Non-invasive BCI: Advanced, Optimal, and Cost-effective Implementation Report

Contents

Sections 1-9 are the main report; Sections 10-11 are controlled appendices.

• Executive summary
• 1. Non-invasive BCI from first principles
• 2. Evaluation method and evidence boundaries
• 3. Global technology landscape
• 4. Frontier implementation
• 5. Optimal implementation
• 6. Cost-effective implementation
• 7. BOM, software stack, and implementation
• 8. Staged roadmap and stop conditions
• 9. Risk, ethics, and regulatory boundaries
• 10. Critical claim audit (source years: 2000-2026)
• 11. Authoritative source index (source years: 2000-2026)

Executive summary

This report separates three questions that are often collapsed into one. If the objective is to observe the widest practical range of physiology, the frontier research architecture is high-density professional EEG plus fNIRS, unified synchronization, and preregistered cross-day validation. It increases observable signals and also increases setup, timing, artifact, and ergonomic burdens. It is a maximum-observability architecture, not evidence that a hybrid system is invariably more accurate. S0063, S0064, and S0065 support acquisition practices, hardware-integration facts, and the limits of the evidence. They do not establish a universal hybrid advantage.

For a research system that can be maintained, audited, and improved over time, the optimal implementation is 32-64-channel professional EEG with interchangeable wet/semi-dry electrodes, hardware events, LSL stream coordination, and an analysis stack based on MNE, MNE-BIDS, MOABB, Braindecode, and pyRiemann. CSP, FBCSP, xDAWN, and Riemannian methods remain strong baselines. Every learned transformation is fit inside the training fold. Results are reported under held-out-subject, held-out-session, and prospective online evaluation. This is a risk-adjusted engineering choice, not a claim that more channels always perform better. Vendor source S0069 establishes only that 32-128-channel R-Net configurations were offered at a point in time.

For a constrained budget, the cost-effective implementation is open 8-16-channel EEG, limited to teaching and low-consequence prototypes. It is appropriate for learning the acquisition chain, event marking, basic decoding, and reproducible experiments. It should not be carried directly into long-term human use, clinical decisions, or consequential closed-loop control. The 2026 point-in-time pages list an 8-channel board at USD 1,249 in S0074 and a 16-channel research bundle at USD 4,499 in S0076. Tax and shipping, consumables and support are additional, configurations can change, and neither page establishes medical certification. Procurement source S0095 is CNY 438,000 for one specific awarded set. S0097 is CNY 1,891,500 for a mixed multi-model package. These amounts cannot be directly compared or treated as general retail prices.

The evidence cutoff is 2026-07-24. The clinical discussion presents together a positive protocol-specific multicenter trial S0194, a meta-analysis with substantial heterogeneity and a non-significant MBI result S0195, and a randomized pilot with no significant between-group difference S0203. The combined evidence does not support universal efficacy. Standards and regulatory statements are also date-bounded: three 2026 GB/T records were not yet effective at the cutoff and are recommended standards; the MIIT item was a submission notice. The NMPA route depends on product facts, intended use, and closed-loop function.

This English edition is an unofficial AI-assisted translation of the controlled Chinese report. It requires professional review before publication or regulatory use. The original sources and official Chinese regulatory texts control where wording differs.

1. Non-invasive BCI from first principles

1.1 What a BCI does

A brain-computer interface, or BCI, converts measurable neural activity into a signal that a machine can use. The full chain begins with physiology, continues through sensors and time-aligned stimulus or behavior records, then passes through cleaning, feature extraction, and a model. An application displays the result or controls a device. When that application stimulates the user in return, the system is closed loop. A high model score cannot repair an untraceable event marker, a drifting clock, or corrupted raw data.

An invasive BCI places electrodes inside the skull or neural tissue. It may obtain signals closer to neuronal populations, while surgery, implanted materials, and longitudinal follow-up introduce a different risk profile. A non-invasive BCI does not penetrate skin or skull. EEG and fNIRS are common modalities. Avoiding surgery reduces one class of risk, but scalp, hair, movement, ambient electrical noise, and person-to-person variability remain. Non-invasive does not mean risk-free, and a consumer product is not automatically suitable for diagnosis or treatment.

1.2 EEG, fNIRS, and hybrid acquisition

EEG measures voltage differences at scalp electrodes. Its high temporal resolution is useful for stimulus-locked activity, oscillatory changes, and online control. Spatial interpretation is limited by volume conduction and head anatomy. Functional near-infrared spectroscopy, or fNIRS, estimates hemodynamic changes through light absorption. Its response is slower and provides a different physiological view. A defensible fNIRS record retains raw intensity, wavelength and geometry metadata, and short-distance channels rather than only processed hemoglobin curves S0063.

EEG-fNIRS places the modalities in the same task. A hybrid system is more than two result files. Optodes and electrodes must coexist without unstable pressure or obstruction. Event paths, device clocks, and correction steps must be recorded. Weight, heat, donning time, and participant movement need evaluation. Existing studies are small and heterogeneous. The report therefore recommends hybrid acquisition only for maximum observability and does not claim that hybrid is necessarily more accurate S0064 S0065.

1.3 Wet, dry, and semi-dry electrodes

Wet electrodes use conductive gel or electrolyte to establish contact. Preparation and cleaning take time, but wet acquisition remains the signal reference. Dry electrodes remove gel and may shorten setup, while hair, local pressure, motion, and contact changes can raise noise. Semi-dry electrodes use a small or slowly released amount of fluid. Evaporation and long-wear behavior still depend on the design S0049 S0051.

Impedance is one indicator of contact; it is not sufficient evidence of signal quality. Active electronics, amplifier input, and environmental noise change the relationship between impedance and usable signal S0052. A dry or semi-dry option is validated against wet electrodes on the same cap, in the same paradigm and participants, over multiple sessions. Acceptance includes usable-data yield, spectra, artifacts, task behavior, setup time, pressure points, comfort, cleaning, and maintenance. Reading a single impedance indicator is not a validation protocol.

1.4 Four common paradigms

An event-related potential, or ERP, is a time-locked EEG change associated with repeated events. P300 is a commonly used ERP component, often elicited when a user attends to an infrequent target. It depends on stimulus design and usually requires calibration. Low-calibration research asks whether personal data can be reduced under a specific protocol; it does not establish calibration-free operation S0200.

