Topics Pioneers Insight Method Author
Back

Industry Watch 45: Brain-Computer Interfaces’ Path to Treating Disease

·高能量

Rendering the long image…

Topics Frontiers in Brain-Computer Interfaces and Neuromodulation
Pioneers 李龙 盛经纬

Summary

  • The near-term commercial opportunity in brain-computer interfaces lies in unmet needs among patients existing treatments do not reach. 李龙 says fewer than 1% of Parkinson’s patients had received an invasive deep brain stimulation (DBS) implant; that figure has risen to around 1.x% over the past 2 years. A procedure typically costs RMB100K–300K, depending on the brand. Risk and cost both constrain adoption. 穹顶 hopes noninvasive neuromodulation can lower those barriers, but the product still has to deliver tangible benefits to hospitals, doctors and patients.
  • Clinical usability is reshaping the product trade-offs technology companies make. 穹顶 once had more than 20 channels and roughly 3–4 mm precision, but fitting the device took 30 minutes for a treatment that lasted just 20 minutes. It later scaled back to 4 or 6 channels, cutting setup time to under 5 minutes. 李龙 says companies cannot focus only on greater precision or deeper stimulation; they also have to account for how the device will be used.
  • The value of invasive versus noninvasive approaches depends both on the information a task requires and on the cost patients are willing to accept. Invasive systems offer precise local recordings and can build a long-term dataset for an individual; noninvasive systems can more readily capture activity across more people and the whole brain. 李龙 compares it to a football match: hearing the crowd cheer from outside the stadium is enough to tell whether there was a goal; identifying who scored and how requires going inside. Near-invasive decoding performance in some tasks does not mean brain magnetic signals can now read all of a person’s thoughts.
  • The registration hurdle for closed-loop neuromodulation is proving that more precise interventions deliver clinical benefit. Personalization, accurate targeting and neuronal response are only part of the evidence chain; a new treatment must also prove efficacy or at least equivalence. 李龙 compares the challenge to precision guidance versus “dropping a nuclear bomb”: the former is more precise, but its superiority may be hard to demonstrate against fixed efficacy measures and observation windows.
  • Data is already creating specific product advantages, but an industry-wide data flywheel has yet to be proven. 穹顶 says models built from nearly 100,000 Chinese brain scans allow it to match patient models even when imaging is low quality. The company is also buying EEG and MRI data, paying a little over RMB90 for each structural scan. 盛经纬 cautions that data remains scarce, standards are inconsistent and noise levels are high: “garbage in, garbage out.” More data alone does not guarantee a breakthrough in models.
  • The near-term test for AI is whether it can link brain states to intervention outcomes in a usable model. 穹顶 plans to buy data from around 1,000 people, with roughly 1,000 hours per person, to compare how the brain changes under different physical interventions. Under the plan at the time, the data would take about 8–9 months to collect, but there was no guarantee it would yield a reliable model. 李龙 says, “Optimism comes from conviction; pessimism comes from the data I’ve collected” (“乐观来自信念,悲观来自我采集到的数据”).
  • China’s clinical resources and faster validation may turn policy support into a data-accumulation advantage. Based on 李龙’s project experience, ethics approvals for new indications and targets take around 2–3 months in China, versus at least 2 years in the UK or California. 盛经纬 also points to the advantages of China’s supply chain and clinical resources. Both acknowledge that the US still leads in original devices and foundational technology, and that Chinese companies need to move beyond following in the footsteps of “China’s Neuralink.”
  • The guests’ 3-year scorecard is a listed company, reimbursement for noninvasive brain-computer interfaces and doctors writing orders on their own initiative. 李龙 believes the sector can make the “final push” to commercialization within 3 years, then use revenue to fund upstream R&D. 盛经纬 puts more weight on doctors voluntarily ordering tests to meet real clinical needs. By then, the industry could be “more than just a concept—time to deliver on performance,” with applications continuing to expand into psychiatric disease and higher cognitive function.

