AIbci Brain-Computer Interface Research Report

Executive summary

As of 23 July 2026, no single non-invasive BCI is the performance, general-purpose, and value winner for every task. The evidence supports three separate decisions: task-specific frontiers, a risk-adjusted research platform, and an auditable low-cost prototype route. Metrics from different paradigms are not pooled into one leaderboard.

Task-specific non-invasive frontier

Recent scalp-EEG studies separately demonstrate online individual-finger class control, continuous two-dimensional pursuit, and low-density dry-electrode code-VEP selection (S0191, S0197, S0192). These are different tasks with different metrics. Selected healthy responders, same-day fine-tuning, accumulated session data, and visual-stimulus constraints prevent extrapolation to unsupervised patient home use or universal cross-day stability.

Best overall implementation

The recommended general research platform remains 32–64-channel professional EEG with hardware event verification, immutable raw data, BIDS, quality gates, classical baselines, compact neural models, confidence rejection, and a deterministic safety state machine. This is a broad platform recommendation, not a claim that more channels win every task.

Best audited value

An OpenBCI-class 8–16-channel wet-electrode system remains the auditable value route for bounded P300, SSVEP, motor-imagery, ERP teaching, and interaction prototypes. A four-channel study supports task feasibility, but does not publish price, supply, interface openness, China delivery, or total cost of ownership (S0199). Lower channel count alone therefore does not establish lower TCO or better value.

Clinical and regulatory balance

Stroke rehabilitation evidence is mixed and must be presented together. A 2025 meta-analysis reported a 3.69-point FMA-UE mean difference with substantial heterogeneity; a 17-centre Chinese RCT reported a 3.35-point between-group difference; a severe-paresis pilot RCT found no significant three-month between-group difference, p=0.382 (S0195, S0194, S0203). This supports hospital-led evaluation of defined protocols, not universal efficacy.

FDA De Novo DEN200046 is a Class II, product-code QOL, 21 CFR 890.5420 regulatory record. NCT05965713 is completed with actual enrolment of 109 but still has no posted results (S0201, S0202). Neither fact establishes Chinese approval or adds an unpublished efficacy result.

Evidence and interpretation rules

Character or word rate, WER, classification accuracy, clinical scales, and task completion time measure different endpoints. Healthy and patient cohorts, random windows and held-out sessions, offline and online tests, and scalp and implanted signals remain separate. Cost ranking is limited to routes that meet the target and have an auditable cost boundary.

Audited scorecard entries

Source catalog

221 normalized source records are available from the bilingual source library. Each entry retains its original title, authoritative URL, evidence boundary, rights status, and download availability. The Chinese controlled report remains the normative detailed edition.