Steady-state visual evoked potential, or SSVEP, uses periodic visual stimulation. code-VEP distinguishes targets with coded sequences. Both may support useful information transfer, but visual comfort, fatigue, stimulus safety, and asynchronous rejection must be evaluated. A five-class code-VEP study with eight dry electrodes demonstrates low-density feasibility for that task, not universal reliability across visual tasks or home populations S0192.

Motor imagery asks a user to imagine movement and decodes changes in sensorimotor rhythms. It avoids continuous flicker but often has substantial individual variability and benefits from training and feedback. A four-channel online study provides low-density feasibility evidence; healthy-participant classification accuracy is not rehabilitation benefit S0199. A passive BCI estimates a state such as attention, workload, or fatigue without an intentional command. Such estimates are easy to overinterpret and require explicit uncertainty and use limits.

1.5 Reading metrics

Classification accuracy is the proportion of correct predictions. With class imbalance it can conceal failure on a minority class, so confusion matrices, balanced measures, or class-specific results are needed. Information transfer rate, or ITR, combines the number of choices, correctness, and selection duration under assumptions that must be disclosed. It should not be compared across incompatible tasks without reconstructing the protocol. Latency is the time from an event to an output; jitter is the variability of that time. They are measured separately. Rejection rate records how often a system declines to output. In asynchronous control, rejection is part of safety and usability rather than merely a lost score.

Setup time runs from seating the participant to achieving recording conditions. Comfort covers pressure, heat, visual fatigue, and acceptable wear duration. Safety includes electrical, stimulation, skin, fall, erroneous-control, and data risks. Total cost of ownership, or TCO, includes hardware, electrodes and consumables, tax and shipping, software and support, spares, training, synchronization validation, safety testing, ethics, governance, and operations. A board list price does not determine TCO. Clinical endpoints concern patient benefit; consumer endpoints often concern convenience or experience. They are not interchangeable.

2. Evaluation method and evidence boundaries

2.1 Different objective functions

“Frontier” is evaluated by research observability and tolerates complexity only when additional information can be synchronized, retained, and audited. “Optimal” is evaluated by risk-adjusted long-term engineering value, including stable acquisition, maintainability, cross-day generalization, and reproducibility. “Cost-effective” is evaluated by learning and prototyping value in low-consequence settings. It cannot become inexpensive by silently dropping safety and validation. Because the objectives differ, the report does not produce a single device leaderboard.

The evidence grades follow the controlled source table. Grade A generally covers official material, cataloged standards, or stronger clinical and human evidence. Grade B covers methods, reviews, datasets, and independent validation. Grade C provides procurement context. Grade D is point-in-time vendor evidence. These grades are not a mechanical score. A grade A clinical trial remains limited to its population, intervention, comparator, endpoint, and follow-up. A grade D page can show that a configuration or quoted price existed; it cannot prove efficacy.

2.2 No cross-paradigm leaderboard

Finger-class control, continuous two-dimensional tracking, code-VEP, P300, and motor imagery differ in stimulation, participant selection, calibration, and endpoint. Classification accuracy, continuous error, and online selection speed do not share a common denominator. Frontier studies S0191 and S0197 demonstrate different forms of real-time feasibility. Low-density studies S0192, S0199, and S0200 answer other questions. Results are interpreted within protocol, not ranked across paradigms.

2.3 Preventing information leakage

Information leakage occurs when test information directly or indirectly influences training. Examples include computing a mean on all sessions before cross-validation, choosing channels after inspecting all data, learning artifact thresholds from the test fold, or repeatedly tuning after viewing test results. Such scores may appear stable and then fail in deployment S0105.

Filtering parameters that are estimated from data, bad-channel handling, artifact thresholds, normalization, feature selection, and hyperparameter search belong inside the training pipeline. Each training fold is fit independently; transformations are then applied to validation or test folds. PREP and Autoreject can improve consistency, but any learned behavior must obey the same fold boundary S0120 S0121.

2.4 Generalization and reproducibility

A random split of neighboring segments from the same session does not answer whether a system works for another person or on another day. Reports include held-out-subject and held-out-session results. An online use case also receives a prospective evaluation with the model frozen before data collection. Multi-day data S0193 and accumulated-session evidence S0197 show why session history and individual state matter.

Data are organized using BIDS principles at acquisition. Raw data are append-only. The event dictionary, channel table, hardware and firmware, reference scheme, exclusion reasons, code version, random seeds, model artifacts, and audit logs travel with the record. Sharing requires a separate check of consent scope, de-identification, access control, and rights. A reproducibility package should let another analyst regenerate the main conclusion from raw data, not merely inspect a figure.

3. Global technology landscape

3.1 Frontier capability remains task-specific

High-density EEG research now covers discrete classes and continuous control. S0191 used screened, healthy, experienced participants and subject-specific same-day fine-tuning for finger classes. S0197 studied multi-session two-dimensional continuous tracking. These results establish feasibility for their real-time chains. They do not establish long-term patient home use or a universal device winner.

Low-density EEG can also work within a suitable paradigm. There is controlled evidence for eight-dry-electrode code-VEP, four-channel motor imagery, and low-calibration P300. The limitations are concrete: visual stimuli, healthy populations, calibration procedures, online endpoints, and device generations differ. Fewer channels may reduce setup time, while reducing spatial coverage and the ability to diagnose artifacts. The tradeoff must be tested over the intended task and multiple sessions.

3.2 Algorithms have no data-independent champion

CSP is a standard motor-imagery spatial-filter baseline S0110. FBCSP combines CSP across frequency bands S0111. xDAWN enhances evoked responses for ERP and P300 tasks S0112. Riemannian methods model covariance structure S0114. Their computational burden is manageable, and they are strong baselines rather than historical decorations.

Deep models can learn temporal and spatial representations. A deep model is a candidate rather than a default winner. A comparison is interpretable only under the same split, tuning budget, and latency constraint. It also reports calibration cost, rejection, cross-session stability, ablation results, and prospective online performance. An improvement in a task-specific paper does not establish that deep learning is universally preferable.