Deep dive

1. Both companies started with unmet medical needs

  • 谢达 breaks brain-computer interfaces into 2 basic questions: Can we tell “what is happening in the brain,” and can we accurately intervene in abnormal activity? 昆迈医疗 focuses mainly on reading signals, while 穹顶 focuses on noninvasive modulation. The discussion centered on how the 2 ends of that process could gradually form a closed loop for diagnosis and treatment.

  • 李龙’s previous team worked on invasive DBS devices. He says there are around 1.8 million Parkinson’s patients in China each year, but fewer than 1% had received an implant; that figure has only reached around 1.x% in the past 2 years. The risks and expense of invasive treatment leave large numbers of patients beyond the reach of an effective technology.

  • Treatment typically costs RMB100K–300K per patient, depending on the brand. Around 2016–2017, 李龙 saw that temporal interference stimulation might deliver a function similar to a brain pacemaker. He founded 穹顶 in 2019 to address the shortcomings of existing neuromodulation: it “can’t reach deep enough, can’t target accurately enough and doesn’t work well enough.”

2. The opportunity in helium-free magnetoencephalography comes from advances in sensor technology

  • 盛经纬 earned a PhD in medical physics from Peking University and has long worked with functional imaging. In 2013, he took part in a task as a research subject and saw his own brain regions “light up”—an experience that left a deep impression. He recalls that early imported systems cost around RMB30M, with liquid helium refills running about RMB1M a year.

  • 昆迈 uses optically pumped magnetometers (OPMs), which exploit interactions between light and quantum systems for precision magnetic sensing. Its goal is to capture high-fidelity neural activity without relying on liquid helium. 盛经纬 says the company’s motivation started with a new technology becoming ready for validation, integration and engineering.

  • 盛经纬 and his team were thrilled when their in-house detector first picked up a brain magnetic signal from a Chinese subject. 谢达 recalls that when he first invested, there were fewer than 20 liquid-helium magnetoencephalography systems installed across China.

3. Engineering can iterate quickly; clinical patterns take case-by-case accumulation

  • 李龙 initially approached R&D with an engineering team’s mindset: keep investing in the lab or company, and iteration can happen quickly. But putting a product to work with people means repeatedly dealing with clinical communication, safety and ethics. Even brains with similar structures differ significantly from one individual to the next.

  • With only 1,000–2,000 cases, every new patient with the same disease still felt like a new problem for 穹顶, requiring the team to adjust targets, parameters and strategies again. Only after its experimental dataset passed 10,000 cases did the team start to develop general templates that could be extended to other clinical centers and patients.

  • 盛经纬 says China’s supply chain, clinical environment and openness to regulatory innovation helped the product reach the clinical-registration milestone relatively smoothly. But commercialization after approval brought a different lesson: “Before starting a company, I didn’t know how hard it was to sell equipment.” Good technology and products alone are not enough to win doctors’ acceptance or hospital procurement.

4. Channel count and precision have to yield to clinical workflow

  • 李龙 divides users into 2 groups: hospitals that pay and doctors who operate the equipment, and patients. The first group needs an easy-to-use tool with low medical overhead; the second needs an outcome existing treatments cannot deliver. Adding channels alone does not automatically make a better product.

  • 穹顶’s early device had more than 20 channels and roughly 3–4 mm precision, but doctors were reluctant to spend half an hour fitting it. Patients also struggled with “20 minutes of treatment and half an hour of setup.” A single session took up 50 minutes, disrupting both clinic schedules and the patient experience.

  • The company later chose to reduce complexity, using 4 or 6 channels to do the job and bringing setup time below 5 minutes, at the cost of lower precision. 李龙 emphasizes: “Simply chasing more precision, better performance or deeper stimulation does not necessarily make a product better.”

