain databases. We have collected data from 100,000 subjects and already have a quarter of a million hours of brain-activity data. “The same brain signal can look completely different in different people, but once you collect data on a large enough scale, you can build an AI model that identifies those patterns and measures brain function accurately.” “Without that data,” explains CEO and co-founder Dr. Hagai Lalazar, 50, “AI gives you nothing. It’s like having a powerful car engine with only two drops of fuel — you’ll travel a meter and stop. For the engine to work, you need the right kind of fuel and enough of it for the size of the engine. The fuel is the data. The engine is the AI model. You have to scale both together, and then you can really take off.” Matching the treatment to the condition And the Israeli company is indeed taking off. Quietly operating until only about a month ago, Hemispheric was recently revealed for the first time in Israeli and international media. The company says it is the first in the world to offer a precise system for objectively measuring and decoding human brain activity, capable of diagnosing a long list of neurological and psychiatric conditions in a test lasting just 15 minutes. The list is extensive: depression, anxiety, post-traumatic stress disorder, or PTSD, schizophrenia risk prediction, Parkinson’s disease, Alzheimer’s disease, invisible traumatic brain injury, obsessive-compulsive disorder, or OCD, tinnitus, attention-deficit/hyperactivity disorder, or ADHD, sleep disorders and early signs of cognitive decline. The system can also assess the severity and degree of these conditions. The patient wears the cap, which is connected to an app on a tablet or cellphone, and the system measures brain activity and generates data. The concept is similar to a blood test, X-ray or echocardiogram, which help physicians diagnose illness, select treatment and track a patient over time. Why does this matter? Because while precise, objective tests have long existed to assess the heart, lungs and other major organs, conditions originating in the human brain have not had equivalent tools. To this day, many are diagnosed through interviews, questionnaires, observation and behavioral assessment. Hemispheric employs 114 people in three countries, including experts in neuroscience, artificial intelligence, software engineering, electronics and clinical-grade medical systems, some of them with doctorates. The company also works with leading advisers in psychiatry, neurology and neuroscience. What will the output your system gives doctors look like? Dr. Hagai Lalazar: “Like the routine report you receive after a blood test.” So if there are different degrees of PTSD, you will be able to identify each one precisely? “That’s right. You’ll be able to know exactly which type you personally have.” From among all the known manifestations of PTSD? “Yes, but not only that. We are also working on defining new subtypes.” But doctors will have to understand those numbers. “That’s right. We will train them. We still cannot disclose the names of all the organizations we are working with. They include organizations in the United States and Israel involved in veteran rehabilitation, pharmaceutical companies and companies developing noninvasive devices to treat brain disorders. “For example, we are in contact with the American Psychiatric Association. They are feeling their way in the dark. They are hungry for this information.” And at the next stage, pharmaceutical companies will begin developing drugs according to the different subtypes you diagnose? “That’s right. We will be able to know, for example, that people suffering from depression type X rather than type Y respond better to a particular treatment at a certain dosage.” Lalazar explains: “The brain is unlike any other part of the body. It is the most important and most complex organ. It is not merely molecular. It is an organ with an electrical language, and it is isolated from the body and from the outside world. “Until now, to truly ‘measure’ it, you had to lift the skin, open the skull and implant a chip, which is an extremely invasive procedure. And even then, you only reach a small area of the brain.” This lack of tools for understanding brain function has serious consequences. Between one in two and one in three people will experience a brain-related problem at some point in life, in childhood, adulthood or old age, and humanity still lacks adequate tools to confront many of these conditions. It will affect all of us in one way or another: At 50, you suddenly forget where you put your keys. In the middle of life, you may suddenly begin suffering from chronic distress. Lalazar says: “In Israel, we are on the verge of a mental-health epidemic following the October 7 massacre, the war with Iran and the events that followed. According to estimates, about 6% of the population here will suffer from chronic PTSD, meaning about half a million people. “The state has woken up to the problem, and we were asked to conduct a massive clinical study on the issue in cooperation with relevant bodies in Israel. We conducted a pilot involving 122 veterans of combat units who served in Gaza, including both healthy participants and those diagnosed with PTSD. “For those dealing with the condition, we showed that their situation is the result of a physical problem, a biological condition in the brain, not a weakness of character. Once people understand that, they become more open to receiving treatment. “Pharmaceutical companies have been unable to develop an effective