3.3 Clinical and consumer boundaries

Stroke rehabilitation findings must be read together. S0194 observed improvement under a specific multicenter protocol. S0195 reported substantial heterogeneity and a non-significant MBI result. S0203 found no significant treatment-group difference in a randomized pilot. Populations, interventions, comparisons, and endpoints differ. The defensible conclusion is to validate each protocol, not to claim universal efficacy.

Consumer devices can support self-experimentation, teaching, and low-consequence interaction prototypes. Independent bench comparison S0198 found differences in bandwidth, ERP waveform, and timing. It used an EEG phantom, not human clinical outcomes. Detecting a waveform does not establish medical electrical safety, clinical equivalence, continued supply, or regulatory status.

4. Frontier implementation

This chapter's answer is a high-density professional EEG plus fNIRS maximum-observability research architecture. It is intended to widen research observation, and it does not claim that hybrid is necessarily more accurate. Preregistration, cross-day validation, and end-to-end loopback are required before formal acquisition.

4.1 Definition and architecture

Frontier means maximum observability here, not the largest equipment count or highest isolated score. The implementation uses high-density professional EEG plus fNIRS, retains both raw streams and raw event records, adds a unified synchronization layer, and preregisters primary endpoints, exclusions, and cross-day analysis before recruitment. The cap, optodes, electrodes, and short-distance channels are checked on a phantom and through human fitting so that mechanical conflicts and fixed pressure points are identified early.

The EEG side records raw voltage, reference, channel locations, contact checks, and hardware events. The fNIRS side retains raw intensity, wavelengths, geometry, and short-distance channels rather than only derived hemoglobin signals. The stimulus machine, behavior device, and acquisition computers retain their local clock records. LSL can discover and coordinate streams and perform time correction, but LSL is not a shared hardware clock or deterministic medical bus S0066. Timing acceptance separates latency from jitter and uses a photodiode, loop wire, or traceable event for an end-to-end loopback S0068.

4.2 Experiment and analysis

The first experiments do not begin with a fused model. They establish EEG-only, fNIRS-only, and synchronization baselines, then compare early, late, or decision-level fusion. If a unimodal timeline cannot be trusted, a fusion result has no auditable foundation. All models use identical participant, session, and trial splits. Preprocessing is fit inside the training fold. Main results show unimodal and hybrid outputs so that an advantageous subset is not silently selected.

Cross-day work measures fitting repeatability, cap displacement, skin and hair conditions, and fatigue. Preregistration defines the endpoint, analysis window, missing-data rule, and stop rule. Raw data, BIDS metadata, code, environment lockfiles, and audit logs form a versioned package. A result that cannot be traced to source events is excluded from the primary conclusion.

4.3 Appropriate and inappropriate use

This architecture is appropriate for studying complementary electrical and hemodynamic signals, comparing fusion methods, or constructing a reusable multimodal dataset. It is a poor match when minimal setup time is the dominant goal. It cannot move into clinical use simply because it has more sensors. If synchronization loopback repeatedly fails, short-distance channels are absent, or fitting burden causes unacceptable participant refusal, the intended observability has not been achieved.

5. Optimal implementation

5.1 Hardware selection

The recommended acquisition system is 32-64-channel professional EEG with hardware event input, stable drivers, and raw-data export. It uses interchangeable wet/semi-dry electrodes; wet electrodes remain the reference. The semi-dry option is admitted only after same-cap, same-paradigm, multi-session paired testing. The range is an engineering selection, not a physiological law and not proof that more channels perform better. Point-in-time vendor record S0069 supports only the existence of R-Net 32-128 configurations and requires procurement verification.

Hardware events should enter the amplifier or a validated trigger interface from the stimulus system. LSL handles stream coordination, metadata, and time correction. A shared trigger or hardware synchronization path is retained when devices provide one. The acquisition computer has controlled operating-system, power, USB topology, and driver versions. End-to-end loopback is rerun for each release; a single laboratory measurement is not a permanent guarantee.

5.2 Software stack and baselines

MNE reads, visualizes, and processes signals. MNE-BIDS writes raw data and metadata into an auditable BIDS layout. MOABB provides standardized evaluation protocols for public benchmarks. Braindecode supplies candidate deep models. pyRiemann supplies Riemannian baselines. Lockfiles or containers preserve the environment, and data and code versions are written into audit logs.

Motor imagery begins with CSP, FBCSP, and Riemannian methods. ERP and P300 begin with xDAWN plus a simple classifier. A deep model advances only if it is stable against strong baselines under identical splits and tuning budgets without unacceptable calibration, latency, or cross-session collapse. Before prospective collection, preprocessing and parameters are frozen.

5.3 Acceptance measures

Reports include ITR, end-to-end latency, jitter, rejection rate, setup time, usable-data yield, comfort, and adverse events in addition to accuracy. Offline evaluation includes held-out-subject and held-out-session protocols. Online evaluation records failures and human intervention. Operational acceptance covers spares, replacement cycles, cleaning, consumables, driver updates, and data recovery. Every metric includes a definition, denominator, and distribution or uncertainty rather than only a mean.

6. Cost-effective implementation

6.1 Use boundary

The recommendation is open 8-16-channel EEG only for teaching and low-consequence prototypes. Suitable activities include observing alpha rhythm, learning event marking, reproducing a simple P300, SSVEP, or motor-imagery baseline, and testing a software stream. A classroom demonstration is not converted into a diagnostic report, and an unvalidated prototype does not control equipment that could injure a user.

“Open” generally makes interfaces, hardware, or software easier to study. It does not itself establish fidelity, electrical safety, clinical equivalence, or durable supply. Independent comparisons and task evidence S0060 S0198 take priority over brand claims. After purchase, the system is tested with a phantom or known signal and timing loopback before any ethics-approved low-risk human work.

6.2 Reading costs

S0074 lists the 8-channel board at USD 1,249 at a 2026 point in time. S0076 lists the 16-channel research bundle at USD 4,499. These vendor prices exclude tax and shipping; consumables and support are additional. Configurations may change. The pages are not evidence of medical certification and do not prove performance.

S0095 records CNY 438,000 for one awarded set at a specific institution. S0097 records CNY 1,891,500 for a mixed package with multiple models and quantities. Dates, service, warranty, and configuration differ. The package total cannot be divided by an assumed count or directly compared with a bare-board web price.