5. Musk’s biggest impact has been to raise the industry’s profile and willingness to invest

  • 谢达 sees Neuralink’s FDA clinical-trial authorization in May 2023 as a major milestone. 盛经纬 says some of the demonstrations were “not a big deal” to researchers, who had already seen papers from labs. The real change was that someone kept investing and showed the public applications they might be willing to accept.

  • 盛经纬 also points to cases from the University of Lausanne, where brain-computer interface closed-loop treatment helped people with high-level spinal cord injuries walk again. The industry education and shift in public awareness these cases have driven are irreversible.

  • 盛经纬’s speculation about why the giants are entering the field centers on their existing businesses: Musk wants to figure out how robots can serve people and how people connect with robots; Meta may explore neural keyboards and new forms of social media. 李龙 offers a more down-to-earth take: consciousness and other ultimate questions have enormous imaginative appeal and are hard to disprove in the short term.

  • 李龙 does not believe that being hard to disprove means an idea has no value; exploration can still produce useful technology and companies. 盛经纬 compares digital immortality to moving to Mars: a distant goal does not prevent people from first addressing brain diseases and practical interface applications.

6. Noninvasive reading means choosing between different physical signals

  • 盛经纬 explains that neural activity can be measured through electrical or magnetic signals, or inferred indirectly from changes in metabolism, blood flow and oxygenation. Neural activation increases energy consumption, which changes demand for oxygen and blood flow; technologies such as near-infrared imaging use these indirect changes.

  • Scalp EEG has been in use for more than a century, dating back to Hans Berger’s work in 1924, and remains the most widely used noninvasive collection method. The skull acts as both a barrier and a filter, distorting signals as they travel and weakening high-frequency components. EEG is still used for sleep staging, rehabilitation and motor imagery because it is easy to obtain.

  • Brain magnetism and EEG share the same source: both observe neural electrical activity. 昆迈 hopes to take advantage of the similar magnetic permeability of the skull, cerebrospinal fluid and other tissues to reduce signal loss and distortion as signals pass through the body, improving fidelity and downstream computation.

  • 盛经纬 says Meta’s team collected brain magnetic signals from Spanish-speaking subjects as they typed, then labeled the data and trained an AI model. The best accuracy was close to 80%. He sees this as evidence that noninvasive decoding can approach invasive performance in some settings; the example was specific to a typing task, not mind-reading in general.

7. There is still a gap between measuring the brain and understanding thoughts

  • 盛经纬 is explicit that recorded signals cannot currently be equated with someone’s thoughts. He distinguishes brain, behavior and mind: instruments measure the brain, behavioral experiments measure performance, and researchers then try to infer mind, intent and cognition from objective data. Models are still needed to connect the two.

  • From the perspective of a neuromodulation device user, 李龙 acknowledges that invasive signals are higher quality and useful for animal experiments and specific patients who have already had a craniotomy and electrode implantation. But ordinary patients are unlikely to have electrodes implanted just to observe treatment feedback, so high-quality invasive recordings cannot directly serve most of his users.

  • Different noninvasive technologies have different roles. Based on 李龙’s experience, fMRI can support structural localization and guide electromagnetic-field navigation, but lacks the temporal resolution for real-time analysis of an individual. Magnetoencephalography has enough temporal resolution to complement it. EEG, meanwhile, has the size and cost advantages to potentially serve home or consumer settings, though its signal quality is relatively weak.

  • 李龙 sets out a specific near-term goal: first, get post-treatment brain activity closer to a normal state. He thinks that could be achieved within 1–2 or 2–3 years, and is already paying signal-collection providers to pursue it. Figuring out exactly what a patient is thinking is not a problem that has to be solved now.

8. Invasive and noninvasive approaches serve different patients and information needs

  • 谢达 cites an industry estimate that around 85% of related products are noninvasive, while interventional and invasive products together account for around 15%; fewer than 10% may actually be implanted in the brain. He believes patients with severe disabilities, such as high-level spinal cord injuries, are more likely to accept implant risks. Scenarios where the benefits clearly outweigh the risks remain limited.