solution because they do not have a ‘digital stethoscope,’ a device that can identify a brain condition, certainly not its various degrees. “When a person goes to a doctor and says, ‘I feel anxious,’ or ‘I’m forgetting things,’ there is currently no way to test and measure that biologically. The doctor has to rely on the patient’s subjective verbal report.” Hemispheric has already presented its system and applications to senior officials at the FDA’s Center for Devices and Radiological Health. The company has already built applications based on the model it developed. To validate its findings and obtain approval from health regulators, it must now collect data from additional patients. Lalazar says: “We have already signed agreements with medical organizations and pharmaceutical companies in both Israel and the United States. “In Tel Aviv, we established a research site that will open at the end of August on Yigal Alon Street. It will have six laboratories, a 15-person team operating them and five interview rooms. “We are also establishing a similar site in Boston, which will open at the end of September. It will have 10 laboratories and the capacity to conduct about 50,000 diagnostic tests a year. Altogether, the two sites will be able to conduct about 85,000 tests annually. “After we validate the applications, we will conduct clinical trials. Once we receive regulatory approvals, as early as next year we will submit the results of our PTSD trial to the FDA. “We will then offer our testing services at these sites, as well as in hospitals and clinics across Israel and the United States and, in the future, in Europe.” To date, the company has raised $52 million from a number of leading venture funds and investors. What is your long-term business model? “At the end of the day, we are an artificial intelligence company. Eventually, the model will sit in the cloud and we will allow companies in other fields — training, instruction, education and gaming — to connect to our service and build their own applications on top of it. We will make money from usage fees.” The Apple connection Hemispheric — a play on the word “hemisphere,” or half of the brain — was founded by Dr. Hagai Lalazar, a computational neuroscientist and entrepreneur, together with artificial intelligence expert Gidi Littwin. Lalazar has worked in the field for 25 years, both in academia and industry. He earned his doctorate at the Hebrew University of Jerusalem and completed his postdoctoral research at Columbia University in New York. He is also intimately familiar with brain implants. “That was the subject of my research,” he says. “The idea was to implant chips in the brain that could help paralyzed people function independently. Today, Elon Musk’s Neuralink is working on this, but 23 years ago we conducted one of the first experiments in the world using this technology. It is very exciting to see where it has gone today, but it is still intended for very few people. After all, it involves brain surgery.” Hemispheric is the fourth startup Lalazar has been involved in. His previous company, which sought drugs for brain diseases, was founded in the United States with colleagues from Columbia University and was later sold to pharmaceutical giant Roche. “When I returned to Israel, I was already thinking about the next thing I wanted to do: quantify brain function. I wanted to find a new, more widely available and cheaper way to do it, one that was noninvasive. “I knew that to build such a company, I needed a partner who was a world-class AI researcher. I met about 75 of them before I found Gidi.” Where? “On LinkedIn.” Co-founder and CTO Gidi Littwin is one of Israel’s most prominent and experienced computer vision experts. Littwin, who spent 15 years in the industry as a researcher or chief scientist at various companies, was previously among the founders of Israeli startup RealFace, which developed facial-recognition technology using deep learning. Apple acquired the company in 2017, and its technology was incorporated into what is now known as Face ID, Apple’s facial-recognition system. Littwin, whose name appears on a key Face ID patent, went on to become a successful researcher at Apple and was among the leaders of AI development for Vision Pro, the company’s mixed-reality headset combining digital content with the physical world. “My team and I led development of the device’s hand-tracking system at Apple, the best solution of its kind in the world,” he says. “At Apple,” Littwin says, “I truly learned for the first time how you build artificial intelligence that reaches a billion people. A researcher publishing an academic paper is motivated simply by getting the paper published. The fact that the product is not perfect matters less. You keep researching. “But when you are developing a device for billions of people, you have to go all the way. For Apple to adopt a particular development, it has to be perfect. And if it isn’t perfect, they tell you, ‘We’re not interested. Go back to the drawing board and do it again.’” And then you met Hagai. Littwin: “He told me about his background. I had zero background in neuroscience. Hagai raised the idea of doing something noninvasive in this field. We started talking about how we could connect the two disciplines.” Lalazar: “Gidi asked me something at the time that shook me: Why aren’t we insisting on 100% accuracy? That idea — that it might even be possible to strive for complete biological decoding — really unsettled me. “I told him, ‘I don’t know. Let’s test it.’ At first we reached 70% accuracy, then jumped to 90% and now we are well beyond that.” No offense, but you are hardly alone. There are already dozens of companies worldwide claiming they can understand brain processes noninvasively. Lalazar: “This field is divided into two camps. Some are developing new sensors because they believe hardware will solve the problem. “We think there are already enough sensors. The sensor cap used to pick up brain waves, which we and others use, is an off-the-shelf product and has existed since 1924. Everyone today can capture brain waves. The problem is understanding and decoding them. “That requires complex algorithms and artificial intelligence on a very large scale. It is painstaking research. You try and fail, again and again.” Littwin: “The brain is an extraordinarily complex machine: 80 billion neurons and around 10 to the 13th power synaptic connections between them. They communicate in their own electrical language, and you need artificial intelligence that can understand that language. “The only way to learn a system that vast is to create a machine powerful enough to match it. You cannot build such a machine with a database collected from only a few dozen people. “If you don’t have the data, you can build some small feature. But if you truly want to decode the brain, to create a model that translates the language of the brain, you need a sufficiently powerful machine.” But your data is also limited. Even 100,000 test subjects is still relatively small compared with the enormous numbers you would need to reach 100% accuracy. Littwin: “Of course, every additional person who comes in for testing and is added to the database improves the accuracy. “In principle, we would obviously like the largest possible database, containing everyone on the planet — 8.3 billion brains. But since that is impossible, artificial intelligence comes into the picture. “That is where proper modeling is needed to achieve accuracy, and that is where our knowledge and expertise in training AI come in.” The AI model developed by Hemispheric is, in effect, a brain equivalent of GPT, based on a new language for reading, measuring and objectively decoding human brain waves. The model already contains six billion parameters — the internal variables it learns and adjusts automatically during training. The greater the number of parameters, the more precise the model can become. It is called Descartes, a tribute to René Descartes, the French philosopher, scientist and mathematician and one of the fathers of the Scientific Revolution, best known for the phrase “I think, therefore I am.” The company describes Descartes as a six-billion-parameter NeuroAI foundation model designed to interpret noninvasive electrical brain activity. Much as large language models “understand” the meaning of text and images, Descartes translates the brain’s electrical language and can “decode” different brain-wave patterns. One output might indicate the level of depression a person is experiencing. Another could reflect the level of anxiety. “We are now preparing to train the model on a much larger scale,” Lalazar says. “Every doubling of the number of parameters means months of research and further improvements in the system’s accuracy.” Between Tel Aviv and the Philippines How does the data collection actually work? Lalazar: “We have two sites: here, at the company’s building in Tel Aviv, and another in the Philippines. We receive the test subjects, connect them to the sensor cap and for about an hour they are asked to perform certain tasks in front of a computer. “The Tel Aviv site is open 13 hours a day, five days a week. The site in the Philippines is open 15 hours a day, seven days a week.” Aren’t you concerned about bias? Are you certain brain activity among Filipinos is comparable to that of Americans or Israelis? Littwin: “Absolutely. We test it constantly. There are indeed slight differences in brain activity among people living in different regions, but most of it is entirely the same. “The main brain activity for all of us relates to how you move your body, how your eyes work and how you perceive the three-dimensional world. Ethnic diversity is not relevant here.” The natural instinct is to assume the opposite. Littwin: “The truth is that differences in human brain activity are mainly expressed from one individual to another, not between ethnic groups. “My brain and yours are very different from each other, regardless of whether I live here or was born in Africa.” In other words, the neurological difference between two Africans, or between an African and an Israeli, is essentially the same kind of difference as between two Israelis living in the same Tel Aviv building. Littwin: “Yes. We proved that. But there are nuances. For example, we speak different languages, so brain activity associated with language will look different.” You sound like Donald Trump’s next nightmare. Imagine if, as part of his annual public health report, he were required to undergo your assessment. Lalazar: “Listen, we are building a ‘digital stethoscope’ for the brain. If Congress or the Israeli Knesset decides — and I hope they do — that the public deserves precise information about the brain function of leaders and politicians as well, that would be a dream for Western democracy.” Someone will eventually have to build on your technology and take the idea further. It could also be used to improve certain functions in healthy people. “Yes. The potential is obvious. There are many possible applications for the technology, including diagnosing and improving learning in many fields.”