6.3 Upgrade triggers

Moving from teaching to long-term wear, patient research, clinical decisions, or closed-loop control triggers a new product-classification, ethics, safety, signal-quality, multi-session, and operations review. An inexpensive device with repeated recollection, manual data repair, or unstable timing may have higher TCO than a professional system. Cost-effectiveness is the total cost of completing an acceptable task, not the lowest acquisition price.

7. BOM, software stack, and implementation

7.1 Controlled BOM

No estimate is invented for professional EEG, fNIRS, or generic electrodes because the ledger has no configuration-matched auditable quote. CNY 438,000 in S0095 is one specific set, and CNY 1,891,500 in S0097 is a mixed multi-model package. They provide procurement context only. Tax and shipping, consumables and support, software licensing, spares, and validation services are separate quote lines. Medical certification is checked for the exact model and intended use.

7.2 Software and data

The acquisition layer preserves device-native raw files and emits coordinated LSL streams. The event layer retains hardware markers and application logs. MNE reads the source data; MNE-BIDS builds the BIDS dataset and sidecars; MOABB applies standardized evaluation protocols to public benchmarks; pyRiemann, CSP/FBCSP, and xDAWN provide conventional baselines; Braindecode supplies deep candidates. The automated pipeline produces quality reports, metrics, configuration snapshots, and audit logs.

The data package contains raw data, BIDS data, an immutable event dictionary, analysis configuration, environment lockfile, models, results, and access records. Identity mappings and neural data have separate permissions. A de-identified copy is created for approved sharing. Deletion, withdrawal, and correction are logged rather than silently overwriting source data.

7.3 Implementation sequence

First freeze intended use and risk: user, population, task, output, consequence of error, and whether the loop acts on the person. Build the bench chain next, connecting known signal, event source, acquisition, LSL, and analysis while measuring latency and jitter. A small healthy-adult pilot then checks cap fit, skin, visual burden, cleaning, data integrity, and stopping procedures. Only after that pilot does the project enter preregistered multi-session acquisition and frozen evaluation.

Algorithm work starts with simple baselines. Splits are generated and versioned before any learned operation. Preprocessing, thresholds, normalization, features, and tuning are fit inside the training fold. Evaluation then moves through within-person cross-session, held-out-session, held-out-subject, and prospective online conditions. If a deep model does not provide stable value at the same budget, the simpler model remains. Handover includes controlled BOM, serial numbers, firmware, data dictionary, code, test records, risk register, and operations manual.

8. Staged roadmap and stop conditions

Each phase lists its deliverables, an entry gate, and stop conditions. These are qualitative control gates rather than invented medical thresholds.

0–3 months: definition and bench loop

Deliverables: an intended-use and risk statement, selection records for the three architectures, controlled BOM, event dictionary, phantom or known-signal tests, end-to-end loopback report, sample BIDS data, and initial ethics and rights review.

Entry gate: devices export raw data; the event source is traceable; the software environment can be rebuilt; owners have documented the research, consumer, and clinical boundaries.

Stop conditions: loopback synchronization does not pass; raw data or events cannot be traced; the supplier cannot provide information required for safety and rights checks; the team cannot produce an actionable ethics submission. The system returns to bench work or a narrower use rather than entering human acquisition.

3–12 months: multi-session validation

Deliverables: a preregistration, same-cap electrode comparison, multi-session dataset, quality report, conventional strong baselines, training-fold-contained pipeline, held-out-subject and held-out-session results, and a frozen prospective test plan.

Entry gate: every open 0–3-month issue has an owner and closure evidence; timing and data-integrity tests pass repeatedly; ethics approval, informed consent, privacy controls, and access controls are ready.

Stop conditions: cross-session performance collapses and cannot be corrected with a predefined maintenance procedure; participant refusal or fitting burden is unacceptable; artifacts make the primary endpoint uninterpretable; an uncontrolled safety event occurs; leakage is found and the frozen raw data cannot support a clean rerun.

12–24 months: prospective and use-context validation

Deliverables: prospective online results from the frozen system, target-context operating records, exception and rejection logs, actual TCO ledger, maintenance and training material, and a written decision about whether a medical-device route is implicated.

Entry gate: multi-session conclusions reproduce; endpoints and stop rules are frozen; target-context safety, ethics, data rights, and operational ownership are approved.

Stop conditions: prospective results fail to reproduce development findings; erroneous output has unacceptable consequences in context; ethics, safety, or rights checks cannot be completed; a closed-loop change alters classification without reassessment; consumables, support, or staffing make TCO unsustainable. Stopping prevents a system with insufficient evidence from widening its impact.

9. Risk, ethics, and regulatory boundaries

9.1 Human participation and data governance

Informed consent explains the collected signals, purpose, plausible errors, retention, sharing, withdrawal route, and post-withdrawal handling in language participants can understand. Patients, minors, and other vulnerable groups require added protection. Identity mappings and neural data are stored separately, access follows least privilege, and exports enter audit logs. A vague research consent does not automatically authorize a new use S0164 S0211.

Model accuracy cannot replace safety engineering. For medical-use EEG, product classification and risk analysis determine the applicable editions and validation for general IEC safety, EEG-specific essential performance, and EMC. The verified direct catalogs are S0091 (IEC 80601-2-26 Edition 1.1), S0092 (IEC 60601-1 Edition 3.2), and S0093 (IEC 60601-1-2 Edition 4.1). These catalog editions were verified on 2026-07-24, but they do not by themselves determine domestic adoption or the product-applicable edition. Ordinary human-subject research does not automatically become a medical device, but institutional ethics review, safety design, and risk control remain necessary.

9.2 Status of 2026 standards

As of 2026-07-24, S0204, S0205, and S0206 were not yet effective and are recommended standards. Catalog entries are not mandatory NMPA rules. S0207 was a submission notice, not a final standard, effective rule, or medical-device approval. Status requires re-verification when effective dates arrive.

9.3 NMPA and clinical statements

The 2026 NMPA announcement S0216, classification-guidance text S0217, and official interpretation S0219 require classification to follow product facts, intended use, and closed-loop function. A research architecture recommendation is not a classification outcome or registration approval.