  • Whole-brain coverage is another dimension of the discussion. Even with many electrodes, invasive systems tend to focus on specific regions. Noninvasive methods can more easily capture activity across brain regions and the whole brain, potentially providing irreplaceable data for higher-level behaviors that depend on coordination across regions.

  • 李龙 uses a stadium to explain the difference between tasks: hearing cheers from outside is enough to tell whether there was a goal; finding out who scored, and whether the ball went into the upper left or upper right corner, requires going inside. “It depends on what you need”—and whether that justifies the cost of entering.

  • 谢达 compares 2 data strategies: Neuralink emphasizes long-term individual use, while Precision focuses on implanting electrodes during a period when a patient needs a craniotomy, then removing them afterward. One approach prioritizes longer duration, the other data collection from specific patients. Whether they can eventually create a multiplier effect depends on data accumulation and downstream applications.

9. TI has to prove that it reaches the target, changes activity and improves function—in that order

  • Temporal interference stimulation (TI) uses 2 high-frequency waves to create a low-frequency envelope. The aim is to minimize effects on mid- and superficial neurons along the signal path while acting on a selected deep nucleus. 李龙 calls it “noninvasive hitting a target from behind a mountain” and says its effects in some clinical applications are very close to those of a brain pacemaker.

  • 谢达 breaks validation into 3 steps: Does the stimulation reach the target brain region? Do neurons respond? Do behavior and disease function improve? 李龙 says target reach has been validated in monkeys and patients with their skulls opened. Whether neurons are affected has been observed using invasive electrodes to record local field potentials and single-neuron spikes in living people or monkeys; noninvasive measurements alone cannot answer that question.

  • 李龙 says some validation paths are similar for electrical stimulation, TMS, deep transcranial magnetic stimulation and ultrasound. The key difference between modulation and signal collection is that modulation ultimately has to produce a significant clinical effect. 穹顶 cites Parkinson’s disease, essential tremor and some psychiatric conditions as applications, but still needs to establish benefits indication by indication.

10. Differences in outcomes for specific indications are what can create pricing power

  • 李龙’s commercial logic is straightforward: if sound, electricity and magnetism can all treat the same disease with similar efficacy, the market will ultimately compare treatment costs. If only one approach works, it may have pricing power. He uses a comparison of one technology scoring 60 and another 90 to emphasize that an advantage has to show up in a specific indication.

  • 李龙 also cautions that transcranial magnetic stimulation, even if it is not deep or precise enough, still solves many problems; China may ship several thousand units a year. In the clinical settings he has observed, relieving anxiety and tension before an exam does not necessarily require precise stimulation of a specific nucleus.

  • New tools for signal collection and neuromodulation give doctors a chance to better understand individual differences and identify more suitable targets. 李龙 says the key to doctors accepting a treatment is whether patients benefit. Once they see that benefit, doctors’ recommendations and word of mouth are more effective than manufacturers’ own education.

  • 李龙 also raises weight loss as a personal need. He has experienced the effects of medication but found it hard to stick with, and was affected by nausea and mood changes. He would like to explore appetite regulation with a lower cost. He offers this as an example of potential demand among healthy people, not as a weight-loss treatment that has already been validated.

11. Closed-loop modulation faces a registration hurdle: proving clinical superiority

  • 谢达 highlights a tension: companies emphasize personalization and closed-loop treatment, while regulators need key aspects of the treatment to be fixed, leaving limited room for variation in practice. 李龙 says that, at the time, no noninvasive Class III brain-computer interface registration certificate had been granted; one of the challenges is proving that closed-loop treatment delivers better outcomes.

  • 李龙 uses a war analogy. Broad modulation in the past was like “dropping a nuclear bomb”: leveling the whole city made the effect obvious. Closed-loop treatment is like precision guidance, destroying only the target building. But when outcomes are assessed externally against fixed criteria, it may not be easy to prove that the precise approach works better than broad intervention in the near term.

