Calibration-Aware Transfer Learning for 125 Hz OpenBCI Motor Imagery BCI With Real-Time Assistive Control

Abstract

Deploying motor-imagery brain-computer interfaces on low-cost hardware remains challenging because practical systems must handle low sampling rates, user variability, and uncertain predictions before triggering real-time actions. This paper evaluates whether knowledge learned from public MI data in BNCI2014_001 can support a 125 Hz OpenBCI platform through short subject-specific calibration and confidence-aware rejection. A Riemannian baseline, EEGNet, and an exploratory pretrained LaBraM branch are compared under native-frequency and 125 Hz conditions, followed by local OpenBCI calibration and bounded assistive execution. At 125 Hz, EEGNet achieved 71.26% within-subject and 65.55% cross-session accuracy; cross-subject accuracy improved on average from 46.18% to 52.69% after rapid calibration. With rejection, accepted EEGNet accuracy reached 74.61% at 71.8% coverage, end-to-end latency stayed within 2.46-2.56 s, and 104/120 real-time scenarios were successful. The results provide controlled proof-of-feasibility for calibration-aware low-cost MI-BCI deployment, while confirming that broader validation is required before claims of generalizable assistive performance.