External clinical statements present the mixed S0194, S0195, and S0203 evidence together and specify population, intervention, comparator, endpoint, and follow-up. Consumer demonstrations do not use claims such as therapeutic effect or medical grade unless the exact product and use have corresponding evidence and approval. This unofficial English version still requires professional review; the controlled Chinese report, original regulatory texts, and original research sources govern.

10. Critical claim audit (source years: 2000-2026)

The following entries reproduce the 21 controlled claims. Source IDs and scope retain the controlled source-table wording. Four vendor limitations are complete English semantic renderings of the controlled Chinese boundaries.

N001 · critical · frontier · 2026-07-24

Frontier EEG has shown two distinct real-time control feasibility results: S0191 used n=21 screened healthy experienced users with subject-specific same-day fine-tuning for finger classes, while S0197 studied n=28 in multi-session two-dimensional tracking.

Source IDs: S0191;S0196;S0197

Scope: Frontier research-grade EEG feasibility

Limitations: These are task-specific healthy-user studies and do not establish patient home-use effectiveness or a universally best device.

N002 · high · frontier · 2026-07-24

Low-density EEG also has within-paradigm feasibility evidence: S0192 used eight dry electrodes for five-class code-VEP, S0199 used four-channel motor imagery, and S0200 studied low-calibration P300; they answer different questions.

Source IDs: S0192;S0199;S0200

Scope: Low-density frontier EEG feasibility

Limitations: Stimuli, populations, calibration, and endpoints differ; none alone establishes long-term clinical or home reliability.

N003 · high · frontier · 2026-07-24

S0191, S0192, S0197, S0199, and S0200 differ in paradigm, participant selection, calibration, and endpoint, so their accuracy or error metrics must not be directly ordered across paradigms.

Source IDs: S0191;S0192;S0197;S0199;S0200

Scope: Evidence-comparison boundary

Limitations: Only within-protocol interpretation is supported; no cross-paradigm performance winner is claimed.

N004 · critical · frontier · 2026-07-24

High-density EEG plus fNIRS can be used as a maximum-observability research architecture: fNIRS should retain raw intensity and short-distance channels, and hybrid hardware, synchronization, and ergonomics must be validated; small heterogeneous studies do not support a universal advantage.

Source IDs: S0063;S0064;S0065

Scope: Maximum-observability EEG-fNIRS research architecture

Limitations: This is a research observability recommendation, not proof of clinical superiority, product certification, or better outcomes than EEG alone.

N005 · high · optimal · 2026-07-24

Wet electrodes should remain the signal reference; dry and semi-dry electrodes require paired validation using the same cap, paradigm, and multiple sessions, and impedance is not a sufficient condition for signal quality.

Source IDs: S0049;S0051;S0052;S0057

Scope: EEG acquisition and electrode validation

Limitations: Comfort, setup time, hair type, motion, wear duration, noise, and task performance must be assessed together.

N006 · high · optimal · 2026-07-24

LSL can coordinate streams and correct time, but it is not a shared hardware clock or a deterministic medical bus; synchronization acceptance must separate latency from jitter and use end-to-end loopback tests.

Source IDs: S0066;S0068

Scope: Multimodal acquisition synchronization

Limitations: Performance remains dependent on devices, operating systems, networks, and the validated end-to-end configuration.

N007 · critical · optimal · 2026-07-24

For medical-use EEG, applicable editions and validation for IEC general safety, EEG-specific essential performance, and EMC should follow product classification and risk analysis. Ordinary human-subject research does not automatically become a medical device, but it still requires institutional ethics review, safety, and risk controls. Algorithm accuracy cannot replace these duties.

Source IDs: S0091;S0092;S0093;S0164

Scope: Safety ethics and medical engineering

Limitations: Catalog records and ethics guidance do not determine the applicable edition, national adoption, product classification, or approval outcome.

N008 · high · optimal · 2026-07-24

Traditional CSP, FBCSP, xDAWN, and Riemannian methods must serve as strong baselines; modern models are comparable only under the same data splits and tuning budget.

Source IDs: S0110;S0111;S0112;S0114

Scope: Algorithm baseline contract

Limitations: Baseline choice remains paradigm-specific and does not imply that any one method wins every task.

N009 · critical · optimal · 2026-07-24

All learned preprocessing, artifact thresholds, normalization, feature selection, and hyperparameter fitting must occur inside the training fold before application to validation or test folds.

Source IDs: S0105;S0120;S0121

Scope: Leakage-resistant algorithm validation

Limitations: A high score from a leaked split is not valid evidence of generalization.

N010 · high · optimal · 2026-07-24

Validation must report held-out-subject and held-out-session results; the multi-day data in S0193 and accumulated-session results in S0197 show that random within-session splits cannot replace cross-person and cross-day tests.

Source IDs: S0193;S0197

Scope: External and temporal validation

Limitations: Healthy research cohorts do not establish patient home deployment or prospective clinical validity.

N011 · medium · optimal · 2026-07-24

Deep models are candidate algorithms rather than default winners; they should be judged using strong baselines, ablations, calibration cost, latency, cross-session stability, and prospective online results.

Source IDs: S0196;S0197

Scope: Frontier algorithm assessment

Limitations: Narrative reviews and task-specific studies do not support a universal architecture winner.

N012 · critical · optimal · 2026-07-24

The point-in-time vendor specification in S0069 records available R-Net 32-128-channel configurations; within that configuration-existence boundary, 32-64-channel professional EEG is the risk-adjusted engineering selection because it balances spatial coverage, maintainable acquisition, synchronization, and cross-session validation. This does not prove that more channels perform better.

Source IDs: S0057;S0066;S0069;S0105;S0193;S0196

Scope: Overall optimal non-invasive BCI engineering architecture

Limitations: S0069 is a point-in-time vendor self-report that supports only the existence of the configuration and requires procurement verification; tax and shipping are excluded; consumables and support are additional; it does not establish medical certification or performance.

N013 · critical · optimal · 2026-07-24

Clinical evidence must present the mixed results together: S0194 found a protocol-specific improvement in a multicenter trial, S0195 reported substantial heterogeneity and a non-significant MBI result, and the S0203 randomized pilot found no significant between-group difference; universal efficacy cannot be claimed.