  • Greater precision does not automatically resolve the registration challenge. A new treatment has to prove efficacy or equivalence; when stimulation is weaker and covers a smaller area, companies must also show results within the specified timeframe. 谢达’s response is that any breakthrough still needs solid evidence of efficacy and safety.

12. The broad reach of noninvasive technology has not automatically created a must-have market

  • 盛经纬 admits that the weakness of noninvasive technology is the lack of clearly defined use cases: it “seems able to do everything and serve a broad population,” but is not always irreplaceable. There is plenty of scope in diagnosing, treating and assessing psychiatric conditions, but medical devices specifically intended for a psychiatric disease are still regulated as Class III.

  • At the other extreme are consumer entertainment products: put on one or two dry electrodes and it can feel as if you have superpowers. 盛经纬 believes that when the technology is not yet mature, these applications can get stuck in a semi-entertainment state, unable to meet the efficacy standards of serious medicine.

  • Rehabilitation is a middle ground, with real medical needs and little suitability for invasive approaches. But 盛经纬 leaves one key question open: are the results due to the brain-computer interface, the conventional rehabilitation that happens at the same time, or the interface raising the ceiling for treatment? The contributions still need to be disentangled.

  • 李龙 believes demand for signal collection and modulation will grow together. When medication is the only option, doctors prescribe based on rating scales and have limited need for fine-grained brain imaging. Precise modulation creates a stronger need to know “where to stimulate and how to get there,” giving diagnostic tools a new reason to be used.

13. Paying customers are shifting from research buyers to clinical departments

  • 李龙 describes how 穹顶’s customer base has evolved: first university research users, then hospital researchers, and now actual clinical departments. Early buyers paid to explore a new technology and push it forward; increasingly, buyers want direct improvements in patient outcomes, lower labor requirements and a path to reimbursement.

  • For hospitals, equipment needs to deliver better service and revenue with less investment. For payers, the ideal is to control costs better than ongoing medication or surgery. 李龙 sees benefits for all 3 parties as the basis for commercialization—not technology novelty alone sustaining purchases.

  • 昆迈 still sells mainly to large institutions. 盛经纬 acknowledges that magnetoencephalography systems are large, expensive medical equipment, and that procurement and use take time. Using brain-reading and neuromodulation equipment together in hospitals could create more paid demand for this infrastructure.

14. Neuromodulation companies are starting to buy data rather than build their own collection infrastructure

  • 李龙 compares 昆迈 to “big infrastructure”; what his own company needs is the data generated on that infrastructure. For example, if he needs brain magnetic data from 2,000 people, with at least 3 hours per person, buying equipment and recruiting and collecting the data himself would be unaffordable. Buying compliant, de-identified data from hospitals can save both money and labor.

  • 穹顶 has already bought data from brain EEG foundation-model companies and structural MRI scans from hospitals, paying a little over RMB90 per case. 李龙 believes buying data directly is more efficient for the company, which needs to advance its algorithms and registration work; building the entire infrastructure itself does not make economic sense.

  • Its data needs include baseline brain magnetic signals and signals collected after TI, to measure the difference between treated and untreated cases, then narrow that gap through target selection, fitting and parameter adjustments. 李龙 wants a comprehensive platform for sound, light, electrical and magnetic data and is willing to pay equipment makers or hospitals for access.

  • 谢达 warns that brain data is highly sensitive; 盛经纬 thinks it may eventually be managed like genetic data. 李龙 stresses that patient identities are not needed—only characteristic parameters from Chinese or Asian populations. How to de-identify the data and ensure compliant transfer and transactions remains a prerequisite for the model.

15. The data flywheel is constrained by quantity, standards and quality

  • 谢达 describes the flywheel as more patients generating more high-quality signals, better models improving products, and better products attracting more patients and doctors. 盛经纬 says it works in theory, but he has yet to see it form a closed loop in practice.

