Top brain-computer interface expert Philip Sabes: The real bottleneck of neuromodulation is "prediction", and treatment and scientific discoveries should occur simultaneously.
According to IPO Exclusive News, on September 14, the BCI Innovation and Transformation Center of Brain Intelligence World successfully hosted the "BCI Inno+ Salon | International Symposium on Brain-Computer Interface". This conference gathered global scientists, entrepreneurs, clinical and regulatory experts in the field of brain-computer interface, to conduct in-depth dialogues on core topics including the technological evolution of high-precision BCI, closed-loop neuromodulation and clinical diagnosis and treatment applications, wearable brain-computer interaction, and the construction of a global industrial innovation ecosystem.
Neuralink founding scientist and current CEO of Arbor Neuroscience Philip Sabesdelivered a speech titled "The Fastest Path to Reliable Neuromodulation: How Neurotechnology Accelerates Therapeutic and Scientific Discoveries" and held discussions with the attendees.
Philip Sabes believes that although neuromodulation technology has been developed for many years, its clinical application is severely lagging behind. To break through this bottleneck, the key lies in two types of technologies: one is implantable electrodes that can record neural signals with high quality over a long period of time; the other is wearable and portable fMRI-like technology to realize long-term monitoring of physiological indicators such as cerebral blood flow.
He pointed out that the existing clinical treatment process should be truly transformed into a systematic scientific experiment, and predictive models should be established by continuously accumulating individualized data, so as to realize the coordinated progress of "treatment" and "scientific discovery", and finally achieve reliable prediction and individualized optimization of neuromodulation efficacy. This is the key path for this field to move towards maturity.
For Neuromodulation to Move Towards Reliable Therapy, a Predictable Causal Chain Needs to Be Established
Sabes first roughly divides brain-computer interfaces into two categories: one is the "outward-facing" interface, which realizes motor control and sensory reconstruction by reading brain signals, and may even achieve a more direct connection between humans and AI in the long run; the other is the "inward-facing" interface, which treats neurological and psychiatric diseases by stimulating or regulating brain activity, that is, neuromodulation.
Over the past few decades, implantable electrical stimulation represented by Deep Brain Stimulation (DBS) has formed a relatively mature product system, and routes such as Transcranial Magnetic Stimulation (TMS), Transcranial Electrical Stimulation (tES), and Focused Ultrasound (FUS) are also developing. But in his view, the speed of technological development does not fully match clinical practice. The so-called "Reliable therapies" first means being able to predict whether a treatment is effective for different patients, and secondly, further optimizing treatment parameters for individuals.
"The real bottleneck is prediction. Because we don't know exactly what will happen after neuromodulation and stimulation, we can't reliably predict treatment outcomes."
This problem is particularly prominent in the fields of psychiatric disorders, chronic pain and other fields. For example, when DBS is used to treat Parkinson's disease tremor, tremor reduction can be observed soon after stimulation; but the efficacy of diseases such as depression, obsessive-compulsive disorder, and chronic pain may only appear after days or even weeks, and there is a long "intervention-response" cycle between one stimulation and the final clinical outcome. At the same time, these diseases usually involve multiple symptoms and multiple brain circuits. Even if the final scale improves, it is difficult to judge which circuit or which symptom has been changed.
In response to this, Sabes proposed three key breakthrough directions: Causal manipulation, Multiscale observations and Longitudinal study, that is, to record physiological changes in time after the implementation of stimulation, and continue to track changes from local brain regions, neural networks to behavior and symptom levels.
Especially in the aspect of causal intervention, it is not enough to find biomarkers that are "correlated" with the curative effect. What is more important is to establish a causal chain between "stimulation-physiological change-final outcome". "What we really need is that when I do this thing, it changes this physiological signal, and this change is related to the final outcome."
Sabes hopes to find signals that appear early enough and can well predict the final curative effect from this process. "If we can know in advance what will happen, we can optimize the treatment earlier, shorten the experimental cycle and get more data. In the final analysis, this is a data problem, and also a problem of how to quickly optimize for individuals."
Enable Clinical Treatment to Continuously Generate Data That Can Be Used for Learning
In the future, neuromodulation devices will be a system that continuously conducts experiments and learning. In reality, doctors are constantly adjusting DBS parameters for patients. In a sense, each parameter adjustment is similar to an experiment: change the stimulation parameters, and then observe the patient's response.
However, Sabes pointed out : "Many of the operations we perform on patients today, which we call parameter adjustment and clinical treatment, are not experiments that allow anyone to really learn something. We do not carry out enough observations, nor do we collect the data." Therefore, the key next step is to improve the effectiveness of clinical data.
The "experiment" here does not need to add additional interventions, but to synchronously record stimulation parameters, brain activity, network changes and clinical outcomes in the normal diagnosis and treatment process, so that one treatment can become the data source for the next treatment decision, and finally form a cycle of "intervention-observation-prediction-adjustment-re-intervention".
While the patient is receiving treatment, the system also continuously learns why this treatment is effective and how the next patient should be treated.