Source IDs: S0194;S0195;S0203

Scope: Clinical evidence balance

Limitations: Effects remain population, intervention, comparator, endpoint, and follow-up specific.

N014 · high · optimal · 2026-07-24

Non-invasive BCI research and implementation must include informed consent, vulnerable groups, neural-data privacy, purpose limitation, withdrawal, and access control in ethics and data governance; neither ethics guidance nor personal-information law is product approval.

Source IDs: S0164;S0211

Scope: Ethics privacy and governance

Limitations: Applicable duties depend on research product and data-processing facts and require case-specific legal and ethics review.

N015 · high · optimal · 2026-07-24

As of 2026-07-24, the three 2026 GB/T records S0204-S0206 were not yet effective and are recommended standards; S0207 is only a submission notice, not a final standard, effective rule, or medical-device approval.

Source IDs: S0204;S0205;S0206;S0207

Scope: Standards and regulatory status

Limitations: Future effective dates or later final texts require re-verification and do not retroactively change this cutoff-date statement.

N016 · high · optimal · 2026-07-24

The 2026 NMPA announcement, classification-guidance text, and official interpretation require the regulatory route to follow product facts, intended use, and closed-loop function; a research architecture recommendation is not a classification conclusion or registration approval.

Source IDs: S0216;S0217;S0219

Scope: Medical-device regulatory boundary

Limitations: The announcement guidance and interpretation do not approve any specific device or make every non-invasive research system a medical device.

N017 · critical · cost_effective · 2026-07-24

The 2026 vendor pages provide only point-in-time quotes and specifications: S0074 lists an 8-channel board at USD 1249, and S0076 lists a 16-channel research bundle at USD 4499.

Source IDs: S0074;S0076

Scope: Point-in-time vendor quote evidence

Limitations: These are point-in-time vendor web quotes; tax and shipping are excluded; consumables and support are additional; the pages do not establish medical certification or performance.

N018 · high · cost_effective · 2026-07-24

S0095 records CNY 438000 for one awarded set, while S0097 records CNY 1891500 for a mixed multi-model package; they provide procurement context and are not general retail prices.

Source IDs: S0095;S0097

Scope: Public procurement context

Limitations: Configuration, quantity, service, warranty, institution, and bidding period differ and prevent direct unit-price comparison.

N019 · high · cost_effective · 2026-07-24

Lower-cost acquisition systems require independent signal-fidelity and task-suitability validation; open hardware may work in specific tasks without establishing medical electrical safety, clinical equivalence, or a universal brand winner.

Source IDs: S0060;S0198

Scope: Independent cost-effective hardware validation

Limitations: Device versions, firmware, phantom tests, and task protocols limit transfer to current human clinical use.

N020 · critical · cost_effective · 2026-07-24

The point-in-time specifications and quotes in S0074/S0076 record 8/16-channel configurations; within that configuration-existence boundary, open 8-16-channel EEG is recommended only for teaching and low-consequence prototypes, while long-term human use, clinical decisions, or closed-loop control require stronger safety, acquisition quality, multi-session validation, and operational support.

Source IDs: S0060;S0074;S0076;S0192;S0198;S0199

Scope: Low-consequence cost-effective architecture

Limitations: The point-in-time vendor specifications and quotes are self-reported; tax and shipping are excluded; consumables and support are additional; they do not establish medical certification or performance.

N021 · high · cost_effective · 2026-07-24

Total cost of ownership cannot be derived directly from a point-in-time vendor board quote or a single procurement price; quote and specification audits must also include electrodes and consumables, tax and shipping, software and support, spares, training, synchronization validation, safety testing, and operations.

Source IDs: S0074;S0076;S0095;S0097;S0198

Scope: Implementation and total-cost boundary

Limitations: Point-in-time vendor quotes cannot be directly combined; tax and shipping are excluded; consumables and support must be itemized; the quotes do not establish medical certification or performance.

11. Authoritative source index (source years: 2000-2026)

This index provides bibliographic metadata, direct original links, and concise evidence boundaries. translated means a licensed local translation is recorded, review_required means professional review is still required, and not_permitted means no derivative full-text translation was made. An entry never presents unlicensed full text as a local mirror.

The ledger retains two official HTTP originals: S0097 and S0211. Read-only probes of their HTTPS candidates on 2026-07-24 produced a 403 anti-bot response and a TLS handshake failure, respectively. The report therefore preserves the ledger URLs. Users should verify the official domain and must not treat these protocol exceptions as permission to use an unofficial mirror.

S0049 · 2014 · B

Dry EEG Electrodes. Original: S0049. Role: Dry-electrode interface review. Limitation: Older review; does not validate a current product or long-duration use. Full-text/rights boundary: translated

S0051 · 2020 · B

Review of semi-dry electrodes for EEG recording. Original: S0051. Role: Semi-dry electrode review. Limitation: Evaporation and long-wear risks remain device-specific. Full-text/rights boundary: not_permitted

S0052 · 2020 · B

Impedance and Noise of Passive and Active Dry EEG Electrodes: A Review. Original: S0052. Role: Impedance and noise boundary. Limitation: Impedance alone is not sufficient evidence of signal quality. Full-text/rights boundary: not_permitted

S0057 · 2018 · B

Signal Quality Evaluation of Emerging EEG Devices. Original: S0057. Role: Independent signal-quality validation. Limitation: Device generations and test conditions limit generalization. Full-text/rights boundary: translated

S0060 · 2016 · B

Comparison of an Open-hardware EEG Amplifier with Medical Grade Device in BCI Applications. Original: S0060. Role: Open-hardware comparison. Limitation: Specific tasks only; does not establish electrical safety or clinical equivalence. Full-text/rights boundary: not_permitted

S0063 · 2021 · B

Best practices for fNIRS publications. Original: S0063. Role: fNIRS reporting and acquisition guidance. Limitation: Reporting guidance is not product certification; raw intensity and short-distance channels must be retained. Full-text/rights boundary: not_permitted

S0064 · 2021 · B

Wearable Integrated EEG-fNIRS Technologies: A Review. Original: S0064. Role: Hybrid hardware integration review. Limitation: Hardware integration synchronization and ergonomics differ among systems. Full-text/rights boundary: translated