  • The first constraint is that brain data remains scarce relative to natural-language, internet and IoT data. The second is wide variation across individuals, collection settings, degrees of invasiveness and modalities. Data from local versus whole-brain recordings, or electrical versus magnetic signals, cannot simply be combined because it all comes from the brain.

  • Third, brain data is noisy and signals can be mixed together; it is also hard to label. 盛经纬 echoes the warning in the discussion: “garbage in, garbage out.” He acknowledges that large models could improve data utilization, but thinks sheer volume may not deliver the kind of breakthrough seen in natural language.

  • 谢达 adds that the industry still lacks a widely accepted model architecture that can represent different brain signals in a unified way. Much language decoding still reads the motor commands needed to speak, which is essentially motor decoding. Moving into sensation, consciousness and logical reasoning will require more high-quality observation and intervention data.

16. Data accumulation is already reducing the hardware requirements at the point of care

  • 李龙 believes China is strong at copying and iterating on engineering, so a pure hardware lead may last only 3 or 5 years. That is why 穹顶 is willing, beyond pursuing revenue, to give up some commercial value in exchange for customer data and convert accumulated experience into better treatment matching.

  • He gives county-hospital tenders as an example: some competitors require 1.5T or 3.0T-and-above MRI systems or high-throughput magnetoencephalography. With models built from nearly 100,000 Chinese brain scans, 穹顶 can find a close match in its database using 0.5T imaging—or even a structural image taken on a phone—and solve the problem.

  • 盛经纬 describes this as data accumulation providing a “dimensionality-reduction method.” For 昆迈, the next step is similarly to move from medical-device hardware manufacturing toward offering solutions, and to turn data into an asset that can support defensibility through models and ongoing services.

17. Decoding higher cognition will require data and computation to advance together

  • 盛经纬 describes the vision as moving from “measure the brain” to “compute the mind.” His premise is that the brain and mind are connected in some way, but whether this can be achieved, and how long it will take, remain open challenges. Progress will require data accumulation and abstract models from computational neuroscience.

  • He uses psychotherapy as an example: therapists do not have especially advanced tools, yet conversations can help people explore their inner lives or resolve some problems. After speaking with computational neuroscientists, he saw a possible path: with the right abstraction, higher cognition could come close to a classification problem, allowing data and AI to be used for training.

  • 谢达 asks whether noninvasive approaches could make breakthroughs in higher cognition that exceed invasive ones. 盛经纬 says that if these behaviors mainly reflect coordination across brain regions, then researchers need to “see the whole picture”; every local region does not necessarily need to be measured with extreme precision. On that assumption, whole-brain collection is a major opportunity for noninvasive technology.

18. Experiments will determine whether multiple physical interventions can fit into a unified model

  • 李龙 says 穹顶 is studying ultrasound not just to compete, but to compare how EEG, brain magnetic signals and MRI change when TI and ultrasound act on the same STN nucleus. The first question is whether different physical interventions on the same target produce identifiable differences.

  • The company is buying data from around 1,000 people, with roughly 1,000 hours per person, covering sound, light, magnetic stimulation and its own electrical stimulation. Under the project plan at the time, the data would arrive in 8–9 months. 李龙 is explicit that “there is no guarantee yet” it will produce reliable, usable results.

  • Early pilot studies can already show clear differences between intervention and no intervention, but the team cannot yet say whether different methods or targets produce stable differences. If they can identify clear biomarkers, 李龙 believes that would explain the significance of combining multiple physical factors for modulation—and would be an important discovery.

  • 李龙 has no detailed roadmap for decoding consciousness. He thinks a breakthrough could come once technology reaches a certain point. He says he is “both pessimistic and optimistic,” explaining: “Optimism comes from conviction; pessimism comes from the data I’ve collected” (“乐观来自信念,悲观来自我采集到的数据”).

19. BrainOS abandoned an approach to psychiatric assessment and classification based on subjective labels

  • 盛经纬 says BrainOS initially aimed to use brain magnetism for more precise psychiatric assessment, early screening and classification, but later abandoned the direction. The issue was not simply that the technology could not do it; training depended on relatively subjective labels, and problems with the labels carried through to the reported accuracy.