Use Experiments and Data to Answer "What Kind of Equipment Should Be Built" First
This line of thinking also affects the development path of BCI products.
Sabes introduced that he had previously founded a company hoping to directly develop related neural interface devices, but ultimately failed to obtain sufficient financing. One of the important reasons is that the team was not really able to determine exactly what kind of equipment was needed in the end at that time.
"This is a 'chicken and egg' problem. If you already have the equipment, you can carry out experiments; if you complete the experiments and have the data, you will know exactly what kind of equipment to build."
His solution is to first build a good enough experimental device, collect data on a certain scale, and then use these data to define the clinical products that are really needed. This is also one of the issues that Arbor Neuroscience (hereinafter referred to as "Arbor") is currently paying more attention to.
Sabes believes that for neuromodulation, two types of technologies are particularly worthy of attention: one is an electrode system that can stimulate and record at the same time, and the other is ultrasound.
In terms of electrodes, although the existing clinical DBS system can already achieve partial brain signal recording, the data dimension is still limited. A device that is really suitable for studying disease mechanisms and individualized treatment needs to be improved in both stimulation and recording, and have the ability to observe and regulate across different spatial scales.
He said that taking Parkinson's disease as an example, even if it is located in the subthalamic nucleus (STN), different symptoms such as tremor, stiffness, and gait disorder may correspond to different subdivided positions. If more precise stimulation can be achieved, the treatment can be optimized for different symptoms respectively.
In more complex diseases such as depression, it is necessary to expand to the brain network level to understand the problem: different symptoms may come from different circuits. In the future, it is necessary to understand and even regulate multiple targets at the same time, instead of simply finding a unified stimulation position.
Ultrasound Stimulation and Functional Imaging Provide New Paths for Neuromodulation
For another important technology, Sabes pointed out that compared with implanted electrodes, one potential advantage of ultrasound is that there is no need to implant the recording electrodes directly into the brain tissue, and the sound beam can be controlled and focused to act on different brain regions.
The functional ultrasound system previously developed by Arbor (developed by its predecessor Forest Neurotech) further attempts to read neural activity through changes in cerebral blood flow. Sabes analogizes the capability of functional ultrasound imaging (fUS/fUSI) to a kind of "more portable fMRI (functional magnetic resonance imaging)".
However, such high-quality functional ultrasound brain imaging is currently difficult to directly penetrate the intact adult skull to obtain sufficiently clear brain function signals. Usually, patients need to have undergone craniotomy decompression and other operations, and have a skull window or implant material that can transmit ultrasound. When these conditions are met, the system can observe cortical activity and further read signals related to deep brain structures.
Therefore, to use functional ultrasound imaging for long-term observation, the portability of the device needs to be improved. Arbor plans to further miniaturize the system that still requires the cooperation of the host computer and computer, so that patients can take it home and record for several hours or even a long time in a more natural environment.
Then if functional ultrasound imaging has skull limitations, can ultrasound BCI technology eventually be applied to ordinary patients or even healthy people?
In response to this, Sabes talked about the two capabilities of "Focused Ultrasound (FUS) for therapy" and "using ultrasound for brain activity imaging" respectively. He said, when focused ultrasound is used for therapy, transcranial transmission can be achieved by reducing the frequency and other methods, and the current targeted positioning can reach about a few millimeters; but the real difficulty is to accurately see the changes in brain activity while treating.
He summarizes many current neural stimulations as "Poke and hope." ——stimulate, then wait and see what happens. One of the breakthroughs in future neuromodulation is whether stimulation, reading and prediction can form a closed loop.
Brain-Computer Interface Needs to Redesign the Way of Human Data Acquisition
When reliable therapy relies on a large amount of human data, where does this data come from?
Sabes said that for consumer electronic products such as bracelets, companies can let tens of thousands of healthy people wear them for hours, continuously accumulate data, and gradually train the algorithm to a level that "can be used as soon as you take it home", but brain implant devices obviously cannot replicate this route. "You can pay 10,000 or 60,000 people to use a bracelet, but you can't let 60,000 people receive a brain implant."
Therefore, brain-computer interfaces and neuromodulation must design different data acquisition methods. On the one hand, obtain very high-precision, long-time, multi-modal data from a small number of patients; on the other hand, cover more patients in a relatively low-cost way, and then combine data of different scales.
He pointed out that relying only on precise experiments on a single patient is still not enough, because there are huge individual differences in disease types, brain structures and neural circuits. Therefore, the real model in the future needs to solve two things at the same time: see deeply enough for a small number of patients, and cover a large number of patients widely enough.
"In the long run, we hope that the brain interface can not only provide treatment, but also tell us why it is effective. Every patient can become part of a continuous experiment, and all this data will eventually be pooled to make the next treatment better."
Therefore, in addition to pursuing higher channel counts, more precise stimulation and more miniaturized devices, efficacy prediction, individualized optimization and continuous learning capabilities will also become important development directions for the next generation of brain-computer interfaces.
This article is from WeChat Official Account"IPO Exclusive News" (ID: ipozaozhidao), Author: Uncle C, 36Kr published this content with authorization.