S0065 · 2021 · B

Systematic review on hybrid EEG-fNIRS in BCI. Original: S0065. Role: Hybrid evidence boundary. Limitation: Studies are small and heterogeneous and do not show a universal hybrid advantage. Full-text/rights boundary: not_permitted

S0066 · 2025 · B

The lab streaming layer for synchronized multimodal recording. Original: S0066. Role: LSL architecture and time correction. Limitation: Software coordination and time correction are not a shared hardware clock or deterministic medical bus. Full-text/rights boundary: not_permitted

S0068 · 2023 · B

Two common issues in synchronized multimodal recordings with EEG: jitter and latency. Original: S0068. Role: Jitter and latency methods evidence. Limitation: Protocol-specific evidence; end-to-end loopback validation remains necessary. Full-text/rights boundary: not_permitted

S0069 · 2026 · D

actiCHamp Plus and R-Net official product information. Original: S0069. Role: Vendor evidence that R-Net 32-128 configurations exist. Limitation: Point-in-time vendor self-report supports configuration existence only; procurement verification, tax and shipping, consumables and support, certification, and performance remain outside the claim. Full-text/rights boundary: not_permitted

S0074 · 2026 · D

OpenBCI Cyton 8-channel biosensing board product page. Original: S0074. Role: 2026 vendor point-in-time Cyton quote. Limitation: Point-in-time USD 1249 page quote; tax and shipping and consumables and support are additional; not medical certification; does not prove performance. Full-text/rights boundary: not_permitted

S0076 · 2026 · D

OpenBCI All-in-One Biosensing R&D Bundle product page. Original: S0076. Role: 2026 vendor point-in-time R&D bundle quote. Limitation: Point-in-time USD 4499 page quote; tax and shipping and consumables and support are additional; not medical certification; does not prove performance. Full-text/rights boundary: not_permitted

S0091 · 2024-02-14 · A

IEC 80601-2-26:2019+AMD1:2024 CSV electroencephalographs particular requirements, Edition 1.1, 2024-02-14. Original: S0091. Role: IEC 80601-2-26:2019+AMD1:2024 CSV EEG-specific safety catalog; Edition 1.1; 2024-02-14. Limitation: Official catalog metadata does not determine domestic adoption or the product-applicable edition; verify both under product classification and risk analysis. The standard full text is not redistributed. Full-text/rights boundary: not_permitted

S0092 · 2020-08-20 · A

IEC 60601-1:2005+AMD1:2012+AMD2:2020 CSV medical electrical equipment general requirements, Edition 3.2, 2020-08-20. Original: S0092. Role: IEC 60601-1:2005+AMD1:2012+AMD2:2020 CSV general safety catalog; Edition 3.2; 2020-08-20. Limitation: Official catalog metadata does not determine domestic adoption or the product-applicable edition; verify both under product classification and risk analysis. The standard full text is not redistributed. Full-text/rights boundary: not_permitted

S0093 · 2020-09-01 · A

IEC 60601-1-2:2014+AMD1:2020 CSV medical electrical equipment EMC requirements, Edition 4.1, 2020-09-01. Original: S0093. Role: IEC 60601-1-2:2014+AMD1:2020 CSV EMC catalog; Edition 4.1; 2020-09-01. Limitation: Official catalog metadata does not determine domestic adoption or the product-applicable edition; verify both under product classification and risk analysis. The standard full text is not redistributed. Full-text/rights boundary: not_permitted

S0095 · 2024-05-11 · C

Wireless digital EEG acquisition system award notice MT-24-04031. Original: S0095. Role: Single-set public procurement evidence. Limitation: CNY 438000 is one awarded set and is not a general retail price. Full-text/rights boundary: not_permitted

S0097 · 2024-11-13 · C

SCUT EEG physiological acquisition and stimulation systems award notice. Original: S0097. Role: Mixed-package public procurement evidence. Limitation: CNY 1891500 covers mixed quantities and systems and is not a single-model retail price. Full-text/rights boundary: not_permitted

S0105 · 2023 · B

Leakage and the reproducibility crisis in ML-based science. Original: S0105. Role: Leakage taxonomy and reproducibility boundary. Limitation: Cross-domain methods source rather than BCI performance evidence. Full-text/rights boundary: not_permitted

S0110 · 2000 · B

Optimal spatial filtering of single trial EEG during imagined hand movement. Original: S0110. Role: CSP baseline. Limitation: Classic method; tuning and split leakage remain risks. Full-text/rights boundary: not_permitted

S0111 · 2012 · B

Filter Bank Common Spatial Pattern algorithm. Original: S0111. Role: FBCSP baseline. Limitation: Hyperparameter selection can leak test information. Full-text/rights boundary: translated

S0112 · 2009 · B

xDAWN algorithm to enhance evoked potentials: application to brain-computer interface. Original: S0112. Role: xDAWN ERP baseline. Limitation: Classic offline evidence; implementation and validation split matter. Full-text/rights boundary: not_permitted

S0114 · 2017 · B

Riemannian geometry for EEG-based BCI: a primer and review. Original: S0114. Role: Riemannian baseline review. Limitation: Covariance methods remain sensitive to contamination and sample size. Full-text/rights boundary: not_permitted

S0120 · 2015 · B

PREP pipeline: standardized preprocessing for large-scale EEG analysis. Original: S0120. Role: Preprocessing pipeline boundary. Limitation: Pipeline parameters and fit boundaries can create leakage. Full-text/rights boundary: not_permitted

S0121 · 2017 · B

Autoreject: automated artifact rejection for MEG and EEG. Original: S0121. Role: Artifact-rejection fit boundary. Limitation: Learned thresholds leak if fitted outside the training fold. Full-text/rights boundary: not_permitted

S0164 · 2024-02-06 · A

脑机接口研究伦理指引发布页及附件. Original: S0164. Role: Official BCI ethics guidance. Limitation: Ethics guidance is not a medical-device registration rule or product approval. Full-text/rights boundary: not_permitted

S0191 · 2025-06-30 · A

EEG-based brain-computer interface enables real-time robotic hand control at individual finger level. Original: S0191. Role: Selected-user finger-control feasibility. Limitation: n=21 selected healthy experienced users after responder screening; subject-specific same-day fine-tuning; class decisions rather than simultaneous trajectories. Full-text/rights boundary: review_required