  • With limited data and clear labels, training—and even overfitting—could produce accuracy of 90% or even 100%, but 盛经纬 does not think those figures necessarily mean anything. Following the classic path of identifying a condition, finding a target and then developing a treatment could also require a very long accumulation of evidence.

  • His alternative idea is to use collection and intervention data to learn the brain’s dynamic responses, define normal states and treatment endpoints, and then move directly into treatment research. For psychiatric conditions, it may be possible in the future to first build an effective black-box treatment model and explain its internal classification and mechanisms later.

  • 谢达 asks whether this is a “world model of the brain,” and 李龙 agrees. 盛经纬 is also clear that the idea poses major challenges for current medical ethics and regulation. It remains a possible future path, not a treatment approach that has been clinically validated.

20. There is no consensus on who will lead the value chain; a complete application still needs every link

  • 李龙 does not agree that one company or one part of the value chain will necessarily emerge as the leader. Trace the chain from materials to molecular design to electrodes, and any link could become a bottleneck. He prefers the bucket analogy: applications can take off only when every critical component clears a basic bar. The sector also needs a variety of approaches, from technical showpieces to practical deployment.

  • 盛经纬 thinks the eventual winner may be “whoever makes the closed loop work,” but says it is too early to decide. For now, the leading players are still companies that control data access or have advanced sensors or electrodes. 谢达 brings the discussion back to efficiency: the most efficient company could emerge at any point in the value chain.

21. China’s clinical-speed advantage coexists with a shortfall in original technology

  • Based on 李龙’s experience running projects in China and overseas, ethics approvals for new indications and targets take around 2–3 months in China, versus at least 2 years in the UK or California. Even noninvasive projects face this difference. He believes earlier access to data creates a knock-on effect and that “iteration speed” is a major advantage for China.

  • 盛经纬 adds that China has abundant clinical resources and that its advantage in generating data could grow over the next few years. The supply chain has also evolved from highly customized parts—sometimes “handmade by ourselves”—into a system capable of supporting an industry. These conditions help research make the transition into products and clinical use.

  • He also acknowledges that the US still leads in original technology and devices. The industry is still accustomed to describing itself as “China’s Neuralink” or “China’s Merge Labs,” a sign that independent innovation in foundational ideas, chips and technology remains necessary.

  • 谢达’s reading of patent activity is that the US focuses more on basic neuroscience, materials, chips and algorithms, while China focuses more on clinical validation. He also describes the pace by saying that “the past 3 years” have produced roughly the clinical case volume of “nearly 50 years” in the US, and argues that clinical results and data should quickly feed into downstream application development.

22. The 3-year scorecard should be reimbursement, doctors ordering tests on their own and actual business performance

  • 李龙’s 3-year milestones include China’s first listed brain-computer interface company and its first reimbursed noninvasive brain-computer interface case. He believes the “final push” to commercialization should be complete by then, and that the shape of noninvasive adoption and competition in China will be clearer.

  • He sees commercial revenue as a condition for further research: Chinese companies first have to survive and get paid, then use the returns to fund data purchases, mechanism research and upstream R&D. A sustainable cycle requires every party to reduce costs or capture value.

  • 盛经纬 also thinks commercialization and registration approvals could arrive within 3 years, but puts the most weight on a specific use case: a department doctor voluntarily ordering a brain-computer interface test to address a real clinical problem. That would show the product has entered the medical workflow; a decline in industry hype would not prevent it from creating value.

  • Both guests hope the industry reaches a point where it is “more than just a concept—time to deliver on performance.” 盛经纬 also hopes that in 3 years or longer, brain-computer interfaces will move beyond rehabilitation departments into the assessment, intervention, diagnosis and treatment of psychiatric conditions, as well as the analysis and application of higher cognitive behavior.