China issues a medical-device standard for AI processing of EEG data in brain-computer interfaces

By Inside BCI · September 17, 2026

Chinese regulators have moved on a slice of BCI oversight that no other jurisdiction has yet addressed: the quality of EEG data that AI algorithms consume when they decode brain-computer interface signals. On Monday 14 September 2026, China’s National Medical Products Administration (NMPA) issued a medical-device standard covering AI processing of EEG data in BCI medical devices. State news agency Xinhua’s English coverage on 16 September characterises the standard as world-first on that narrow slice. The standard was drafted by the National Institutes for Food and Drug Control (NIFDC) under an expedited procedure. Effective date is 1 September 2027.

What the standard covers

The standard applies to medical devices that use AI algorithms to post-process electroencephalogram (EEG) signals in a BCI setting. Its subject is the training and evaluation data itself: how EEG signals are collected, processed, annotated, stored, and accessed across the R&D, production, and quality-control stages of a medical device’s lifecycle. In effect, the standard defines what a compliant AI-BCI training dataset looks like, and what evaluation methods a manufacturer must apply to those datasets before an AI algorithm can be judged fit for a medical BCI.

The framing of the problem, from named NIFDC official Li Shu (Deputy Director of the NIFDC medical device institute and standard drafter), is that the field’s EEG datasets were fragmented at every level. “Some used 64 channels, others 128 channels. Sampling frequencies also varied, with some using 2 Hz and others using 1,000 Hz.” The standard’s function is to move medical-BCI AI training out of that fragmented state.

Yuan Peng, Deputy Director of the NMPA Department of Medical Device Registration, framed the intent: “Simply speaking, this standard moves the emerging BCI industry from ‘everyone speaking their own language’ to ‘a unified dialogue,’ allowing these medical devices to enter the market faster and in a more standardized way.”

The China regulatory sequence this fits into

This is China’s third BCI medical-device standard in twelve months. Prior 2025 standards covered BCI terminology (a definitional standard) and test methods for the sensing and response performance of implantable neural stimulators with closed-loop functions (a technical performance standard for a specific device class). The 14 September 2026 standard is the first to address AI processing of the EEG data feeding the decoder, rather than terminology or hardware performance.

Read alongside earlier 2025 and 2026 regulatory milestones, the standard sits inside a coordinated architecture. In June 2025 the Ministry of Industry and Information Technology (MIIT) established a BCI standardisation committee. In August 2025 seven Chinese ministries issued joint development guidelines for the BCI industry. In March 2026 the NMPA granted the world’s first commercial invasive BCI clearance, to Neuracle Medical Technology’s NEO-ONE SCI device. In June 2026 the NMPA issued classification and naming guidelines for BCI medical devices. In August 2026 MIIT’s 15th Five-Year Plan named BCIs as a future industry. The 14 September AI-EEG data standard is the latest substantive move in that architecture, arriving roughly fifteen months after the June 2025 MIIT standardisation committee that opened it.

The FDA and EU comparison holds up on the narrow slice

The claim that the US FDA has no equivalent is defensible on the narrow AI-EEG-BCI slice. The FDA has a 2021 finalised guidance on implanted BCI devices for paralysis and amputation (non-clinical testing, not AI-specific), a January 2025 draft guidance on AI-enabled medical device software functions (cross-device, not BCI-specific), and a December 2024 final guidance on Predetermined Change Control Plans for AI/ML-based medical device software. None of these is a data-quality standard specifically for EEG training data in AI-BCI medical devices.

The EU AI Act (Regulation 2024/1689) treats medical BCI as high-risk under Article 6(1) and Annex I via its intersection with the Medical Device Regulation, and the European Commission’s February 2025 Guidelines on prohibited AI practices name BCI as a vector under the Article 5(1)(a) subliminal-manipulation prohibition. No EU BCI-specific data-quality standard exists. The framework is horizontal.

Internationally, there is no equivalent BCI-specific AI training-data-quality standard published by ISO/TC 215 (health informatics) or ISO/IEC JTC 1/SC 42 (AI), the two international bodies most likely to address this. UNESCO’s November 2025 Recommendation on the Ethics of Neurotechnology is non-binding principles. On the narrow AI-EEG-BCI training-data-quality slice, China’s world-first framing survives adversarial pressure.

What the standard is and is not

The standard is a policy signal today. It becomes effective 1 September 2027, roughly one year after publication. Chinese manufacturers of medical-BCI AI systems have that window to bring their EEG training-data pipelines into compliance. In the interim, the standard functions as guidance for Chinese BCI developers and as a public benchmark against which regulators in other jurisdictions can be measured.

The standard’s operative status is not explicitly stated in available English-language coverage. Some downstream reports characterise it as “mandatory” or “Class III”; those framings are editorial and do not appear in the Xinhua release. Whether the standard triggers device-classification consequences (Class II versus Class III) is not yet clear.

Named Chinese BCI operators (Neuracle Medical, StairMed, NeuroXess, NeuCyber, Weiling Medical, BrainCo, Arfysica Innovation) were not named as consulted parties in the release. The standard was developed inside NIFDC, not through public industry consultation as visible in the announcement.