S0192 · 2024-12-03 · A

Leveraging textured flickers: a leap toward practical, visually comfortable, and high-performance dry EEG code-VEP BCI. Original: S0192. Role: Eight-dry-electrode code-VEP feasibility. Limitation: n=24 healthy participants; five-class stimulus-specific task; does not establish patient home long-term or cross-device performance. Full-text/rights boundary: not_permitted

S0193 · 2025-03-23 · B

A multi-day and high-quality EEG dataset for motor imagery brain-computer interface. Original: S0193. Role: Multi-day MI dataset and benchmark. Limitation: Healthy cohort; reported cross-validation does not prove held-out cross-subject or cross-session generalization. Full-text/rights boundary: review_required

S0194 · 2024-04-19 · A

Rehabilitation with brain-computer interface and upper limb motor function in ischemic stroke: A randomized controlled trial. Original: S0194. Role: Multicenter stroke RCT positive signal. Limitation: Open-label blank-controlled protocol with blinded outcomes; cannot infer effectiveness outside the protocol. Full-text/rights boundary: not_permitted

S0195 · 2025-03-03 · A

Effects of brain-computer interface based training on post-stroke upper-limb rehabilitation: a meta-analysis. Original: S0195. Role: Stroke rehabilitation meta-analysis. Limitation: Substantial heterogeneity and possible publication bias; MBI was not significant. Full-text/rights boundary: not_permitted

S0196 · 2025-01-28 · B

Non-Invasive Brain-Computer Interfaces: State of the Art and Trends. Original: S0196. Role: Non-invasive BCI state-of-the-art review. Limitation: Narrative review; cannot establish a best device or clinical effectiveness. Full-text/rights boundary: review_required

S0197 · 2024-04-30 · A

Continuous tracking using deep learning-based decoding for noninvasive brain-computer interface. Original: S0197. Role: Multi-session continuous tracking study. Limitation: n=28 healthy experienced users; two-dimensional virtual tracking; subject-specific accumulated-session models; not patient or home evidence. Full-text/rights boundary: review_required

S0198 · 2024-08-13 · B

Evaluation of consumer-grade wireless EEG systems for brain-computer interface applications. Original: S0198. Role: Independent consumer-device comparison. Limitation: EEG phantom rather than clinical outcomes; no universal brand winner; firmware and generations may change. Full-text/rights boundary: not_permitted

S0199 · 2024-11-18 · B

A Wearable Brain-Computer Interface With Fewer EEG Channels for Online Motor Imagery Detection. Original: S0199. Role: Four-channel online MI feasibility. Limitation: Healthy participants; task accuracy is not rehabilitation benefit or long-term reliability; price and China TCO are not established. Full-text/rights boundary: not_permitted

S0200 · 2024-09-25 · B

Subject-Independent Wearable P300 Brain-Computer Interface Based on Convolutional Neural Network and Metric Learning. Original: S0200. Role: Wearable P300 low-calibration feasibility. Limitation: n=100 P300 speller participants; accuracy is not characters per minute or errors per hour; patient and home evidence absent. Full-text/rights boundary: not_permitted

S0203 · 2024-01-20 · A

Brain computer interface training with motor imagery and functional electrical stimulation for patients with severe upper limb paresis after stroke: a randomized controlled pilot trial. Original: S0203. Role: Stroke BCI-FES randomized pilot balance. Limitation: n=40 randomized and n=35 analyzed; no significant treatment-group difference; protocol-specific severe-paresis subgroup. Full-text/rights boundary: review_required

S0204 · 2026-01-28 · A

Information technology—Brain-computer Interfaces—Reference architecture. Original: S0204. Role: Reference architecture standard status. Limitation: GB/T 47023-2026 is recommended and was not yet effective on 2026-07-24; catalog entry is not a mandatory NMPA rule. Full-text/rights boundary: not_permitted

S0205 · 2026-01-28 · A

Information technology—Brain-computer interfaces—Multi-modal data format. Original: S0205. Role: Multimodal data-format standard status. Limitation: GB/T 47127-2026 is recommended and was not yet effective on 2026-07-24; catalog entry is not a mandatory NMPA rule. Full-text/rights boundary: not_permitted

S0206 · 2026-03-31 · A

Information technology—Brain-computer interfaces—Visual evoked potential data coding and decoding. Original: S0206. Role: Visual-evoked-potential standard status. Limitation: GB/T 47346-2026 is recommended and was not yet effective on 2026-07-24; catalog entry is not a mandatory NMPA rule. Full-text/rights boundary: not_permitted

S0207 · 2026-06-25 · A

MIIT 2026-06-25 notice of five communications industry-standard submissions. Original: S0207. Role: Industry-standard submission notice status. Limitation: Submission notice is not a final standard effective rule or medical-device approval. Full-text/rights boundary: not_permitted

S0211 · 2021-08-20 · A

Personal Information Protection Law of the People's Republic of China. Original: S0211. Role: Personal information law boundary. Limitation: Applies according to processing facts and is not a product approval instrument. Full-text/rights boundary: not_permitted

S0216 · 2026-06-30 · A

NMPA Announcement No. 24 of 2026 Issuing Two Medical Device Product Guiding Principles, Including the Guiding Principles for Classification of Brain-Computer Interface Medical Device Products. Original: S0216. Role: NMPA BCI classification announcement. Limitation: Classification guidance does not convert a research architecture recommendation into product approval. Full-text/rights boundary: not_permitted

S0217 · 2026-06-30 · A

Guiding Principles for Classification of Brain-Computer Interface Medical Device Products. Original: S0217. Role: NMPA BCI classification guidance attachment. Limitation: Guidance defines classification conditions and routes but is not product-specific approval or four mandatory standards. Full-text/rights boundary: not_permitted

S0219 · 2026-06-30 · A

Interpretation of the Guiding Principles for Classification of Brain-Computer Interface Medical Device Products. Original: S0219. Role: Official interpretation of BCI classification guidance. Limitation: Official interpretation explains classification conditions and routes but does not approve a specific product. Full-text/rights boundary: not_permitted