Where this sits on the four-jurisdiction posture map

Issue 01 of the Dargentic Intelligence Report classified four global regulatory postures on brain-computer interfaces: Rights-First (Chile), Horizontal Regulator (EU with AI Act and GDPR), State-Patchwork (US with Colorado, California, Montana, Connecticut neural-data laws enacted), and Industrial Builder (China and South Korea). The 14 September standard is a textbook Industrial Builder move.

The pattern here is that Chinese regulators are building a purpose-built BCI-specific regulatory architecture in parallel with, and ahead of, an active commercial development pipeline (NMPA March 2026 first invasive approval, Neuracle STAR Market prospectus accepted June 2026, Arfysica pre-IPO tutoring September 2026, BrainCo HK confidential filing January 2026). The regulatory framework is being purpose-built to serve national-industry commercial deployment, rather than repurposing horizontal medical-device frameworks that were designed for other technology classes.

This is what makes the “FDA has no equivalent” framing analytically substantive rather than state-media chest-beating. The US regulatory architecture governs medical BCI through general FDA CDRH pathways (Breakthrough Device Designation, IDE, De Novo, PMA) and general AI/ML software guidance. Those pathways are technology-agnostic by design. The Chinese pathway is technology-specific by design. That is a real difference in regulatory posture, not a superlative claim.

The commercial implication for Chinese BCI operators

Neuracle Medical’s STAR Market prospectus, filed with the Shanghai Stock Exchange, describes a company whose commercial base is entirely non-invasive EEG products (20-plus SKUs, 8 to 1024 channels, deployed at 500-plus hospitals and 300-plus research institutions), plus a first-mover invasive product with zero disclosed revenue through the reporting period. The 14 September AI-EEG standard directly touches the non-invasive commercial base that funds Neuracle’s IPO pathway. Any AI decoder Neuracle ships on EEG data captured by its own 8 to 1024 channel systems will, by September 2027, need to be trained on compliant data pipelines. The same is true for Arfysica Innovation’s non-invasive EEG-plus-exoskeleton rehabilitation platform, and for the consumer-facing BrainCo Focus product lines.

The commercial takeaway is that the standard formalises the training-data pipeline as a regulatory checkpoint for Chinese non-invasive BCI operators. It affects the invasive operator field (StairMed, NeuroXess, NeuCyber, Weiling) only indirectly at this stage, since intracortical decoders are trained on different data. But as invasive operators move to hybrid decoders that combine intracortical and non-invasive signals, the standard’s scope will grow.

Not disclosed

The exact NMPA standard number, Chinese-language title, and direct NMPA notice URL were not located at publication. Whether the standard is voluntary, recommended, or mandatory, and which device classification tiers (Class II, Class III, both) it applies to, is not explicit in available English-language coverage. Which Chinese BCI operators, if any, were consulted during the drafting process is not stated. Whether a companion Chinese-language technical annex is being released alongside the standard is not stated. Whether NMPA plans additional AI-BCI standards on adjacent domains (invasive decoder training data, closed-loop stimulation algorithm validation, non-invasive consumer-BCI safety) is not stated.

What to watch

Whether Chinese BCI operators publicly acknowledge the standard and describe their compliance readiness in the coming quarters. Neuracle Medical is the most directly exposed operator, and its STAR Market prospectus process (currently at CSRC inquiry stage) is the natural forum for a public statement. A Neuracle prospectus supplement or CSRC response addressing the AI-EEG data standard would be the first material commercial signal.

Whether the FDA or EU responds with a matching move. The February 2025 EU Commission Guidelines on prohibited AI practices already name BCI as a vector; a follow-on European Data Protection Board or MDCG guidance specifically on AI-BCI data quality is the closest EU analogue that could emerge in the coming quarters. The FDA CDRH’s next AI/ML Software Action Plan update is the closest US analogue.

Whether ISO/TC 215 (health informatics) or ISO/IEC JTC 1/SC 42 (AI) takes up the AI-BCI training-data quality question at the international standards level. National regulators moving faster than international standards is a common pattern; a subsequent international standard that closely tracks the Chinese specification would be a notable outcome.

Whether the 1 September 2027 effective date is enforced strictly, extended, or accompanied by transitional exceptions for named operator classes. NMPA’s 12-month compliance runway is typical; deviations from that runway would signal how tightly the regulator intends to hold the field to the new expectation.

Whether Yuan Peng, Li Shu, or Zheng Jia give further public interviews framing the standard’s intent, or whether NMPA publishes an English translation of the standard text to support international regulatory dialogue. Both would extend the standard’s soft-power reach beyond the Chinese domestic market.

Sources

Primary and near-primary coverage:

Regulatory framework references:

Cross-reference to prior InsideBCI coverage:

policy China NMPA NIFDC AI-BCI-standard EEG-data-quality Yuan-Peng Li-Shu Zheng-Jia medical-device-standard Industrial-Builder-posture US-China-regulatory-competition FDA EU-AI-Act MIIT-15th-Five-Year-Plan Neuracle STAR-Market Issue-01